{
 "schema_version": "1.0.0",
 "dataset": "Digital Fly Lab controls ledger",
 "version": "2026-10-06-run9",
 "generated_at": "2026-10-06T11:04:41Z",
 "run_id": "run:c0ecce13-d7a1-40c8-b30e-661b847df700",
 "as_of": "2026-10-06",
 "answer": {
  "one_line": "Sometimes: fly wiring beat scrambled wiring in 15 of 33 studies, mostly reflex, sensory and steering circuits; in trained, reservoir, ML and our walking tests it tied or lost (10).",
  "short": "Of 41 control studies in our catalogue, 33 compare the real wiring with a scrambled or rewired copy: the real wiring helps in 15, makes no difference in 8, does worse in 2, gives mixed results in 7 and is not yet scored in 1. The clear wins are untrained reflex, sensory and steering circuits, while networks trained for machine-learning or game tasks usually do as well with scrambled wiring (a chess-learning mushroom body: 1680 vs 1675 Elo), and a brainless autopilot or a simple rule often matches the fly (Doom 53.3 s vs 51.9 s). The verdict depends on the null: in our own pre-registered test of the sugar reflex, degree-preserving and weight shuffles silence it (0 Hz vs 85 Hz), keeping the sensory and motor boundary keeps about 46% of it, and a sign shuffle makes the whole brain fire. In our own walking test (a FlyGym body driven by the FlyWire brain model through our hand-made mapping), scrambled wiring and a brainless constant drive walked as far as the real brain (14.1 and 14.3 vs 13.9 mm). In our looming test (5-6 Oct 2026), a shadow on one eye turned the fly away in 4 of 4 real-brain body runs and 0 of 3 scrambled body runs (3 more scrambled drives are identical to the no-brain floor, so their walk follows by determinism); scrambled maps silence the steering neurons, scrambled maps turned up to the real map's activity still gave no turn command (3 of 3), and scrambling only the inside of the map kept about -9% of the turn command (none on the far side in all 6 runs) and 69% of the escape signal. A steering study on an aircraft rudder (flight-test-the-fly) also finds the wiring helps, yet a classical yaw damper does far better. Most studies are small, single-author and not replicated.",
  "counts": {
   "studies": 41,
   "with_wiring_null": 33,
   "by_family": {
    "reflex-circuit": {
     "helps": 3,
     "no_difference": 0,
     "worse": 0,
     "mixed": 1,
     "not_tested": 0
    },
    "sensory-model": {
     "helps": 2,
     "no_difference": 0,
     "worse": 0,
     "mixed": 0,
     "not_tested": 1
    },
    "body-steering": {
     "helps": 4,
     "no_difference": 2,
     "worse": 0,
     "mixed": 1,
     "not_tested": 2
    },
    "game-play": {
     "helps": 4,
     "no_difference": 4,
     "worse": 0,
     "mixed": 3,
     "not_tested": 4
    },
    "ml-benchmark": {
     "helps": 1,
     "no_difference": 2,
     "worse": 1,
     "mixed": 1,
     "not_tested": 0
    },
    "reservoir": {
     "helps": 0,
     "no_difference": 0,
     "worse": 1,
     "mixed": 0,
     "not_tested": 0
    },
    "language-model": {
     "helps": 0,
     "no_difference": 0,
     "worse": 0,
     "mixed": 0,
     "not_tested": 1
    },
    "forecasting": {
     "helps": 0,
     "no_difference": 0,
     "worse": 0,
     "mixed": 0,
     "not_tested": 1
    },
    "graph-analysis": {
     "helps": 1,
     "no_difference": 0,
     "worse": 0,
     "mixed": 0,
     "not_tested": 0
    },
    "evolution": {
     "helps": 0,
     "no_difference": 0,
     "worse": 0,
     "mixed": 1,
     "not_tested": 0
    }
   }
  }
 },
 "null_types": [
  {
   "id": "weight-shuffle",
   "label": "Weight shuffle",
   "keeps": "the graph (who connects to whom), degrees, the weight distribution; usually signs",
   "destroys": "which synapse strength sits on which edge",
   "question_answered": "Does the pattern of synapse strengths matter, given the wiring diagram?",
   "pitfall": "Keeps topology, so it tests weights, not wiring; papers often do not say whether topology is kept.",
   "is_wiring_null": true
  },
  {
   "id": "degree-preserving",
   "label": "Degree-preserving rewiring",
   "keeps": "every neuron's in- and out-degree (and often its weights and signs)",
   "destroys": "which specific partners each neuron has",
   "question_answered": "Does the specific partner choice matter beyond each neuron's number of connections?",
   "pitfall": "Rewires the sensory and motor interface too, which can open shortcuts from inputs to outputs (flyconnectome-nulls: 10.7% vs 0.01%); one realisation is not enough.",
   "is_wiring_null": true
  },
  {
   "id": "random-graph",
   "label": "Random graph (Erdos-Renyi)",
   "keeps": "neuron count and edge count (density)",
   "destroys": "degree distribution, motifs, modules, partner identity",
   "question_answered": "Does the fly graph beat any random graph of the same size and density?",
   "pitfall": "Easy to beat: also destroys hubs and degree structure, so a win overstates how specific the wiring is.",
   "is_wiring_null": true
  },
  {
   "id": "target-permutation",
   "label": "Target permutation",
   "keeps": "each source neuron's out-degree, weights and signs",
   "destroys": "in-degrees and which targets each neuron reaches",
   "question_answered": "Does it matter where each neuron's outputs go?",
   "pitfall": "In-degrees change, so some neurons get far more or less input; activity levels shift, not just routing.",
   "is_wiring_null": true
  },
  {
   "id": "sign-shuffle",
   "label": "Sign shuffle",
   "keeps": "topology and synapse sizes",
   "destroys": "which synapses are excitatory and which inhibitory (Dale's law per neuron)",
   "question_answered": "Do the transmitter signs matter?",
   "pitfall": "Changes the E/I balance and activity level; a loss may be due to runaway or silent activity, not lost computation.",
   "is_wiring_null": true
  },
  {
   "id": "boundary-preserving",
   "label": "Boundary-preserving (interior-only) null",
   "keeps": "all sensory-input and descending/motor-output edges; degrees of interior neurons",
   "destroys": "interior (central) wiring only",
   "question_answered": "Does the central wiring matter once the input/output interface is held fixed?",
   "pitfall": "Fairest test of central wiring but weaker: in flyconnectome-nulls it erased the nulls' advantage; interior definition must be declared in advance.",
   "is_wiring_null": true
  },
  {
   "id": "no-edges",
   "label": "No edges / zero weights",
   "keeps": "the neurons, their models and the input/readout",
   "destroys": "all synaptic transmission",
   "question_answered": "How much of the result comes from the network at all?",
   "pitfall": "A floor, not a wiring control: almost any network beats it.",
   "is_wiring_null": false
  },
  {
   "id": "lesion",
   "label": "Lesion / knockout",
   "keeps": "the rest of the connectome",
   "destroys": "named neurons or cell types (silenced or removed)",
   "question_answered": "Is this particular neuron or pathway needed?",
   "pitfall": "A null result can mean the controller bypasses the neuron (cyber-larva MDNa: 0.0% change), not that the neuron is unimportant.",
   "is_wiring_null": false
  },
  {
   "id": "no-graph-model",
   "label": "No-graph model (MLP, CNN, unconstrained RNN)",
   "keeps": "the task, inputs, outputs and training budget",
   "destroys": "the connectome constraint entirely",
   "question_answered": "Does the fly wiring beat a generic learned model?",
   "pitfall": "Parameter counts and tuning effort rarely match; a bigger generic model winning says little about wiring.",
   "is_wiring_null": false
  },
  {
   "id": "untrained",
   "label": "Untrained / random parameters",
   "keeps": "the wiring and neuron model",
   "destroys": "learned or fitted parameters",
   "question_answered": "Is the result due to the wiring alone or to training on top of it?",
   "pitfall": "Tests training, not wiring; a connectome-constrained model with random parameters can still show some features (flyvis contrast preference).",
   "is_wiring_null": false
  },
  {
   "id": "replay",
   "label": "Replay (open-loop playback)",
   "keeps": "the recorded input or activity sequence",
   "destroys": "closed-loop coupling between brain and body/world",
   "question_answered": "Does behaviour depend on live feedback through the network?",
   "pitfall": "Replayed activity can look like a working controller; must be compared against live runs on the same scenes.",
   "is_wiring_null": false
  },
  {
   "id": "no-brain-baseline",
   "label": "No-brain baseline (scripted, PD, random policy)",
   "keeps": "the body, task and scoring",
   "destroys": "the brain model entirely",
   "question_answered": "Does the fly network beat a simple controller without a brain?",
   "pitfall": "Baselines often share scaffolding (pilots, CPGs, steering rules) with the brain run, so both may be carried by the same code.",
   "is_wiring_null": false
  },
  {
   "id": "stimulus-absent",
   "label": "Stimulus absent",
   "keeps": "the network and body",
   "destroys": "the sensory stimulus",
   "question_answered": "Is the behaviour driven by the stimulus rather than by spontaneous or built-in activity?",
   "pitfall": "Shows input dependence, not that the fly wiring is needed; any network can pass it.",
   "is_wiring_null": false
  },
  {
   "id": "label-shuffle",
   "label": "Label / target shuffle",
   "keeps": "inputs and network",
   "destroys": "the pairing of inputs with labels, targets or directions",
   "question_answered": "Is the readout above chance?",
   "pitfall": "Only a chance level; says nothing about wiring.",
   "is_wiring_null": false
  },
  {
   "id": "fly-data",
   "label": "Real fly data",
   "keeps": "nothing simulated",
   "destroys": "-",
   "question_answered": "Does the model match animal behaviour or physiology?",
   "pitfall": "A reference, not a null; literature summary statistics may also be the fitting target (not held out).",
   "is_wiring_null": false
  },
  {
   "id": "other",
   "label": "Other / unclassified",
   "keeps": "depends on the study",
   "destroys": "depends on the study",
   "question_answered": "Study-specific (e.g. all-excitatory, binarised weights, input-specificity sets, unverified shuffles)",
   "pitfall": "Must be described in the arm label; not counted as a wiring null.",
   "is_wiring_null": false
  }
 ],
 "studies": [
  {
   "id": "byo-shiu-shuffle",
   "name": "Build your own: sugar to MN9 against four scrambled-wiring nulls (our run)",
   "evidence_grade": "A",
   "made_by_us": true,
   "task_family": "reflex-circuit",
   "loop": "open-loop",
   "wiring": {
    "dataset": "FlyWire FAFB",
    "release": "v783 (Shiu et al. Completeness_783.csv, Connectivity_783.parquet)",
    "scope": "whole brain, 138,639 neurons"
   },
   "trained_parts": "none",
   "metric": {
    "name": "MN9 (proboscis motor neuron) firing rate with 21 right sugar GRNs at 150 Hz, 1 s trials",
    "unit": "Hz",
    "higher_is_better": true
   },
   "real": {
    "value": 85.2,
    "n": 5,
    "spread": "± 3.56 sd",
    "source": "Digital Fly Lab Build your own run of 2026-09-28, results.json conditions[0] (trials 82, 89, 81, 86, 88 Hz); trial seeds 0-2 re-run in run 5 bit-exact (Digital Fly Lab fair test of 2026-10-01, trial files real-0..2.json)"
   },
   "floor": {
    "label": "no stimulus (network silent)",
    "value": 0,
    "source": "Digital Fly Lab Build your own run of 2026-09-28, results.json: conditions[1] no_stimulus mn9_rate_hz_mean"
   },
   "arms": [
    {
     "arm_id": "D-degree-preserving",
     "null_type": "degree-preserving",
     "label": "Degree-preserving shuffle, 5 shuffle(s) (seeds 1, 2, 3, 4, 5); whole-brain spikes 4021 vs real 13,372-14,284; neurons active 102.7",
     "value": 0,
     "n": 5,
     "samples": 5,
     "spread": "± 0.00 sd across 5 shuffles",
     "readout_refit": null,
     "source": "Digital Fly Lab fair test of 2026-10-01 (report: https://shaduf.ai/p/digital-fly-catalog/reports/2026-10-01-run5/), summary.json arms[D]; seed 1: 0 Hz (run 1 results.json shuffled_wiring_seed1 (trial seed 0; all 3 trials 0 Hz; 126 neurons active over 3 trials)); seed 2: 0 Hz (run 1 results.json shuffled_wiring_seed2 (trial seed 0; all 3 trials 0 Hz; 120 neurons active over 3 trials)); seed 3: 0 Hz (trial file D-3.json); seed 4: 0 Hz (trial file D-4.json); seed 5: 0 Hz (trial file D-5.json)",
     "recomputed_by_us": true,
     "activity_ratio": 0.291,
     "activity_class": "reduced"
    },
    {
     "arm_id": "W-weight-shuffle",
     "null_type": "weight-shuffle",
     "label": "Weight shuffle, 5 shuffle(s) (seeds 101, 102, 103, 104, 105); whole-brain spikes 5363 vs real 13,372-14,284; neurons active 109.0",
     "value": 0,
     "n": 5,
     "samples": 5,
     "spread": "± 0.00 sd across 5 shuffles",
     "readout_refit": null,
     "source": "Digital Fly Lab fair test of 2026-10-01 (report: https://shaduf.ai/p/digital-fly-catalog/reports/2026-10-01-run5/), summary.json arms[W]; seed 101: 0 Hz (trial file W-101.json); seed 102: 0 Hz (trial file W-102.json); seed 103: 0 Hz (trial file W-103.json); seed 104: 0 Hz (trial file W-104.json); seed 105: 0 Hz (trial file W-105.json)",
     "recomputed_by_us": true,
     "activity_ratio": 0.388,
     "activity_class": "reduced"
    },
    {
     "arm_id": "S-sign-shuffle",
     "null_type": "sign-shuffle",
     "label": "Sign shuffle (Dale's law kept), 5 shuffle(s) (seeds 201, 202, 203, 204, 205); whole-brain spikes 3.56261e+06 vs real 13,372-14,284; neurons active 60501.4",
     "value": 214.2,
     "n": 5,
     "samples": 5,
     "spread": "± 43.27 sd across 5 shuffles",
     "readout_refit": null,
     "source": "Digital Fly Lab fair test of 2026-10-01 (report: https://shaduf.ai/p/digital-fly-catalog/reports/2026-10-01-run5/), summary.json arms[S]; seed 201: 268 Hz (trial file S-201.json); seed 202: 188 Hz (trial file S-202.json); seed 203: 191 Hz (trial file S-203.json); seed 204: 253 Hz (trial file S-204.json); seed 205: 171 Hz (trial file S-205.json)",
     "recomputed_by_us": true,
     "activity_ratio": 257.637,
     "activity_class": "runaway"
    },
    {
     "arm_id": "B-boundary-preserving",
     "null_type": "boundary-preserving",
     "label": "Boundary-preserving shuffle, 5 shuffle(s) (seeds 301, 302, 303, 304, 305); whole-brain spikes 8967 vs real 13,372-14,284; neurons active 215.2",
     "value": 39,
     "n": 5,
     "samples": 5,
     "spread": "± 1.22 sd across 5 shuffles",
     "readout_refit": null,
     "source": "Digital Fly Lab fair test of 2026-10-01 (report: https://shaduf.ai/p/digital-fly-catalog/reports/2026-10-01-run5/), summary.json arms[B]; seed 301: 38 Hz (trial file B-301.json); seed 302: 39 Hz (trial file B-302.json); seed 303: 39 Hz (trial file B-303.json); seed 304: 41 Hz (trial file B-304.json); seed 305: 38 Hz (trial file B-305.json)",
     "recomputed_by_us": true,
     "activity_ratio": 0.648,
     "activity_class": "comparable"
    }
   ],
   "effect": {
    "headline_arm": "D-degree-preserving",
    "best_null_arm": "S-sign-shuffle",
    "retention": 0,
    "retention_method": "floor",
    "baseline_arm": null,
    "baseline_retention": null,
    "difference": 85.2,
    "ci": null,
    "p": null,
    "stat_note": "Pre-registered 2026-10-01T10:51:13Z (design published in https://shaduf.ai/p/digital-fly-catalog/reports/2026-10-01-run5/#prereg-heading). One trial per shuffle at trial seed 0; spread is across shuffles. Degree-preserving shuffle: 0 ± 0 Hz (n 5, retention 0.0); Weight shuffle: 0 ± 0 Hz (n 5, retention 0.0); Sign shuffle (Dale's law kept): 214.2 ± 43.27 Hz (n 5, retention 2.514); Boundary-preserving shuffle: 39 ± 1.22 Hz (n 5, retention 0.458). Gain-matched arm (G addendum, pre-registered 2026-10-02T09:35:33Z): not matched (bracket: shuffle 3 x1.5: 5,063 spikes, MN9 0 Hz; shuffle 3 x2: 6,865 spikes, MN9 0 Hz; shuffle 3 x2.5: 10,138 spikes, MN9 0 Hz; shuffle 3 x3: 24,024 spikes, MN9 0 Hz; shuffle 4 x2.5: 24,277 spikes, MN9 0 Hz; shuffle 5 x2.5: 29,297 spikes, MN9 0 Hz; band 9,680-17,976; x 2.5 matched shuffle 3 only, shuffles 4 and 5 were above the band at that gain)."
   },
   "wiring_effect": "mixed",
   "baseline_beaten": "not-tested",
   "fairness": {
    "keeps_degree": true,
    "keeps_io_boundary": true,
    "tuned_equally": false,
    "null_samples": 5,
    "has_no_brain_baseline": false,
    "verdict": "unfair",
    "reason": "Four null types with the same stimulus, readout, neuron model, edge count and synapse count; degree-preserving and boundary-preserving arms keep every neuron's degrees, and the boundary arm keeps every sensory/ascending output and motor/descending/endocrine input. Not tuned equally: the degree-preserving null is much less excitable than the real wiring. The pre-registered gain search matched shuffle 3 at w_syn x 2.5 (10,138 spikes, within 30% of the real 13,828), but shuffles 4 and 5 at the same gain fired 24,277 and 29,297 spikes (above the band), so the arm is not matched on 3 shuffles. MN9 stayed at 0 Hz in all 6 gain trials (x1.5 to x3, 5,063 to 29,297 spikes), which points to mis-routing rather than low excitability."
   },
   "quality": {
    "n_per_arm": 5,
    "held_out": null,
    "uncertainty_reported": true,
    "results_in_repo_files": true,
    "independent_test": true,
    "peer_review": "none",
    "recomputed_by_us": true,
    "grade": "weak",
    "reason": "Our own pre-registered measurement; per-trial files with property checks; one reflex, one model, one trial per shuffle; all 5 pre-registered shuffles per arm ran."
   },
   "extraction": "files",
   "one_line": "Our run: real wiring 85 Hz; degree-preserving 0 Hz; weight shuffle 0 Hz; interior-only shuffle 39 Hz; sign shuffle runaway.",
   "commit": "philshiu/Drosophila_brain_model@91bdd1e7",
   "checked_at": "2026-10-02",
   "sources": [
    {
     "url": "https://github.com/philshiu/Drosophila_brain_model",
     "label": "direct"
    }
   ]
  },
  {
   "id": "drosophila-brain-mlx",
   "name": "drosophila-brain-mlx",
   "evidence_grade": "A",
   "made_by_us": false,
   "task_family": "reflex-circuit",
   "loop": "open-loop",
   "wiring": {
    "dataset": "FlyWire FAFB",
    "release": "v630",
    "scope": "whole brain (127,400 neurons)"
   },
   "trained_parts": "none",
   "metric": {
    "name": "MN9 firing rate, 21 right sugar GRNs at 100 Hz, 30 x 1 s trials",
    "unit": "Hz",
    "higher_is_better": true
   },
   "real": {
    "value": 67.3,
    "n": 30,
    "spread": null,
    "source": "e417b33 README.md:L50-53 (\"On the FlyWire wiring MN9 fires at 67.30 Hz\"); docs/figures/control-demo.svg text \"MN9 67.3 Hz; brain: 290,963 spikes, 408 neurons\""
   },
   "floor": null,
   "arms": [
    {
     "arm_id": "dp-shuffle",
     "null_type": "degree-preserving",
     "label": "copy of the connectome keeping each neuron's in/out degree and outgoing signs and sizes, targets random (seeds 0-4)",
     "value": 0,
     "n": 30,
     "samples": 5,
     "spread": null,
     "readout_refit": false,
     "source": "e417b33 README.md:L53-60 (seed 0: \"MN9 fires no spike at all\"; \"four more shuffles (seeds 1 to 4) leave MN9 silent as well\"); docs/figures/control-demo.svg \"MN9 0.0 Hz ... no MN9 spike in 30 trials\"",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "dp-shuffle",
    "best_null_arm": "dp-shuffle",
    "retention": 0,
    "retention_method": "ratio",
    "baseline_arm": null,
    "baseline_retention": null,
    "difference": 67.3,
    "ci": null,
    "p": null,
    "stat_note": "5 of 5 degree-preserving shuffles give 0 spikes in 30 trials each; input GRNs fire 99.40 vs 99.38 Hz so the input is matched. No spread reported for the real rate."
   },
   "wiring_effect": "helps",
   "baseline_beaten": "not-tested",
   "fairness": {
    "keeps_degree": true,
    "keeps_io_boundary": null,
    "tuned_equally": true,
    "null_samples": 5,
    "has_no_brain_baseline": false,
    "verdict": "fair",
    "reason": "Degree-, sign- and weight-preserving target shuffle, 5 seeds, identical untrained model and identical input spikes; sensory/motor boundary edges are shuffled too."
   },
   "quality": {
    "n_per_arm": 30,
    "held_out": null,
    "uncertainty_reported": false,
    "results_in_repo_files": false,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": false,
    "grade": "weak",
    "reason": "Committed SVG + README only (no raw result table); 5 shuffle seeds but seeds 1-4 reported only as \"silent\"; reproduces Shiu model (Brian2 parity), not peer reviewed."
   },
   "extraction": "files",
   "one_line": "Reimplemented Shiu model: MN9 67.3 Hz on real wiring, silent on all 5 degree-preserving shuffles with identical input.",
   "commit": "e417b33616513ef350b1b1c3cdf2b5b7a1799c8e",
   "checked_at": "2026-09-30T09:45:34Z",
   "sources": [
    {
     "url": "https://github.com/Kisame76/drosophila-brain-mlx/blob/e417b33616513ef350b1b1c3cdf2b5b7a1799c8e/README.md#L40-L60",
     "label": "direct"
    },
    {
     "url": "https://github.com/Kisame76/drosophila-brain-mlx/blob/e417b33616513ef350b1b1c3cdf2b5b7a1799c8e/docs/figures/control-demo.svg",
     "label": "direct"
    }
   ]
  },
  {
   "id": "fly-brain-lulzx",
   "name": "fly-brain",
   "evidence_grade": "A",
   "made_by_us": false,
   "task_family": "reflex-circuit",
   "loop": "open-loop",
   "wiring": {
    "dataset": "MaleCNS",
    "release": "v1.0",
    "scope": "whole CNS (165,122 neurons, edges >= 6 synapses)"
   },
   "trained_parts": "nine global brain parameters fitted by cross-entropy search to the same 17-assay benchmark; flyvis front end pretrained",
   "metric": {
    "name": "17-assay literature benchmark score (mean over 12 paired seeds)",
    "unit": "score 0-1",
    "higher_is_better": true
   },
   "real": {
    "value": 0.7961,
    "n": 12,
    "spread": "± 0.0057 sem",
    "source": "08cf866 public/data/ablation_ladder.json:rungs[key=baseline].score / scoreSem"
   },
   "floor": null,
   "arms": [
    {
     "arm_id": "w_shuffle",
     "null_type": "weight-shuffle",
     "label": "w_shuffle: same synapse-count distribution permuted across retained edges (no refit)",
     "value": 0.4041,
     "n": 12,
     "samples": null,
     "spread": "± 0.0132 sem",
     "readout_refit": false,
     "source": "08cf866 public/data/ablation_ladder.json:rungs[key=w_shuffle].score / scoreSem; paired d -0.392 ± 0.0129 se",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "w_shuffle_refit",
     "null_type": "weight-shuffle",
     "label": "w_shuffle after refitting the nine global parameters (12 gens x 20 pop)",
     "value": 0.5591,
     "n": 12,
     "samples": null,
     "spread": "± 0.0182 sem",
     "readout_refit": true,
     "source": "08cf866 public/data/ablation_refit.json:rungs[key=w_shuffle].refit / refitSem (refit baseline 0.7952 ± 0.0062)",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "sign_free",
     "null_type": "other",
     "label": "sign_free: every neuron excitatory (\"the floor the benchmark must be able to detect\")",
     "value": 0.5241,
     "n": 12,
     "samples": null,
     "spread": "± 0.0021 sem",
     "readout_refit": false,
     "source": "08cf866 public/data/ablation_ladder.json:rungs[key=sign_free].score / scoreSem (refit 0.5515, 08cf866 public/data/ablation_refit.json)",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "w_binary",
     "null_type": "other",
     "label": "w_binary: graded synapse counts replaced by their mean (topology only)",
     "value": 0.3373,
     "n": 12,
     "samples": null,
     "spread": "± 0.0014 sem",
     "readout_refit": false,
     "source": "08cf866 public/data/ablation_ladder.json:rungs[key=w_binary].score / scoreSem (refit 0.5277, 08cf866 public/data/ablation_refit.json)",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "w_shuffle_refit",
    "best_null_arm": "w_shuffle_refit",
    "retention": 0.7023,
    "retention_method": "ratio",
    "baseline_arm": null,
    "baseline_retention": null,
    "difference": 0.237,
    "ci": null,
    "p": null,
    "stat_note": "Author SEMs over 12 paired seeds; paired difference without refit -0.392 ± 0.013 se. With refit, the refit baseline is 0.7952 so the gap is 0.236. Benchmark was also the fitting target (not held out)."
   },
   "wiring_effect": "helps",
   "baseline_beaten": "not-tested",
   "fairness": {
    "keeps_degree": true,
    "keeps_io_boundary": true,
    "tuned_equally": true,
    "null_samples": null,
    "has_no_brain_baseline": false,
    "verdict": "partly-fair",
    "reason": "Weight shuffle keeps topology, degrees and the IO boundary; refit arm gives the shuffle the same parameter search as the real model. Number of distinct shuffle realisations is not stated in the JSON."
   },
   "quality": {
    "n_per_arm": 12,
    "held_out": false,
    "uncertainty_reported": true,
    "results_in_repo_files": true,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": false,
    "grade": "moderate",
    "reason": "Committed JSON with SEMs and 12 paired seeds and a refit arm; but the score is the same benchmark the model was fitted to, shuffle realisations unclear, not peer reviewed."
   },
   "extraction": "files",
   "one_line": "Weight-shuffled MaleCNS scores 0.40 (0.56 after refit) vs 0.80 on the fitted 17-assay benchmark; real counts matter.",
   "commit": "08cf8666bd3cb405c803f95821ebe06d22b3e5ab",
   "checked_at": "2026-09-30T09:45:34Z",
   "sources": [
    {
     "url": "https://github.com/Lulzx/fly-brain/blob/08cf8666bd3cb405c803f95821ebe06d22b3e5ab/public/data/ablation_ladder.json",
     "label": "direct"
    },
    {
     "url": "https://github.com/Lulzx/fly-brain/blob/08cf8666bd3cb405c803f95821ebe06d22b3e5ab/public/data/ablation_refit.json",
     "label": "direct"
    }
   ]
  },
  {
   "id": "shiu-drosophila-brain-model",
   "name": "Drosophila_brain_model (Shiu et al. 2024)",
   "evidence_grade": "A",
   "made_by_us": false,
   "task_family": "reflex-circuit",
   "loop": "open-loop",
   "wiring": {
    "dataset": "FlyWire FAFB",
    "release": "v630",
    "scope": "whole brain (127,400 neurons)"
   },
   "trained_parts": "none (one hand-tuned global w_syn = 0.275 mV)",
   "metric": {
    "name": "fraction of simulations in which MN9 is activated (sugar GRNs at 100 Hz)",
    "unit": "fraction",
    "higher_is_better": true
   },
   "real": {
    "value": 1,
    "n": null,
    "spread": null,
    "source": "paper PMC11446845 Results (feeding-initiation section): \"robust activation of MN9 in 100% of simulations\"; Supplementary Table 1d"
   },
   "floor": null,
   "arms": [
    {
     "arm_id": "weight-shuffle",
     "null_type": "weight-shuffle",
     "label": "connectivity weights shuffled randomly (global weight distribution maintained)",
     "value": 0.01,
     "n": 100,
     "samples": null,
     "spread": null,
     "readout_refit": false,
     "source": "paper PMC11446845 Results: \"only 1 of 100 shuffled simulations did (Supplementary Table 1d)\"",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "weight-shuffle",
    "best_null_arm": "weight-shuffle",
    "retention": 0.01,
    "retention_method": "ratio",
    "baseline_arm": null,
    "baseline_retention": null,
    "difference": 0.99,
    "ci": null,
    "p": null,
    "stat_note": "Counts only (100% vs 1/100); no test or CI in the main text. Number of distinct shuffles and number of real-wiring runs not stated in the text."
   },
   "wiring_effect": "helps",
   "baseline_beaten": "not-tested",
   "fairness": {
    "keeps_degree": null,
    "keeps_io_boundary": null,
    "tuned_equally": true,
    "null_samples": null,
    "has_no_brain_baseline": false,
    "verdict": "partly-fair",
    "reason": "Same model, same w_syn and input for both arms (nothing trained). Paper text does not say whether topology/degrees are kept or how many distinct shuffles were drawn."
   },
   "quality": {
    "n_per_arm": 100,
    "held_out": null,
    "uncertainty_reported": false,
    "results_in_repo_files": false,
    "independent_test": false,
    "peer_review": "peer-reviewed",
    "recomputed_by_us": false,
    "grade": "weak",
    "reason": "Peer-reviewed, 100 shuffled runs, but only counts in paper text; no spread; shuffle design details only in Supplementary Table 1d (not read)."
   },
   "extraction": "paper",
   "one_line": "Real FlyWire weights activate MN9 in every run; shuffled weights in 1 of 100: a large, clean wiring effect on one reflex.",
   "commit": "91bdd1e7dcf193f3e7ca5a8933497fcef63b7960",
   "checked_at": "2026-09-30T09:45:34Z",
   "sources": [
    {
     "url": "https://www.ebi.ac.uk/europepmc/webservices/rest/PMC11446845/fullTextXML",
     "label": "direct"
    },
    {
     "url": "https://github.com/philshiu/Drosophila_brain_model",
     "label": "direct"
    }
   ]
  },
  {
   "id": "fly-ocr",
   "name": "Fly OCR",
   "evidence_grade": "A",
   "made_by_us": false,
   "task_family": "sensory-model",
   "loop": "offline",
   "wiring": {
    "dataset": "MaleCNS",
    "release": "v1.0",
    "scope": "whole CNS, 166,700 neurons / 25,582,938 edges"
   },
   "trained_parts": "readout only (digit pilot: logistic regression on spike counts of 512 feature cells)",
   "metric": {
    "name": "10-class digit accuracy, 200-image pilot",
    "unit": "percent",
    "higher_is_better": true
   },
   "real": {
    "value": 82,
    "n": 200,
    "spread": null,
    "source": "48cf341 reports/controls/controls.json: intact_matched_subset.accuracy (164/200)"
   },
   "floor": {
    "label": "chance (10 classes); edge lesion also 10.0%",
    "value": 10,
    "source": "48cf341 reports/controls/controls.json: edge_lesion_original_head.accuracy (20/200)"
   },
   "arms": [
    {
     "arm_id": "degree_preserving_target_perm",
     "null_type": "degree-preserving",
     "label": "randomized (uniform permutation of edge target vector, exact in/out degree)",
     "value": 61,
     "n": 200,
     "samples": 3,
     "spread": "± 5.68 sd",
     "readout_refit": true,
     "source": "48cf341 reports/controls/controls.json: randomized[seed 41,42,43].neural.accuracy = 65.0/63.5/54.5",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "edge_lesion",
     "null_type": "no-edges",
     "label": "edge lesion, original head",
     "value": 10,
     "n": 200,
     "samples": 1,
     "spread": null,
     "readout_refit": false,
     "source": "48cf341 reports/controls/controls.json: edge_lesion_original_head.accuracy",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "degree_preserving_target_perm",
    "best_null_arm": "degree_preserving_target_perm",
    "retention": 0.7083,
    "retention_method": "floor",
    "baseline_arm": null,
    "baseline_retention": null,
    "difference": 21,
    "ci": null,
    "p": null,
    "stat_note": "Intact 82.0% (Wilson 76.1-86.7) vs rewirings 65.0/63.5/54.5 (each Wilson ~58-71 or lower): all three below the intact interval. Separate no-brain baselines on the final digit task (docs/results.md:L9-18, n=500): fly 89.0%, raw-pixel linear 99.0%, retinal-sample linear 91.8%, CNN 100%."
   },
   "wiring_effect": "helps",
   "baseline_beaten": "not-tested",
   "fairness": {
    "keeps_degree": true,
    "keeps_io_boundary": true,
    "tuned_equally": true,
    "null_samples": 3,
    "has_no_brain_baseline": true,
    "verdict": "fair",
    "reason": "Rewiring keeps exact in/out degree and source weight/sign lists; head refit per seed with same C grid. Lesion uses the original head (distribution shift). Baselines were on the main 500-image set, not the pilot."
   },
   "quality": {
    "n_per_arm": 200,
    "held_out": true,
    "uncertainty_reported": true,
    "results_in_repo_files": true,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": true,
    "grade": "strong",
    "reason": "3 rewiring seeds with refit heads, Wilson intervals, committed JSON; but small 200-image pilot, not the final letter model, and pixel baselines beat the fly."
   },
   "extraction": "files",
   "one_line": "Intact MaleCNS digit pilot 82% beats three degree-preserving target rewirings (mean 61%), but raw-pixel linear readout (99%) beats the fly circuit.",
   "commit": "48cf341e99c17dc911fb09fdc8419d98b4d0ea86",
   "checked_at": "2026-09-30T09:39:44Z",
   "sources": [
    {
     "url": "https://github.com/jerryjliu/fly_ocr",
     "label": "direct"
    },
    {
     "url": "https://raw.githubusercontent.com/jerryjliu/fly_ocr/48cf341e99c17dc911fb09fdc8419d98b4d0ea86/reports/controls/controls.json",
     "label": "direct"
    },
    {
     "url": "https://raw.githubusercontent.com/jerryjliu/fly_ocr/48cf341e99c17dc911fb09fdc8419d98b4d0ea86/docs/results.md",
     "label": "direct"
    }
   ]
  },
  {
   "id": "flydoom-mutkuoz",
   "name": "flydoom",
   "evidence_grade": "A",
   "made_by_us": false,
   "task_family": "sensory-model",
   "loop": "open-loop",
   "wiring": {
    "dataset": "FlyWire (FAFB)",
    "release": "v783",
    "scope": "whole brain"
   },
   "trained_parts": "none (one global synaptic gain calibrated on a non-visual reflex)",
   "metric": {
    "name": "odour-evoked change in lateral-horn neuron (LHN) rate, nose on minus nose off, same Doom level and seed, player frozen",
    "unit": "Hz",
    "higher_is_better": true
   },
   "real": {
    "value": 14.06,
    "n": 10,
    "spread": "± 5.76 sd",
    "source": "86d99a4 paper/data/m8_intact.json deltas.LHN (recomputed)"
   },
   "floor": null,
   "arms": [
    {
     "arm_id": "shuffled",
     "null_type": "degree-preserving",
     "label": "degree-preserving shuffle of the same brain (shuffle_graph=True, graph seed 0)",
     "value": 0.009,
     "n": 10,
     "samples": 1,
     "spread": "± 0.010 sd",
     "readout_refit": false,
     "source": "86d99a4 paper/data/m8_shuffled.json deltas.LHN (recomputed)",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "shuffled",
    "best_null_arm": "shuffled",
    "retention": 0.0006,
    "retention_method": "ratio",
    "baseline_arm": null,
    "baseline_retention": null,
    "difference": 14.051,
    "ci": null,
    "p": null,
    "stat_note": "Same pattern for DNp01 (-17.7 vs -0.002 Hz) and DNa02 (+11.7 vs +0.11 Hz); LC4 control ~0 in both. README.md:L120-125 shows an older run (LHN 0.31 -> 17.63, scrambled +0.01). Separate game test vs command-matched random agent (no-brain baseline; README.md:L232-235, 120 held-out levels): fly arena +5.7 health, +21.0 s survival (95% CI excludes 0), mirrored eyes -0.4/-7.1, sky arena -4.3/-8.2; no wiring null in the game test."
   },
   "wiring_effect": "helps",
   "baseline_beaten": "not-tested",
   "fairness": {
    "keeps_degree": true,
    "keeps_io_boundary": null,
    "tuned_equally": false,
    "null_samples": 1,
    "has_no_brain_baseline": true,
    "verdict": "unfair",
    "reason": "Degree-preserving shuffle but a single shuffled graph; global gain calibrated on the real wiring only (shuffled baseline rates differ: LHN off 25.8 vs 0.32 Hz). Wiring null tests a neural response, not game play; the game test has only a command-matched random agent."
   },
   "quality": {
    "n_per_arm": 10,
    "held_out": null,
    "uncertainty_reported": true,
    "results_in_repo_files": true,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": true,
    "grade": "weak",
    "reason": "Per-seed JSON committed and recomputed; 10 seeds per arm but 1 shuffle; README reports intervals for the game test only."
   },
   "extraction": "files",
   "one_line": "Smell drives the lateral horn +14 Hz with real wiring, ~0 with a degree-preserving shuffle; in-game the fly narrowly beats a command-matched random agent.",
   "commit": "86d99a4f10193203759c3b4266535190e3b09672",
   "checked_at": "2026-09-30T09:44:53Z",
   "sources": [
    {
     "url": "https://github.com/mutkuoz/flydoom/blob/86d99a4f10193203759c3b4266535190e3b09672/paper/data/m8_intact.json",
     "label": "direct"
    },
    {
     "url": "https://github.com/mutkuoz/flydoom/blob/86d99a4f10193203759c3b4266535190e3b09672/paper/data/m8_shuffled.json",
     "label": "direct"
    },
    {
     "url": "https://github.com/mutkuoz/flydoom/blob/86d99a4f10193203759c3b4266535190e3b09672/README.md",
     "label": "direct"
    }
   ]
  },
  {
   "id": "flyvis",
   "name": "flyvis",
   "evidence_grade": "A",
   "made_by_us": false,
   "task_family": "sensory-model",
   "loop": "offline",
   "wiring": {
    "dataset": "other (FlyEM FIB-25 + FIB-19 optic-lobe columns)",
    "release": "fib25-fib19_v2.2",
    "scope": "64 cell types, ~45,669 neurons on 721 columns"
   },
   "trained_parts": "734 parameters (time constants, resting potentials, synapse strength scales) + optic-flow decoder; synapse counts and signs fixed",
   "metric": {
    "name": "characterised cell types whose contrast preference (ON/OFF) is predicted correctly",
    "unit": "cell types (of 32)",
    "higher_is_better": true
   },
   "real": {
    "value": 32,
    "n": 50,
    "spread": null,
    "source": "paper PMC11525180 (Lappalainen et al. 2024 Nature) Results: ensemble median predicts preferred contrast for 32/32 characterised cell types (best single model 30/32); ensemble of 50 task-optimised DMNs"
   },
   "floor": null,
   "arms": [
    {
     "arm_id": "random-dmn",
     "null_type": "untrained",
     "label": "random DMN: full connectome, random single-cell and synapse parameters (not task-optimised)",
     "value": null,
     "n": null,
     "samples": null,
     "spread": null,
     "readout_refit": null,
     "source": "paper PMC11525180 (Lappalainen et al. 2024 Nature) Results \"ablation\" paragraph and Extended Data Fig. 2b-d: accurate contrast preference, poor direction selectivity; values only in figure",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "cell-type-connectivity-only",
     "null_type": "other",
     "label": "task-optimised with only cell-type (not single-neuron) connectivity",
     "value": null,
     "n": null,
     "samples": null,
     "spread": null,
     "readout_refit": true,
     "source": "paper PMC11525180 (Lappalainen et al. 2024 Nature) Results ablation paragraph / Extended Data Fig. 2: \"predicted neural activity poorly\"; values only in figure",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "merge-e-i",
     "null_type": "other",
     "label": "Full DMN Merge E/I: 37 excitatory / 22 inhibitory / 4 mixed types merged",
     "value": null,
     "n": null,
     "samples": null,
     "spread": null,
     "readout_refit": true,
     "source": "paper PMC11525180 (Lappalainen et al. 2024 Nature) Results: \"poor performance on par with the random DMN\"; Extended Data Fig. 2",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "decoder-alone",
     "null_type": "no-brain-baseline",
     "label": "decoder network alone (no DMN)",
     "value": null,
     "n": null,
     "samples": null,
     "spread": null,
     "readout_refit": true,
     "source": "paper PMC11525180 (Lappalainen et al. 2024 Nature) Results: ensemble \"exhibited superior task performance to both the decoder network alone and models with random parameter configurations (Extended Data Fig. 2a)\"; task error only in figure",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "unconstrained-cnn",
     "null_type": "no-graph-model",
     "label": "unconstrained CNN, 5 models x 414,666 parameters (lower bound on task error)",
     "value": null,
     "n": 5,
     "samples": null,
     "spread": null,
     "readout_refit": true,
     "source": "paper PMC11525180 (Lappalainen et al. 2024 Nature) Methods \"Unconstrained CNN\"; task error only in Extended Data Fig. 2a",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": null,
    "best_null_arm": null,
    "retention": null,
    "retention_method": "none",
    "baseline_arm": null,
    "baseline_retention": null,
    "difference": null,
    "ci": null,
    "p": null,
    "stat_note": "No numeric null values in the paper text; Extended Data Fig. 2 shows them graphically only. Qualitative: connectome-constrained task-optimised DMN beats decoder-alone and random-parameter models on task error; random-parameter DMN keeps contrast but loses direction selectivity; cell-type-only connectivity predicts poorly. DSI prediction vs task error r = -0.60, P = 2.6e-6 (df 48)."
   },
   "wiring_effect": "not-tested",
   "baseline_beaten": "not-tested",
   "fairness": {
    "keeps_degree": null,
    "keeps_io_boundary": true,
    "tuned_equally": true,
    "null_samples": null,
    "has_no_brain_baseline": true,
    "verdict": "partly-fair",
    "reason": "Ablation models are task-optimised with the same pipeline, but no wiring-shuffle null exists (ablations remove connectome detail rather than shuffle it); values unreadable from text."
   },
   "quality": {
    "n_per_arm": 50,
    "held_out": true,
    "uncertainty_reported": false,
    "results_in_repo_files": false,
    "independent_test": false,
    "peer_review": "peer-reviewed",
    "recomputed_by_us": false,
    "grade": "weak",
    "reason": "Numbers quoted from the full text (Europe PMC XML, read 2026-10-03). The random-parameter and ablation nulls are shown only in Extended Data Fig. 2, with no number in the text, the SI or the peer-review file and no Source Data, so we record no null values. There is no shuffled-wiring null; a referee asked for a generic network with the same sparseness (peer-review file)."
   },
   "extraction": "paper",
   "one_line": "Connectome-constrained optic-lobe model beats decoder-alone and random-parameter models; ablation values are figure-only, no wiring shuffle.",
   "commit": "92b3845cc426dd309a1a0e1b3890156c42e14021",
   "checked_at": "2026-10-03",
   "sources": [
    {
     "url": "https://www.ebi.ac.uk/europepmc/webservices/rest/PMC11525180/fullTextXML",
     "label": "direct"
    },
    {
     "url": "https://github.com/TuragaLab/flyvis",
     "label": "direct"
    },
    {
     "url": "https://static-content.springer.com/esm/art%3A10.1038%2Fs41586-024-07939-3/MediaObjects/41586_2024_7939_MOESM4_ESM.pdf",
     "label": "direct"
    }
   ]
  },
  {
   "id": "byo-flygym-body",
   "name": "Build your own, Add a body: FlyWire brain drives FlyGym walking (our run)",
   "evidence_grade": "A",
   "made_by_us": true,
   "task_family": "body-steering",
   "loop": "open-loop",
   "wiring": {
    "dataset": "FlyWire FAFB",
    "release": "v783 (Shiu et al. Completeness_783.csv, Connectivity_783.parquet)",
    "scope": "whole brain, 138,639 neurons"
   },
   "trained_parts": "none (hand-made rate-to-drive mapping)",
   "metric": {
    "name": "forward displacement of the FlyGym 2.1.0 fly in 1.0 s of driven body time, DNp09 (P9) stimulated at 150 Hz",
    "unit": "mm",
    "higher_is_better": true
   },
   "real": {
    "value": 13.9252,
    "n": 3,
    "spread": "± 0.58 sd",
    "source": "Digital Fly Lab body test of 2026-10-02 (report: https://shaduf.ai/p/digital-fly-catalog/reports/2026-10-02-run6/), summary.json groups.real; body/P9-real-0..2.json; brain/P9-real-0..2.json"
   },
   "floor": {
    "label": "zero drive (no command)",
    "value": 0,
    "source": "Digital Fly Lab body test of 2026-10-02, body run file zero.json"
   },
   "arms": [
    {
     "arm_id": "D-degree-preserving",
     "null_type": "degree-preserving",
     "label": "degree-preserving shuffles 3, 4, 5 (run 5 nulls.py), DNp09 stimulated, same mapping and body",
     "value": 14.1489,
     "n": 3,
     "samples": 3,
     "spread": "± 0.00 sd",
     "readout_refit": false,
     "source": "Digital Fly Lab body test of 2026-10-02 (report: https://shaduf.ai/p/digital-fly-catalog/reports/2026-10-02-run6/), summary.json groups.D; body/P9-D3..5.json; brain/P9-D-3..5.json",
     "recomputed_by_us": true,
     "activity_ratio": 0.818,
     "activity_class": "comparable"
    },
    {
     "arm_id": "matched-constant",
     "null_type": "no-brain-baseline",
     "label": "no brain: each side's mean real drive, constant for 1 s",
     "value": 14.2782,
     "n": 1,
     "samples": 1,
     "spread": null,
     "readout_refit": null,
     "source": "Digital Fly Lab body test of 2026-10-02 (report: https://shaduf.ai/p/digital-fly-catalog/reports/2026-10-02-run6/), summary.json groups.matched; body/matched.json",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "random-drive",
     "null_type": "no-brain-baseline",
     "label": "no brain: per-bin random drives with the real drive's per-side mean and sd (seeds 0-2)",
     "value": 13.9578,
     "n": 3,
     "samples": 3,
     "spread": "± 0.46 sd",
     "readout_refit": null,
     "source": "Digital Fly Lab body test of 2026-10-02 (report: https://shaduf.ai/p/digital-fly-catalog/reports/2026-10-02-run6/), summary.json groups.random; body/random-0..2.json",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "D-degree-preserving",
    "best_null_arm": "D-degree-preserving",
    "retention": 1.0161,
    "retention_method": "floor",
    "baseline_arm": "matched-constant",
    "baseline_retention": 1.0253,
    "difference": -0.2237,
    "ci": null,
    "p": null,
    "stat_note": "Pre-registered 2026-10-02T09:35:33Z (design published in https://shaduf.ai/p/digital-fly-catalog/reports/2026-10-02-run6/#prereg-heading; deviation 1: the P9 turning sign, fixed before the first body run). One body run per brain trial; physics is deterministic, so identical drives give identical walks (the three D shuffles sent identical drives: only the stimulated P9 neurons reached the mapped DNs). Retention D 1.0161, matched constant 1.0253, random 1.0023. The rate-to-drive mapping is ours (hand-made, not fitted)."
   },
   "wiring_effect": "no-difference",
   "baseline_beaten": "no",
   "fairness": {
    "keeps_degree": true,
    "keeps_io_boundary": false,
    "tuned_equally": true,
    "null_samples": 3,
    "has_no_brain_baseline": true,
    "verdict": "fair",
    "reason": "Same stimulus (DNp09 at 150 Hz), the same fixed mapping and the same FlyGym controller for every arm; nothing was tuned for any arm; brain activity of the scrambled brains is comparable to the real one (whole-brain spikes ratio 0.818). The shuffle does not keep the sensory/motor boundary."
   },
   "quality": {
    "n_per_arm": 3,
    "held_out": null,
    "uncertainty_reported": true,
    "results_in_repo_files": true,
    "independent_test": true,
    "peer_review": "none",
    "recomputed_by_us": true,
    "grade": "strong",
    "reason": "Our own pre-registered measurement with per-run files; one stimulus, one model, one body, 3 trial seeds and 3 shuffles; 1 s of walking."
   },
   "extraction": "files",
   "one_line": "Our test: FlyGym fly walks 13.9 mm with real brain, 14.1 mm scrambled, 14.3 mm with a brainless constant drive; the mapping is ours.",
   "commit": "philshiu/Drosophila_brain_model@91bdd1e7; flygym==2.1.0",
   "checked_at": "2026-10-02",
   "sources": [
    {
     "url": "https://github.com/philshiu/Drosophila_brain_model",
     "label": "direct"
    },
    {
     "url": "https://github.com/NeLy-EPFL/flygym",
     "label": "direct"
    }
   ]
  },
  {
   "id": "byo-flygym-loom",
   "name": "Build your own, Add a sense: a looming shadow turns the FlyGym fly (our run)",
   "evidence_grade": "A",
   "made_by_us": true,
   "task_family": "body-steering",
   "loop": "open-loop",
   "wiring": {
    "dataset": "FlyWire FAFB",
    "release": "v783 (Shiu et al. Completeness_783.csv, Connectivity_783.parquet)",
    "scope": "whole brain, 138,639 neurons"
   },
   "trained_parts": "none (hand-made rate-to-drive mapping)",
   "metric": {
    "name": "heading change away from the looming side, deg, 1.0 s after onset (FlyGym 2.1.0, constant external forward drive 0.8; LC4+LPLC2 of one eye at 80 Hz)",
    "unit": "deg",
    "higher_is_better": true
   },
   "real": {
    "value": 51.6463,
    "n": 4,
    "spread": "± 14.81 sd",
    "source": "Digital Fly Lab looming test of 2026-10-05 (report: https://shaduf.ai/p/digital-fly-catalog/reports/2026-10-05-run8/), summary.json body.mean_away_deg.real; body/{L,R}-real-*.json"
   },
   "floor": {
    "label": "constant external drive (0.8, 0.8), no brain; signed per looming side and pooled",
    "value": 0,
    "source": "Digital Fly Lab looming test of 2026-10-05, body run file F.json"
   },
   "arms": [
    {
     "arm_id": "D-degree-preserving",
     "null_type": "degree-preserving",
     "label": "degree-preserving shuffles (run 5 nulls.py), same stimulus, mapping and body",
     "value": 0,
     "n": 6,
     "samples": 3,
     "spread": "± 1.49 sd",
     "readout_refit": false,
     "source": "Digital Fly Lab looming test of 2026-10-05 (report: https://shaduf.ai/p/digital-fly-catalog/reports/2026-10-05-run8/), summary.json; body/{L,R}-D*.json (deduplicated drives: L-D3=F, L-D4=F, R-D3=F); run 9: Digital Fly Lab looming follow-up of 2026-10-06, drive files  (floor-identical by sha256: LOOML-D-5-f, LOOMR-D-4-f, LOOMR-D-5-f)",
     "recomputed_by_us": true,
     "activity_ratio": 0.4972,
     "activity_class": "reduced"
    },
    {
     "arm_id": "random-steering",
     "null_type": "no-brain-baseline",
     "label": "no brain: random turning term per 100 ms bin, N(0, s) with s = RMS of the real turning terms; no MDN term",
     "value": 6.23,
     "n": 6,
     "samples": 6,
     "spread": "± 12.10 sd",
     "readout_refit": null,
     "source": "Digital Fly Lab looming test of 2026-10-05, body run file rand-*.json",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "G-degree-preserving-gain-matched",
     "null_type": "degree-preserving",
     "label": "degree-preserving shuffles with the recurrent weight scaled until spikes outside the stimulated set match the real brain (+-30%) (recurrent gain 2.25)",
     "value": 1.487,
     "n": 3,
     "samples": 3,
     "spread": "± 0.00 sd",
     "readout_refit": false,
     "source": "Digital Fly Lab looming follow-up of 2026-10-06 (report: https://shaduf.ai/p/digital-fly-catalog/reports/2026-10-06-run9/), summary.json; drives/ (all floor-identical by sha256: LOOML-G-3-g2.25-f, LOOML-G-4-g2.25-f, LOOML-G-5-g2.25-f)",
     "recomputed_by_us": true,
     "activity_ratio": 0.9864,
     "activity_class": "comparable"
    }
   ],
   "effect": {
    "stat_note": "Pre-registered 2026-10-05T09:42:47Z (design published in https://shaduf.ai/p/digital-fly-catalog/reports/2026-10-05-run8/#prereg-heading; deviations 3-5 logged before any body run: the brain trials took about 160 s, so the pre-registered time limits left real seeds 0-1 and shuffles [3, 4] per side). Reading by the pre-registered rule: 'the turn away needs the wiring'. Real runs turned away in 4 of 4 (threshold 10 deg over the floor), shuffles in 0 of 3, random in 3 of 6. Floor-referenced retention D 0.0, random 0.1206. Clip share 0.0. Brain gate (fly67's rule) reproduced in our Brian2 run. The rate-to-drive mapping is ours (hand-made, not fitted); the 0.8 forward drive is external. 6 Oct 2026 brain-level follow-up (pre-registered 2026-10-06T09:42:54Z; fast Poisson input validated against the 5 Oct trials, gate V passed): boundary-preserving scramble B: turn command 'needs-wiring-in-between' (LI retention -0.0853, DNa02 rule 0 of 6), escape signal 'partly' (GF retention 0.6887); giant fibre silenced: 'steering-survives-gf-silencing'; gain-matched G: 'needs-wiring-even-at-same-activity' (m 3, k 0); plain D shuffles: spikes outside the stimulated set 0.072 of real. Brain-level readings do not enter the effect; the 6 Oct drives that differ from the floor wait for body runs (7 pending).",
    "ci": null,
    "p": null,
    "headline_arm": "D-degree-preserving",
    "best_null_arm": "G-degree-preserving-gain-matched",
    "baseline_arm": "random-steering",
    "retention": 0,
    "retention_method": "floor",
    "baseline_retention": 0.1206,
    "difference": 51.6463
   },
   "fairness": {
    "keeps_degree": true,
    "keeps_io_boundary": false,
    "tuned_equally": true,
    "null_samples": 3,
    "has_no_brain_baseline": true,
    "reason": "Same stimulus, the same fixed mapping and the same FlyGym controller for every arm; the only tuned setting is the recurrent gain of the gain-matched arm, set to match the real brain's activity as pre-registered. A gain-matched degree-preserving arm matched the real brain's activity outside the stimulated neurons on 3 shuffles. The plain degree-preserving shuffles are much quieter outside the stimulated neurons.",
    "verdict": "fair"
   },
   "quality": {
    "n_per_arm": 4,
    "held_out": null,
    "uncertainty_reported": true,
    "results_in_repo_files": true,
    "independent_test": true,
    "peer_review": "none",
    "recomputed_by_us": true,
    "reason": "Our own pre-registered measurement with per-run files; one stimulus, one model, one body; 4 real-brain body runs; scrambled arm n 6 over 3 shuffles = 3 body runs plus 3 drives identical to the no-brain floor, whose walk follows by determinism; 1 s after onset.",
    "grade": "strong"
   },
   "extraction": "files",
   "one_line": "Our test: a looming shadow turned the fly away in 4 of 4 real-brain runs, 0 of 3 scrambled body runs (+3 floor-identical); our mapping.",
   "commit": "philshiu/Drosophila_brain_model@91bdd1e7; flygym==2.1.0",
   "checked_at": "2026-10-06",
   "sources": [
    {
     "url": "https://github.com/philshiu/Drosophila_brain_model",
     "label": "direct"
    },
    {
     "url": "https://github.com/znatgost/fly67",
     "label": "direct"
    },
    {
     "url": "https://github.com/NeLy-EPFL/flygym",
     "label": "direct"
    }
   ],
   "wiring_effect": "helps",
   "baseline_beaten": "yes"
  },
  {
   "id": "flight-test-the-fly",
   "name": "Flight-test the fly",
   "evidence_grade": "A",
   "made_by_us": false,
   "task_family": "body-steering",
   "loop": "closed-loop",
   "wiring": {
    "dataset": "FlyWire FAFB",
    "release": "v783 (Shiu et al. 2024 packaging), edges >= 5 synapses",
    "scope": "T4/T5-to-DNa02/DNg02 subcircuit within 4 hops: 20,556 neurons, 251,358 simulated edges"
   },
   "trained_parts": "none; one scalar rudder gain K per wiring chosen on tuning seeds 0-9 by the same margin-constrained grid rule as the classical yaw damper (Shiu LIF parameters unchanged)",
   "metric": {
    "name": "mean RMS yaw rate on 20 held-out turbulence seeds (100-119), JSBSim c172x 6-DOF, moderate MIL-spec turbulence, 60 s flights",
    "unit": "deg/s",
    "higher_is_better": false
   },
   "real": {
    "value": 3.195,
    "n": 20,
    "spread": null,
    "source": "d3fa996 results/phase2b_metrics.json: controllers.fly_real.mean_rms_r = 0.05576 rad/s (README table: 3.19 deg/s)"
   },
   "floor": {
    "label": "bare airframe, no yaw controller (wings-level roll hold only)",
    "value": 3.923,
    "source": "d3fa996 results/phase2b_metrics.json: controllers.bare.mean_rms_r = 0.06848 rad/s (README table: 3.92 deg/s)"
   },
   "arms": [
    {
     "arm_id": "shuffle-median",
     "null_type": "degree-preserving",
     "label": "median of 10 degree- and sign-preserving shuffles (Maslov-Sneppen swaps of post endpoints; in/out degree, edge weights and each neuron's sign kept), each with its own K from the same tuning rule",
     "value": 3.922,
     "n": 20,
     "samples": 10,
     "spread": null,
     "readout_refit": true,
     "source": "d3fa996 results/phase2b_metrics.json: H3_H2b.mean_rms_by_wiring, median of fly_shuffle_00..09 (computed by us from the committed per-wiring means)",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "shuffle-best",
     "null_type": "degree-preserving",
     "label": "best of the 10 shuffles (fly_shuffle_08; K = 0.268, rudder saturated in 99.97% of steps, a relay controller)",
     "value": 3.309,
     "n": 20,
     "samples": 1,
     "spread": null,
     "readout_refit": true,
     "source": "d3fa996 results/phase2b_metrics.json: controllers.fly_shuffle_08.mean_rms_r = 0.05775 rad/s (README table: best scrambled wiring 3.31 deg/s)",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "yaw-damper",
     "null_type": "no-brain-baseline",
     "label": "classical washout yaw damper on measured yaw rate (K = 7.2 per rad/s, same tuning rule and budget)",
     "value": 1.689,
     "n": 20,
     "samples": 1,
     "spread": null,
     "readout_refit": null,
     "source": "d3fa996 results/phase2b_metrics.json: controllers.yaw_damper.mean_rms_r (README table: 1.69 deg/s)",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "shuffle-median",
    "best_null_arm": "shuffle-best",
    "retention": 0.0014,
    "retention_method": "floor",
    "baseline_arm": "yaw-damper",
    "baseline_retention": 3.0687,
    "difference": -0.727,
    "ci": [
     -0.8,
     -0.64
    ],
    "p": 0.0909,
    "stat_note": "CI: 95% percentile bootstrap CI of real minus per-seed median shuffle, -0.72 deg/s [-0.80, -0.64] over 20 test seeds (phase2b_metrics.json H3_H2b, -0.01255 rad/s [-0.01397, -0.01113]; README H2b/H3 row). p = 1/11 is the rank p (real lowest of 11 wirings), the floor with 10 shuffles. Shuffles 01, 02 and 08 selected the top gain and act as saturated relays (RMS sideslip 7.1-7.4 deg vs 1.77 real). 2-state Dutch-roll plant: real 2.65, best shuffle 2.91, bare 3.09, damper 1.50 deg/s (README). Open loop, 1 Hz coherence: real 0.98 vs 19 degree-preserving shuffles median 0.31 and vs 19 type-preserving (boundary-preserving) shuffles max 0.97, rank 1/20, p = 0.05 (results/phase1_summary.md; different metric, not entered as an arm)."
   },
   "fairness": {
    "keeps_degree": true,
    "keeps_io_boundary": null,
    "tuned_equally": true,
    "null_samples": 10,
    "has_no_brain_baseline": true,
    "verdict": "fair",
    "reason": "Degree-, weight- and sign-preserving shuffles on the simulated edges, each tuned with the same gain grid and margin rule as the real wiring and the damper; 10 closed-loop shuffles, held-out test seeds. Whether the shuffle keeps the T4/T5 input and DNa02 readout boundary was not checked (the type-preserving variant that does was run open loop only)."
   },
   "quality": {
    "n_per_arm": 20,
    "held_out": true,
    "uncertainty_reported": true,
    "results_in_repo_files": true,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": true,
    "grade": "strong",
    "reason": "Pre-registered, 20 held-out seeds per arm, 10 shuffles, bootstrap CIs in a committed metrics file; one author, not peer reviewed."
   },
   "extraction": "files",
   "one_line": "Fly yaw-damper circuit: RMS yaw rate 3.19 deg/s vs 3.92 median of 10 degree-preserving shuffles (best 3.31); classical damper 1.69, bare airframe 3.92.",
   "commit": "d3fa996e943e8fc6949c60c79e77444ce533cb08",
   "checked_at": "2026-10-05",
   "sources": [
    {
     "url": "https://github.com/aryawidjaja/flight-test-the-fly/blob/d3fa996e943e8fc6949c60c79e77444ce533cb08/results/phase2b_metrics.json",
     "label": "direct"
    },
    {
     "url": "https://github.com/aryawidjaja/flight-test-the-fly/blob/d3fa996e943e8fc6949c60c79e77444ce533cb08/README.md",
     "label": "direct"
    },
    {
     "url": "https://github.com/aryawidjaja/flight-test-the-fly/blob/d3fa996e943e8fc6949c60c79e77444ce533cb08/results/phase1_summary.md",
     "label": "direct"
    }
   ],
   "wiring_effect": "helps",
   "baseline_beaten": "no"
  },
  {
   "id": "fly-brain-zero-shot",
   "name": "Are fruit flies zero-shot adapters?",
   "evidence_grade": "A",
   "made_by_us": false,
   "task_family": "body-steering",
   "loop": "closed-loop",
   "wiring": {
    "dataset": "FlyWire",
    "release": "v783",
    "scope": "whole brain (Shiu et al. LIF model)"
   },
   "trained_parts": "none in the brain; handoff scale per cell type and turn gain k hand-tuned on separate flies; FlyVis pretrained by its authors",
   "metric": {
    "name": "poles reached per 10 s window (intact flies, 0-10 s)",
    "unit": "poles / 10 s",
    "higher_is_better": true
   },
   "real": {
    "value": 2.958,
    "n": 48,
    "spread": "± 1.13 sd",
    "source": "0c63ba5 assets/results_per_fly.csv (recomputed): intact, real wiring; README.md:L59"
   },
   "floor": {
    "label": "turn neurons unplugged (DNa02 disconnected from the legs)",
    "value": 0.125,
    "source": "0c63ba5 assets/results_per_fly.csv (recomputed): intact, turn neurons unplugged"
   },
   "arms": [
    {
     "arm_id": "scrambled",
     "null_type": "degree-preserving",
     "label": "wiring scrambled (postsynaptic targets permuted across all connections; presynaptic neuron, sign and synapse count kept)",
     "value": 0.344,
     "n": 32,
     "samples": 2,
     "spread": "± 0.65 sd",
     "readout_refit": true,
     "source": "0c63ba5 assets/results_per_fly.csv (recomputed): intact, wiring scrambled (runs shuffle0, shuffle1); README.md:L59",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "unplugged",
     "null_type": "lesion",
     "label": "turn neurons unplugged",
     "value": 0.125,
     "n": 48,
     "samples": 1,
     "spread": "± 0.53 sd",
     "readout_refit": false,
     "source": "0c63ba5 assets/results_per_fly.csv (recomputed): intact, turn neurons unplugged",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "scrambled",
    "best_null_arm": "scrambled",
    "retention": 0.0773,
    "retention_method": "floor",
    "baseline_arm": null,
    "baseline_retention": null,
    "difference": 2.614,
    "ci": null,
    "p": null,
    "stat_note": "After the two-leg cut (5-15 s window, recomputed from the same CSV): real 3.141 ± 1.76 sd (n=64), scrambled 0.125 ± 0.50 sd (n=16, cloned at the cut, 1 run), unplugged 0.521 ± 0.71 sd (n=48). No CI or test reported by the author; gap is > 2 sd."
   },
   "wiring_effect": "helps",
   "baseline_beaten": "not-tested",
   "fairness": {
    "keeps_degree": true,
    "keeps_io_boundary": false,
    "tuned_equally": true,
    "null_samples": 2,
    "has_no_brain_baseline": false,
    "verdict": "partly-fair",
    "reason": "Global target permutation keeps in/out degree, sign and synapse count; turn gain k re-swept (1,2,4,8) on scrambled wiring (vision_loop.py:301-322) but handoff scales not re-tuned. Only 2 shuffles; no brainless steering baseline."
   },
   "quality": {
    "n_per_arm": 32,
    "held_out": true,
    "uncertainty_reported": true,
    "results_in_repo_files": true,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": true,
    "grade": "moderate",
    "reason": "Per-fly table committed and recomputed (matches README); gain chosen on separate training flies; 2 shuffle seeds; no uncertainty reported by the author (sd recomputed by us)."
   },
   "extraction": "files",
   "one_line": "Scrambled wiring silences DNa02 steering: 0.34 vs 2.96 poles per 10 s for real wiring, barely above unplugged turn neurons (0.12).",
   "commit": "0c63ba5cdd123f8d38eea62deb574353f0ec2632",
   "checked_at": "2026-09-30T09:44:53Z",
   "sources": [
    {
     "url": "https://github.com/vib2810/fly-brain-zero-shot/blob/0c63ba5cdd123f8d38eea62deb574353f0ec2632/assets/results_per_fly.csv",
     "label": "direct"
    },
    {
     "url": "https://github.com/vib2810/fly-brain-zero-shot/blob/0c63ba5cdd123f8d38eea62deb574353f0ec2632/README.md",
     "label": "direct"
    },
    {
     "url": "https://github.com/vib2810/fly-brain-zero-shot/blob/0c63ba5cdd123f8d38eea62deb574353f0ec2632/src/flyloop/wiring.py",
     "label": "direct"
    }
   ]
  },
  {
   "id": "fly-connectome-adds-to-body",
   "name": "FLY-lab: What a fly connectome adds to controlling a body",
   "evidence_grade": "A",
   "made_by_us": false,
   "task_family": "body-steering",
   "loop": "closed-loop",
   "wiring": {
    "dataset": "FlyWire",
    "release": "v783",
    "scope": "whole brain (Shiu et al. model, pinned)"
   },
   "trained_parts": "none (3 readout constants grid-searched on calibration seeds)",
   "metric": {
    "name": "test-seed success rate, stimulus left: turn left",
    "unit": "% of episodes",
    "higher_is_better": true
   },
   "real": {
    "value": 100,
    "n": 30,
    "spread": null,
    "source": "c9ee19f results/cd-test-v1/compare-C-vs-B-ab-test-v1/summary.json left.C_successes 30/30; README.md:L14"
   },
   "floor": {
    "label": "A: constant command (no brain)",
    "value": 0,
    "source": "c9ee19f results/cd-test-v1/compare-C-vs-A-ab-test-v1/summary.json left.A_successes 0/30"
   },
   "arms": [
    {
     "arm_id": "E_recal",
     "null_type": "degree-preserving",
     "label": "E: shuffled (double-edge swap per sign), readout recalibrated",
     "value": 52.67,
     "n": 150,
     "samples": 5,
     "spread": "± 50.4 sd across 5 networks",
     "readout_refit": true,
     "source": "c9ee19f results/e-compare-recal-v1/summary.json left (recomputed pooled 79/150)",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "E_fixed",
     "null_type": "degree-preserving",
     "label": "E: shuffled (double-edge swap per sign), fixed readout",
     "value": 31.33,
     "n": 150,
     "samples": 5,
     "spread": "± 34.9 sd across 5 networks",
     "readout_refit": false,
     "source": "c9ee19f results/e-compare-fixed-v1/summary.json left (recomputed pooled 47/150)",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "D_replay",
     "null_type": "replay",
     "label": "D: replay of another episode's network output",
     "value": 0,
     "n": 30,
     "samples": 1,
     "spread": null,
     "readout_refit": false,
     "source": "c9ee19f README.md:L14",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "B_rule",
     "null_type": "no-brain-baseline",
     "label": "B: two-line rule",
     "value": 100,
     "n": 30,
     "samples": 1,
     "spread": null,
     "readout_refit": null,
     "source": "c9ee19f results/cd-test-v1/compare-C-vs-B-ab-test-v1/summary.json left.B_successes 30/30",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "A_constant",
     "null_type": "no-brain-baseline",
     "label": "A: constant command",
     "value": 0,
     "n": 30,
     "samples": 1,
     "spread": null,
     "readout_refit": null,
     "source": "c9ee19f results/cd-test-v1/compare-C-vs-A-ab-test-v1/summary.json left.A_successes",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "E_recal",
    "best_null_arm": "E_recal",
    "retention": 0.5267,
    "retention_method": "floor",
    "baseline_arm": "B_rule",
    "baseline_retention": 1,
    "difference": 47.33,
    "ci": [
     8,
     87
    ],
    "p": null,
    "stat_note": "ci = author's hierarchical bootstrap 95% for E minus C (networks then episodes), sign flipped to real minus null: recal [-0.87,-0.08] -> [8,87] pp; fixed readout -69 pp [-96,-40]. Turn right: all 5 shuffles 0/30 (-100 pp) with fixed and recalibrated readout. No stimulus: all arms except replay 30/30. C vs B: 0 difference, conservative 95% +/-13.6 pp."
   },
   "wiring_effect": "helps",
   "baseline_beaten": "no",
   "fairness": {
    "keeps_degree": true,
    "keeps_io_boundary": false,
    "tuned_equally": true,
    "null_samples": 5,
    "has_no_brain_baseline": true,
    "verdict": "fair",
    "reason": "Double-edge swap keeps in/out degree per sign and weight multiset (brain_loop.py:28-80, self-checked); readout re-fitted on shuffled networks in the recal arm; 5 shuffles; input/readout neuron sets unchanged but their edges rewired."
   },
   "quality": {
    "n_per_arm": 30,
    "held_out": true,
    "uncertainty_reported": true,
    "results_in_repo_files": true,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": true,
    "grade": "strong",
    "reason": "Separate calibration and test seeds, 30 test episodes per arm, 5 shuffles, hierarchical bootstrap CIs, summaries committed; but a two-line rule does as well as the connectome."
   },
   "extraction": "files",
   "one_line": "Connectome beats degree-preserving shuffles on turning (100% vs 53% after readout re-fit, 0% right turns), but a two-line rule ties it; replay fails.",
   "commit": "c9ee19f4f7704d7bf3a26099719c2b85861911db",
   "checked_at": "2026-09-30T09:44:53Z",
   "sources": [
    {
     "url": "https://github.com/Recluse/FLY-lab/tree/c9ee19f4f7704d7bf3a26099719c2b85861911db/results/e-compare-recal-v1",
     "label": "direct"
    },
    {
     "url": "https://github.com/Recluse/FLY-lab/tree/c9ee19f4f7704d7bf3a26099719c2b85861911db/results/e-compare-fixed-v1",
     "label": "direct"
    },
    {
     "url": "https://github.com/Recluse/FLY-lab/tree/c9ee19f4f7704d7bf3a26099719c2b85861911db/results/cd-test-v1",
     "label": "direct"
    },
    {
     "url": "https://github.com/Recluse/FLY-lab/blob/c9ee19f4f7704d7bf3a26099719c2b85861911db/README.md",
     "label": "direct"
    }
   ]
  },
  {
   "id": "fly-exe",
   "name": "Fly.exe (MaleCNS Virtual Fly)",
   "evidence_grade": "B",
   "made_by_us": false,
   "task_family": "body-steering",
   "loop": "closed-loop",
   "wiring": {
    "dataset": "MaleCNS",
    "release": "v1.0",
    "scope": "whole CNS (165,122 neurons, 25,563,197 edges)"
   },
   "trained_parts": "none in the showcase runs (STP/completeness fitted in separate validation)",
   "metric": {
    "name": "giant-fibre lateralisation swing index (left vs right looming object; max +2.0)",
    "unit": "index",
    "higher_is_better": true
   },
   "real": {
    "value": 2,
    "n": 1,
    "spread": null,
    "source": "cc25411 docs/evidence/SHUFFLE_CONTROL_REGIME_MATCHED.md \"The result\" table: intact @ gain 0.40, (120, 0) / (0, 163), swing +2.000"
   },
   "floor": null,
   "arms": [
    {
     "arm_id": "dp-shuffle-matched",
     "null_type": "degree-preserving",
     "label": "degree-preserving target shuffle at regime-matched gain 0.15 (0.97x intact descending rate)",
     "value": -0.098,
     "n": 1,
     "samples": 1,
     "spread": null,
     "readout_refit": false,
     "source": "cc25411 docs/evidence/SHUFFLE_CONTROL_REGIME_MATCHED.md table row \"shuffled @ matched gain 0.15 (0, 104) (4, 78) -0.098\"; src/flysim/multifly_live.py:262 _shuffled_graph",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "dp-shuffle-same-gain",
     "null_type": "degree-preserving",
     "label": "same shuffle at the intact gain 0.40 (6x intact descending rate)",
     "value": 0.169,
     "n": 1,
     "samples": 1,
     "spread": null,
     "readout_refit": false,
     "source": "cc25411 docs/evidence/SHUFFLE_CONTROL_REGIME_MATCHED.md table row \"shuffled @ gain 0.40 (23, 160) (5, 117) +0.169\"",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "eon-shuffled-connectome",
     "null_type": "other",
     "label": "eon showcase shuffled-connectome (diagnostic only): event targets reached, different metric (2 of 4 vs 4, 4, 3 exact)",
     "value": null,
     "n": 1,
     "samples": 1,
     "spread": null,
     "readout_refit": false,
     "source": "cc25411 artifacts/showcase/eon-showcase-v1/acceptance.json:diagnostic_controls.shuffled-connectome.event_targets (value on another scale)",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "dp-shuffle-matched",
    "best_null_arm": "dp-shuffle-same-gain",
    "retention": null,
    "retention_method": "none",
    "baseline_arm": null,
    "baseline_retention": null,
    "difference": 2.098,
    "ci": null,
    "p": null,
    "stat_note": "One shuffle seed, one route. Matched shuffle drives the giant fibre about as hard (104 and 78 spikes vs 120 and 163) but loses side selectivity (-0.098 vs +2.000). Author withdrew the earlier \"topology-specific\" claim based on the unmatched shuffle."
   },
   "wiring_effect": "mixed",
   "baseline_beaten": "not-tested",
   "fairness": {
    "keeps_degree": true,
    "keeps_io_boundary": false,
    "tuned_equally": true,
    "null_samples": 1,
    "has_no_brain_baseline": false,
    "verdict": "partly-fair",
    "reason": "Degree-preserving target shuffle with gain re-tuned to match descending activity (a fair regime match), but a single seed and single route; interface edges are shuffled too."
   },
   "quality": {
    "n_per_arm": 1,
    "held_out": null,
    "uncertainty_reported": false,
    "results_in_repo_files": false,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": true,
    "grade": "weak",
    "reason": "One shuffle seed, no uncertainty, numbers in a committed evidence markdown (no raw table), not peer reviewed; careful regime matching."
   },
   "extraction": "files",
   "one_line": "Regime-matched degree-preserving shuffle still drives the giant fibre but loses left/right selectivity (+2.0 vs -0.1); one seed.",
   "commit": "cc25411e00f945e7330cb6f400372cce315c0802",
   "checked_at": "2026-09-30T09:45:34Z",
   "sources": [
    {
     "url": "https://github.com/Ibtisam-Mohammad/Fly.exe/blob/cc25411e00f945e7330cb6f400372cce315c0802/docs/evidence/SHUFFLE_CONTROL_REGIME_MATCHED.md",
     "label": "direct"
    },
    {
     "url": "https://raw.githubusercontent.com/Ibtisam-Mohammad/Fly.exe/cc25411e00f945e7330cb6f400372cce315c0802/artifacts/showcase/eon-showcase-v1/acceptance.json",
     "label": "direct"
    },
    {
     "url": "https://raw.githubusercontent.com/Ibtisam-Mohammad/Fly.exe/cc25411e00f945e7330cb6f400372cce315c0802/artifacts/showcase/swarm3d-v1/run-exact-summary.json",
     "label": "direct"
    },
    {
     "url": "https://raw.githubusercontent.com/Ibtisam-Mohammad/Fly.exe/cc25411e00f945e7330cb6f400372cce315c0802/artifacts/showcase/swarm3d-v1/run-stimulus-absent-summary.json",
     "label": "direct"
    }
   ]
  },
  {
   "id": "flyarm",
   "name": "FlyArm",
   "evidence_grade": "B",
   "made_by_us": false,
   "task_family": "body-steering",
   "loop": "closed-loop",
   "wiring": {
    "dataset": "MaleCNS",
    "release": "v1.0, edges >= 3 contacts",
    "scope": "whole CNS rate network: 166,700 neurons, 10,520,377 edges; input 1,846 ascending neurons, readout 1,314 descending + 708 motor neurons"
   },
   "trained_parts": "encoder into ascending neurons and linear decoder (behaviour cloning with DAgger); connectome frozen; identical pipeline for connectome, shuffle and GRU",
   "metric": {
    "name": "pick-and-place lift success, first protocol, 6 training seeds x 24 episodes",
    "unit": "% of episodes",
    "higher_is_better": true
   },
   "real": {
    "value": 72.2,
    "n": 144,
    "spread": null,
    "source": "ab3922e docs/results/pick-place-first-protocol-topology.json: outcomes.lift.connectome_rate = 0.7222 (104/144)"
   },
   "floor": null,
   "arms": [
    {
     "arm_id": "shuffle",
     "null_type": "degree-preserving",
     "label": "degree-preserving shuffle (in/out degree, contact multiset, sign and input/output sets kept), 9 shuffles over 6 seeds, same pipeline",
     "value": 55.9,
     "n": 144,
     "samples": 9,
     "spread": null,
     "readout_refit": true,
     "source": "ab3922e docs/results/pick-place-first-protocol-topology.json: outcomes.lift.shuffled_rate = 0.5590 (shuffles averaged within seed); src/flyarm/whole_brain/shuffle.py",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "gru",
     "null_type": "no-graph-model",
     "label": "GRU controller, same encoder/decoder pipeline",
     "value": 56.9,
     "n": 144,
     "samples": 1,
     "spread": null,
     "readout_refit": true,
     "source": "ab3922e docs/results/pick-place-first-protocol-topology.json: outcomes.lift.per_seed[*].gru, pooled by us (82/144)",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "shuffle",
    "best_null_arm": "shuffle",
    "retention": 0.7742,
    "retention_method": "ratio",
    "baseline_arm": "gru",
    "baseline_retention": 0.7881,
    "difference": 16.3,
    "ci": null,
    "p": 0.125,
    "stat_note": "p = seed-level exact sign-flip p over 6 seeds (4 better, 2 worse); episode-level p 0.061 (same file). Other outcomes in the same file: grasp 86.1% vs 88.9% shuffle (p 0.72), place 24.3% vs 19.1% (p 0.14). The earlier 3-seed result (58/72 vs 27/72, RESEARCH_LOG E3) did not replicate in seeds 3-5. Other tracks (RESEARCH_LOG at ab3922e, not entered as arms because metrics differ): kitchen fly above shuffle 25 vs 0, 21.25 vs 15, 25 vs 0 (p 0.125); dexterous hand seed 0 tie 62 vs 63 of 64. Shuffle runs for the headline PPO manipulation protocol (76.3%) were stopped unfinished (STATUS.md), so no shuffle arm exists for it."
   },
   "fairness": {
    "keeps_degree": true,
    "keeps_io_boundary": true,
    "tuned_equally": true,
    "null_samples": 9,
    "has_no_brain_baseline": false,
    "verdict": "fair",
    "reason": "Degree-preserving shuffle keeps in/out degree, contacts, signs and interface sets and goes through the identical training pipeline; 9 shuffles over 6 seeds; trained encoder and decoder can absorb much of the skill."
   },
   "quality": {
    "n_per_arm": 144,
    "held_out": null,
    "uncertainty_reported": true,
    "results_in_repo_files": true,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": true,
    "grade": "strong",
    "reason": "6 seeds x 24 episodes per arm, 9 shuffles, seed- and episode-level p in a committed results file; author says the wiring advantage is not established."
   },
   "extraction": "files",
   "one_line": "Frozen MaleCNS arm controller: pick-and-place lift 72% vs 56% for degree-preserving shuffles over 6 seeds (p 0.125); grasp tied, kitchen ahead, dexterous tie.",
   "commit": "ab3922e7d3dacb8057f95d29bee7f0394a96b7a7",
   "checked_at": "2026-10-05",
   "sources": [
    {
     "url": "https://github.com/yusenthebot/FlyArm/blob/ab3922e7d3dacb8057f95d29bee7f0394a96b7a7/docs/results/pick-place-first-protocol-topology.json",
     "label": "direct"
    },
    {
     "url": "https://github.com/yusenthebot/FlyArm/blob/ab3922e7d3dacb8057f95d29bee7f0394a96b7a7/docs/RESEARCH_LOG.md",
     "label": "direct"
    },
    {
     "url": "https://github.com/yusenthebot/FlyArm/blob/ab3922e7d3dacb8057f95d29bee7f0394a96b7a7/STATUS.md",
     "label": "direct"
    }
   ],
   "wiring_effect": "no-difference",
   "baseline_beaten": "no"
  },
  {
   "id": "flyhard",
   "name": "Flyhard (The Driving Fly)",
   "evidence_grade": "B",
   "made_by_us": false,
   "task_family": "body-steering",
   "loop": "closed-loop",
   "wiring": {
    "dataset": "MaleCNS",
    "release": "v1.0",
    "scope": "whole traced graph topology (165,122 neurons, 25,563,197 edges), unsigned"
   },
   "trained_parts": "all ~25.7 M edge gains and per-neuron leaks (behaviour cloning of an IK teacher, 600 Adam steps, one seed); frozen random input/output projections",
   "metric": {
    "name": "held-out wheel-steering targets reached (within 0.13 rad for the last second of a 3 s trial)",
    "unit": "% of 100 targets",
    "higher_is_better": true
   },
   "real": {
    "value": 100,
    "n": 100,
    "spread": null,
    "source": "328906f reports/2026-09-09/e03-wheel-pilot/metrics.json trained_successes"
   },
   "floor": null,
   "arms": [
    {
     "arm_id": "untrained_core",
     "null_type": "untrained",
     "label": "untrained core (before training, same fixed interfaces)",
     "value": 0,
     "n": 100,
     "samples": 1,
     "spread": null,
     "readout_refit": false,
     "source": "328906f reports/2026-09-09/e03-wheel-pilot/metrics.json initial_successes",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": null,
    "best_null_arm": null,
    "retention": null,
    "retention_method": "none",
    "baseline_arm": null,
    "baseline_retention": null,
    "difference": null,
    "ci": null,
    "p": null,
    "stat_note": "Only a before/after-training comparison; no shuffled/rewired graph and no non-connectome model trained with the same recipe. Catalogue also cites parking 0/50 for learned and reset core (mean final error 3.71 m vs 18.31 m) and three-point turn 6/8 vs 0/2 for reset core (docs only, not opened)."
   },
   "wiring_effect": "not-tested",
   "baseline_beaten": "not-tested",
   "fairness": {
    "keeps_degree": null,
    "keeps_io_boundary": null,
    "tuned_equally": false,
    "null_samples": 0,
    "has_no_brain_baseline": false,
    "verdict": "partly-fair",
    "reason": "The only control is the untrained core; it tests training, not wiring."
   },
   "quality": {
    "n_per_arm": 100,
    "held_out": true,
    "uncertainty_reported": false,
    "results_in_repo_files": true,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": false,
    "grade": "weak",
    "reason": "Committed metrics.json with 100 held-out targets, but one training seed, no uncertainty and no wiring or no-brain control."
   },
   "extraction": "files",
   "one_line": "Trained fly-topology rate network steers a wheel on 100/100 held-out targets vs 0/100 untrained; no wiring null or non-fly baseline.",
   "commit": "328906f4a0e62c8f9fc18805cf6edae6989b82a5",
   "checked_at": "2026-09-30T09:44:53Z",
   "sources": [
    {
     "url": "https://github.com/MarkUnthank/flyhard/blob/328906f4a0e62c8f9fc18805cf6edae6989b82a5/reports/2026-09-09/e03-wheel-pilot/metrics.json",
     "label": "direct"
    }
   ]
  },
  {
   "id": "neurocraft-fly",
   "name": "NeuroCraft fly",
   "evidence_grade": "U",
   "made_by_us": false,
   "task_family": "body-steering",
   "loop": "closed-loop",
   "wiring": {
    "dataset": "not verifiable",
    "release": "not stated in inspectable code",
    "scope": "not stated"
   },
   "trained_parts": "earlier trained readout replaced by a direct controller (README, self-reported)",
   "metric": {
    "name": "behaviour under shuffled weights and without input (shown in the video only)",
    "unit": "not reported",
    "higher_is_better": null
   },
   "real": {
    "value": null,
    "n": null,
    "spread": null,
    "source": "not reported (media/README.md:24-29 describes conditions only)"
   },
   "floor": null,
   "arms": [
    {
     "arm_id": "shuffled-weights",
     "null_type": "weight-shuffle",
     "label": "shuffled weights (video)",
     "value": null,
     "n": null,
     "samples": null,
     "spread": null,
     "readout_refit": null,
     "source": "media/README.md:24-29: 'some responses remain under shuffled weights'; numbers not reported",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "no-input",
     "null_type": "stimulus-absent",
     "label": "sensory input switched off (video)",
     "value": null,
     "n": null,
     "samples": null,
     "spread": null,
     "readout_refit": null,
     "source": "media/README.md:24-29: baseline cruise moves the body with input off; numbers not reported",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": null,
    "best_null_arm": null,
    "retention": null,
    "retention_method": "none",
    "baseline_arm": null,
    "baseline_retention": null,
    "difference": null,
    "ci": null,
    "p": null,
    "stat_note": "Numbers not reported; the code is not released."
   },
   "wiring_effect": "not-tested",
   "baseline_beaten": "not-tested",
   "fairness": {
    "keeps_degree": null,
    "keeps_io_boundary": null,
    "tuned_equally": null,
    "null_samples": null,
    "has_no_brain_baseline": false,
    "verdict": "unclear",
    "reason": "Only described in a README and shown in a video; the shuffle method cannot be inspected."
   },
   "quality": {
    "n_per_arm": null,
    "held_out": null,
    "uncertainty_reported": false,
    "results_in_repo_files": false,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": false,
    "grade": "weak",
    "reason": "No numbers, no code; claim-level only."
   },
   "extraction": "catalogue-text",
   "one_line": "A shuffled-weights run is shown on video, but no numbers or code are published, so the control cannot be checked.",
   "commit": "d121466b3ba2f11498e6da498c74acb061bb00eb",
   "checked_at": "2026-09-29",
   "sources": [
    {
     "url": "https://github.com/evnsnclr/neurocraft-fly-public",
     "label": "direct"
    }
   ]
  },
  {
   "id": "brain-runners",
   "name": "Brain Runners: untrained FlyWire brain (fly2) vs the same rule with no brain and with shuffled wiring",
   "evidence_grade": "A",
   "made_by_us": false,
   "task_family": "game-play",
   "loop": "closed-loop",
   "wiring": {
    "dataset": "FlyWire FAFB",
    "release": "v783 (Shiu et al. repository at 91bdd1e7, sha256-checked: bakeoff/fly/data.py:10-24)",
    "scope": "whole brain, untrained Shiu et al. model"
   },
   "trained_parts": "none (input channel, readout and thresholds chosen on practice seeds, then frozen)",
   "metric": {
    "name": "rows survived on a seeded lane-runner track (of 150), mean over held-out seeds 1200-1399",
    "unit": "rows",
    "higher_is_better": true
   },
   "real": {
    "value": 82.34,
    "n": 200,
    "spread": null,
    "source": "https://github.com/zack-maz/brain-runners/blob/79d80398637909e10687f09f4457b79125203dc6/docs/calibration/FLY2_REPORT.md 'Controls' table: fly2 held-out mean"
   },
   "floor": null,
   "arms": [
    {
     "arm_id": "shuffled-seed1",
     "null_type": "degree-preserving",
     "label": "fly2 on shuffled wiring, seed 1 (targets permuted among connections of the same sign; in/out degree by sign kept; bakeoff/fly/shuffle.py:4-6,16-24)",
     "value": 27.84,
     "n": 200,
     "samples": 1,
     "spread": null,
     "readout_refit": false,
     "source": "https://github.com/zack-maz/brain-runners/blob/79d80398637909e10687f09f4457b79125203dc6/docs/calibration/FLY2_REPORT.md 'Controls' table: shuffled wiring (seed 1)",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "no-brain",
     "null_type": "no-brain-baseline",
     "label": "the same input mapping and rule with no brain",
     "value": 74.05,
     "n": 200,
     "samples": 1,
     "spread": null,
     "readout_refit": null,
     "source": "https://github.com/zack-maz/brain-runners/blob/79d80398637909e10687f09f4457b79125203dc6/docs/calibration/FLY2_REPORT.md 'Controls' table: no brain",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "shuffled-seed1",
    "best_null_arm": "shuffled-seed1",
    "retention": 0.3381,
    "retention_method": "ratio",
    "baseline_arm": "no-brain",
    "baseline_retention": 0.8993,
    "difference": 54.5,
    "ci": null,
    "p": null,
    "stat_note": "No spread reported. Floors on practice seeds (not held-out): random 23.93, always-jump 32.13 ('Floors' table). Authors' own reading: 'the shuffled control is weak evidence' (docs/DECISIONS.md:236-275, decision 43). Read by us at commit 79d8039 on 2026-10-01."
   },
   "wiring_effect": "helps",
   "baseline_beaten": "no",
   "fairness": {
    "keeps_degree": true,
    "keeps_io_boundary": false,
    "tuned_equally": false,
    "null_samples": 1,
    "has_no_brain_baseline": true,
    "verdict": "unfair",
    "reason": "Same frozen input channel, readout and rule for all arms on the same held-out seeds; the shuffle keeps degrees by sign, but the readout was chosen on the real wiring (the shuffled brain's readout neurons never fire), one shuffle only."
   },
   "quality": {
    "n_per_arm": 200,
    "held_out": true,
    "uncertainty_reported": false,
    "results_in_repo_files": true,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": false,
    "grade": "weak",
    "reason": "Pre-specified by the author with held-out seeds, but one shuffle seed and no spread; held-out seeds had been used earlier in candidate research."
   },
   "extraction": "files",
   "one_line": "An untrained fly brain survives 82 rows against 28 on shuffled wiring, but a brainless version of the same rule reaches 74.",
   "commit": "zack-maz/brain-runners@79d8039",
   "checked_at": "2026-10-01",
   "sources": [
    {
     "url": "https://github.com/zack-maz/brain-runners",
     "label": "direct"
    }
   ]
  },
  {
   "id": "chessfly",
   "name": "ChessFly",
   "evidence_grade": "A",
   "made_by_us": false,
   "task_family": "game-play",
   "loop": "closed-loop",
   "wiring": {
    "dataset": "FlyWire FAFB",
    "release": "v783 (Shiu et al. packaging), PN->KC edges >= 5 synapses",
    "scope": "mushroom-body input layer: 285 PNs, 5,177 KCs, 21,509 connections"
   },
   "trained_parts": "KC->MBON weights and a board-channel shortcut by a prediction-error rule from self-play (10,000 games per fly); identical schedule for every wiring",
   "metric": {
    "name": "Chess exam rating after 10,000 self-play games (Elo scale, Random bot = 400; 40-game exams against 4 fixed opponents)",
    "unit": "Elo",
    "higher_is_better": true
   },
   "real": {
    "value": 1680,
    "n": 1,
    "spread": "single fly (seed 1); no spread reported",
    "source": "52276aa brains/adult.json: curve, exam at 10,000 games (wiring 'flywire')"
   },
   "floor": {
    "label": "untrained egg (0 games), same for every wiring",
    "value": 660,
    "source": "52276aa brains/adult.json: curve, exam at 10,000 games: curve[0]"
   },
   "arms": [
    {
     "arm_id": "shuffled",
     "null_type": "degree-preserving",
     "label": "shuffled claws: same claw count per KC, partners re-drawn from the pool of real PN partners (PN popularity kept in expectation)",
     "value": 1675,
     "n": 1,
     "samples": 1,
     "spread": "single fly",
     "readout_refit": true,
     "source": "52276aa brains/adult-shuffled.json: curve, exam at 10,000 games; src/brain.js:90-105",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "random-design",
     "null_type": "random-graph",
     "label": "the original random design: 4,000 KCs with 7 random inputs straight from the board channels",
     "value": 1850,
     "n": 1,
     "samples": 1,
     "spread": "single fly",
     "readout_refit": true,
     "source": "52276aa brains/adult-random.json: curve, exam at 10,000 games",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "shuffled",
    "best_null_arm": "random-design",
    "retention": 0.9951,
    "retention_method": "floor",
    "baseline_arm": null,
    "baseline_retention": null,
    "difference": 5,
    "ci": null,
    "p": null,
    "stat_note": "One fly per wiring and one shuffle (seed 1); ratings from 40-game exams, so differences of tens of Elo are within noise. README (not in files): 200-game exams 1690 / 1695 / 1755 and real vs shuffled 53% to 47% head to head (100 games)."
   },
   "wiring_effect": "no-difference",
   "baseline_beaten": "not-tested",
   "fairness": {
    "keeps_degree": true,
    "keeps_io_boundary": true,
    "tuned_equally": true,
    "null_samples": 1,
    "has_no_brain_baseline": false,
    "reason": "Shuffle keeps each KC's claw count and the PN pool; same self-play schedule and learning rule for all three flies; only one shuffle and one fly per arm.",
    "verdict": "partly-fair"
   },
   "quality": {
    "n_per_arm": 1,
    "held_out": false,
    "uncertainty_reported": false,
    "results_in_repo_files": true,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": false,
    "grade": "weak",
    "reason": "Single fly and single shuffle per arm, no uncertainty; exam values are committed in the brain files."
   },
   "extraction": "files",
   "one_line": "Chess-learning mushroom body: real FlyWire wiring rates 1680, shuffled claws 1675 and a random design 1850 after 10,000 games (one fly each).",
   "commit": "52276aa",
   "checked_at": "2026-10-03",
   "sources": [
    {
     "url": "https://github.com/znatgost/chessfly",
     "label": "direct"
    }
   ]
  },
  {
   "id": "doom-fly-control",
   "name": "Is the fly brain actually playing DOOM? (control experiments)",
   "evidence_grade": "A",
   "made_by_us": false,
   "task_family": "game-play",
   "loop": "closed-loop",
   "wiring": {
    "dataset": "MaleCNS",
    "release": "v1.0",
    "scope": "whole brain (166,700 neurons, 25.6 M edges; DOOMFLY graph)"
   },
   "trained_parts": "none (DOOMFLY's fixed hand-set 'bci' decoder, unchanged)",
   "metric": {
    "name": "survival time per round, combat_survival arena (capped at 90 s)",
    "unit": "s",
    "higher_is_better": true
   },
   "real": {
    "value": 53.34,
    "n": 20,
    "spread": "± 8.40 sd",
    "source": "6b22922 results/a2/real.jsonl (recomputed); results/a2/summary.json real.survival_mean"
   },
   "floor": {
    "label": "all edge weights zero (noconn)",
    "value": 5.5,
    "source": "6b22922 results/a2/noconn.jsonl (recomputed)"
   },
   "arms": [
    {
     "arm_id": "shuf_pooled",
     "null_type": "degree-preserving",
     "label": "shufK: postsynaptic targets permuted across all edges (exact in/out degree), 3 seeds pooled",
     "value": 6.75,
     "n": 12,
     "samples": 3,
     "spread": "± 2.36 sd",
     "readout_refit": false,
     "source": "6b22922 results/a2/shuf0-2.jsonl (recomputed); summary.json 'shuffled (3 seeds pooled)'",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "noconn",
     "null_type": "no-edges",
     "label": "noconn (all edges disconnected)",
     "value": 5.5,
     "n": 8,
     "samples": 1,
     "spread": "± 0.40 sd",
     "readout_refit": false,
     "source": "6b22922 results/a2/noconn.jsonl (recomputed)",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "blind",
     "null_type": "stimulus-absent",
     "label": "blind (photoreceptors zeroed)",
     "value": 10.46,
     "n": 8,
     "samples": 1,
     "spread": "± 3.22 sd",
     "readout_refit": false,
     "source": "6b22922 results/a2/blind.jsonl (recomputed)",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "spray_matched",
     "null_type": "no-brain-baseline",
     "label": "spray_matched: brainless autopilot with the fly's average turn/forward, trigger held",
     "value": 51.93,
     "n": 20,
     "samples": 1,
     "spread": "± 14.04 sd",
     "readout_refit": null,
     "source": "6b22922 results/a2/spray_matched.jsonl (recomputed); summary.json spray_matched",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "shuf_pooled",
    "best_null_arm": "shuf_pooled",
    "retention": 0.0261,
    "retention_method": "floor",
    "baseline_arm": "spray_matched",
    "baseline_retention": 0.9705,
    "difference": 46.59,
    "ci": null,
    "p": null,
    "stat_note": "Author 95% CIs (summary.json): real [49.7, 56.8], shuffled pooled [5.6, 8.1], spray_matched [45.3, 57.3] (overlaps real). Separate doomfly-rl test (results/a1/summary.json, 100 games/arm, trained end to end): basic real 80.57 vs shuf0 80.72 (Welch p 0.92) vs noconn 80.72; rewiring after training -148.44. So wiring matters for the untrained LIF but not for the trained RL agent."
   },
   "wiring_effect": "helps",
   "baseline_beaten": "no",
   "fairness": {
    "keeps_degree": true,
    "keeps_io_boundary": false,
    "tuned_equally": true,
    "null_samples": 3,
    "has_no_brain_baseline": true,
    "verdict": "fair",
    "reason": "Global target permutation keeps in/out degree, sign and synapse mass; neuron identities kept but input/output edges rewired. No tuning in either arm (fixed decoder). Only 4 rounds per shuffle seed."
   },
   "quality": {
    "n_per_arm": 8,
    "held_out": null,
    "uncertainty_reported": true,
    "results_in_repo_files": true,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": true,
    "grade": "strong",
    "reason": "Smallest arm n=8 (real 20, shuffled 12 = 3 seeds x 4, noconn 8, blind 8, spray 20). Per-round JSONL committed and recomputed; bootstrap CIs reported; 3 shuffle seeds but few rounds per seed; a brainless autopilot matches the fly."
   },
   "extraction": "files",
   "one_line": "Untrained fly brain beats shuffled wiring in Doom (53 s vs 7 s), but a brainless autopilot nearly matches; trained RL shows no wiring effect.",
   "commit": "6b2292286af9634bb967fcc672b47ff0ad89cbb6",
   "checked_at": "2026-09-30T09:44:53Z",
   "sources": [
    {
     "url": "https://github.com/gabrycina/doom-fly-control/tree/6b2292286af9634bb967fcc672b47ff0ad89cbb6/results/a2",
     "label": "direct"
    },
    {
     "url": "https://github.com/gabrycina/doom-fly-control/blob/6b2292286af9634bb967fcc672b47ff0ad89cbb6/src/run_a2.py",
     "label": "direct"
    },
    {
     "url": "https://github.com/gabrycina/doom-fly-control/blob/6b2292286af9634bb967fcc672b47ff0ad89cbb6/results/a1/summary.json",
     "label": "direct"
    }
   ]
  },
  {
   "id": "doomfly",
   "name": "DOOMFLY",
   "evidence_grade": "B",
   "made_by_us": false,
   "task_family": "game-play",
   "loop": "closed-loop",
   "wiring": {
    "dataset": "MaleCNS",
    "release": "v1.0",
    "scope": "whole retained graph (166,700 neurons, 25.6 M edges)"
   },
   "trained_parts": "none (hand-set 'bci' decoder gains; optional dopamine-gated KC->MBON11 plasticity in the learning pilot)",
   "metric": {
    "name": "decoded forward command from DNpe017 rate in one 500 ms BCI validation trial",
    "unit": "a.u. (rate x 0.4)",
    "higher_is_better": true
   },
   "real": {
    "value": 16.69,
    "n": 1,
    "spread": null,
    "source": "71ecf53 outputs/doom/bci-validation.json trials[condition=intact].action.forward"
   },
   "floor": {
    "label": "all edges disconnected",
    "value": 0,
    "source": "71ecf53 outputs/doom/bci-validation.json trials[condition=all_edges_disconnected].action.forward"
   },
   "arms": [
    {
     "arm_id": "all_edges_disconnected",
     "null_type": "no-edges",
     "label": "all edges disconnected",
     "value": 0,
     "n": 1,
     "samples": 1,
     "spread": null,
     "readout_refit": false,
     "source": "71ecf53 outputs/doom/bci-validation.json trials[all_edges_disconnected]",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "blank_vision",
     "null_type": "stimulus-absent",
     "label": "black pixels (blank vision)",
     "value": 5.56,
     "n": 1,
     "samples": 1,
     "spread": null,
     "readout_refit": false,
     "source": "71ecf53 outputs/doom/bci-validation.json trials[blank_vision]",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "retina_disconnected",
     "null_type": "lesion",
     "label": "retina disconnected",
     "value": 5.56,
     "n": 1,
     "samples": 1,
     "spread": null,
     "readout_refit": false,
     "source": "71ecf53 outputs/doom/bci-validation.json trials[retina_disconnected]",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": null,
    "best_null_arm": null,
    "retention": null,
    "retention_method": "none",
    "baseline_arm": null,
    "baseline_retention": null,
    "difference": null,
    "ci": null,
    "p": null,
    "stat_note": "Single-trial causal-loop check (turn 1.67 intact vs 0.95 blank/retina-cut vs 0 no edges); file marks learning_demonstrated false. v6 survival pilot (doom-ui/public/learning-iterations.json, 2 held-out episodes each, 8 s cap): plastic 3.66 s, frozen 5.83 s (1 censored), 'shuffled' 3.66 s, where 'shuffled' is a dose-matched time-shifted punishment control for the plasticity rule, not a wiring shuffle. Wiring shuffle of DOOMFLY is tested in doom-fly-control (separate study)."
   },
   "wiring_effect": "not-tested",
   "baseline_beaten": "not-tested",
   "fairness": {
    "keeps_degree": null,
    "keeps_io_boundary": null,
    "tuned_equally": null,
    "null_samples": 0,
    "has_no_brain_baseline": false,
    "verdict": "unclear",
    "reason": "Only lesion/no-edge/blank-input checks of the causal loop; no wiring null and no brainless baseline in this repo."
   },
   "quality": {
    "n_per_arm": 1,
    "held_out": null,
    "uncertainty_reported": false,
    "results_in_repo_files": true,
    "independent_test": true,
    "peer_review": "none",
    "recomputed_by_us": false,
    "grade": "weak",
    "reason": "One trial per condition in a committed JSON; no uncertainty. An independent project (doom-fly-control) later tested this brain with shuffled wiring and a brainless autopilot."
   },
   "extraction": "files",
   "one_line": "DOOMFLY's controls only show that its outputs depend on edges and input (forward 16.7 intact, 0 without edges); wiring specificity untested here.",
   "commit": "71ecf53d78eaffaf1a57ed7b0ccf5d458abc9f33",
   "checked_at": "2026-09-30T09:44:53Z",
   "sources": [
    {
     "url": "https://github.com/nftechie/DOOMFLY/blob/71ecf53d78eaffaf1a57ed7b0ccf5d458abc9f33/outputs/doom/bci-validation.json",
     "label": "direct"
    },
    {
     "url": "https://github.com/nftechie/DOOMFLY/blob/71ecf53d78eaffaf1a57ed7b0ccf5d458abc9f33/doom-ui/public/learning-iterations.json",
     "label": "direct"
    }
   ]
  },
  {
   "id": "doomfly-rl",
   "name": "doomfly-rl",
   "evidence_grade": "A",
   "made_by_us": false,
   "task_family": "game-play",
   "loop": "closed-loop",
   "wiring": {
    "dataset": "MaleCNS",
    "release": "v1.0 49k subgraph (FlyWire 783 also used elsewhere)",
    "scope": "49,393-neuron part, 9.05 M edges"
   },
   "trained_parts": "conv stem, decoder, policy/value heads, per-synapse log-gains and per-neuron scale/shift; distilled from PPO teachers then GRPO fine-tuned (wiring and signs frozen)",
   "metric": {
    "name": "Freedoom MAP01 free-play episode return after 500 GRPO iterations",
    "unit": "return (reward units)",
    "higher_is_better": true
   },
   "real": {
    "value": 6.565,
    "n": 30,
    "spread": "± 1.60 sd",
    "source": "334d915 docs/results/freeplay_heldout_conn_final.json: episodes[].R (30 held-out MAP01 episodes)"
   },
   "floor": null,
   "arms": [
    {
     "arm_id": "shuffled_s0",
     "null_type": "degree-preserving",
     "label": "malecns49k_shuffled_s0 (--shuffle-edges 0)",
     "value": 6.708,
     "n": 30,
     "samples": 1,
     "spread": "± 1.65 sd",
     "readout_refit": true,
     "source": "334d915 docs/results/freeplay_heldout_shuffled_s0_final.json: episodes[].R (30 held-out MAP01 episodes)",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "noconn",
     "null_type": "no-graph-model",
     "label": "malecns49k_noconn (--no-connectome, stem to decoder)",
     "value": 7.394,
     "n": 30,
     "samples": 1,
     "spread": "± 1.28 sd",
     "readout_refit": true,
     "source": "334d915 docs/results/freeplay_heldout_noconn_final.json: episodes[].R (30 held-out MAP01 episodes)",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "shuffled_s0",
    "best_null_arm": "shuffled_s0",
    "retention": 1.0218,
    "retention_method": "ratio",
    "baseline_arm": "noconn",
    "baseline_retention": 1.1263,
    "difference": -0.143,
    "ci": [
     -0.98,
     0.7
    ],
    "p": null,
    "stat_note": "Welch t=-0.34 (real vs shuffle), -2.21 (real vs no-connectome), recomputed by us from per-episode returns; CI approximate. Deaths 13/14/10 of 30. Distillation table docs/controls.md:L20-25 (10 episodes, 5 scenarios, shuffles s0 and s1): every gap inside one std."
   },
   "wiring_effect": "no-difference",
   "baseline_beaten": "no",
   "fairness": {
    "keeps_degree": true,
    "keeps_io_boundary": true,
    "tuned_equally": true,
    "null_samples": 1,
    "has_no_brain_baseline": true,
    "verdict": "partly-fair",
    "reason": "Same recipe, steps and hyperparameters for every backbone; degree-preserving edge shuffle keeps the same neurons and I/O sets. Free play has one shuffle; distillation has two."
   },
   "quality": {
    "n_per_arm": 30,
    "held_out": true,
    "uncertainty_reported": true,
    "results_in_repo_files": true,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": true,
    "grade": "moderate",
    "reason": "30 held-out episodes per arm in committed JSON, recomputed; but one training seed and one shuffle per arm in free play; training logs in S3 only."
   },
   "extraction": "files",
   "one_line": "Doom agent with MaleCNS-49k backbone scores like its degree-preserving shuffle and below a no-connectome model after GRPO; negative result.",
   "commit": "334d9150745d8a65bd03bfaef5bc7bc30cc27d28",
   "checked_at": "2026-09-30T09:40:20Z",
   "sources": [
    {
     "url": "https://github.com/nonatofabio/doomfly-rl",
     "label": "direct"
    },
    {
     "url": "https://raw.githubusercontent.com/nonatofabio/doomfly-rl/334d9150745d8a65bd03bfaef5bc7bc30cc27d28/docs/controls.md",
     "label": "direct"
    },
    {
     "url": "https://raw.githubusercontent.com/nonatofabio/doomfly-rl/334d9150745d8a65bd03bfaef5bc7bc30cc27d28/docs/results/freeplay_heldout_conn_final.json",
     "label": "direct"
    }
   ]
  },
  {
   "id": "fly-dino",
   "name": "Fly Dino (flyjump)",
   "evidence_grade": "C",
   "made_by_us": false,
   "task_family": "game-play",
   "loop": "closed-loop",
   "wiring": {
    "dataset": "MaleCNS",
    "release": "v1.0",
    "scope": "80 neurons (32 visual, 16 descending, 32 bridge), 1,296 edges"
   },
   "trained_parts": "243-parameter 16-12-3 MLP readout (CEM); circuit weights fixed contact counts",
   "metric": {
    "name": "survival time on held-out 180 s Dino courses",
    "unit": "s",
    "higher_is_better": true
   },
   "real": {
    "value": 179.37,
    "n": 100,
    "spread": "± 6.28 sd",
    "source": "e34c661 public/benchmarks/benchmark.json results[name='connectome + trained readout'] meanSeconds 179.372, survived 99/100 (sd recomputed from runs)"
   },
   "floor": {
    "label": "idle (no key presses)",
    "value": 4.51,
    "source": "e34c661 public/benchmarks/benchmark.json results[name='idle'] meanSeconds 4.5085"
   },
   "arms": [
    {
     "arm_id": "direct-8-20-3",
     "null_type": "no-graph-model",
     "label": "direct readout 8-20-3 on the 8 raw observations, no connectome (243 parameters, as many as the connectome readout), same CEM budget and training seed",
     "value": 180,
     "n": 100,
     "samples": 1,
     "spread": "± 0.00 sd",
     "readout_refit": true,
     "source": "e34c661 public/benchmarks/benchmark.json results[name='direct readout 8-20-3 (no connectome)'] meanSeconds 180.0, survived 100/100 (recomputed)",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "direct-8-12-3",
     "null_type": "no-graph-model",
     "label": "direct readout 8-12-3 on the 8 raw observations, no connectome (147 parameters), same CEM budget and training seed",
     "value": 176.89,
     "n": 100,
     "samples": 1,
     "spread": "± 11.78 sd",
     "readout_refit": true,
     "source": "e34c661 public/benchmarks/benchmark.json results[name='direct readout 8-12-3 (no connectome)'] meanSeconds 176.888, survived 92/100 (recomputed)",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "silenced",
     "null_type": "no-edges",
     "label": "circuit output zeroed (same trained readout)",
     "value": 4.51,
     "n": 100,
     "samples": 1,
     "spread": "± 0.01 sd",
     "readout_refit": false,
     "source": "e34c661 public/benchmarks/benchmark.json results[name='circuit output zeroed'] (recomputed)",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "untrained_readout",
     "null_type": "untrained",
     "label": "untrained readout (random readout weights)",
     "value": 4.49,
     "n": 100,
     "samples": 1,
     "spread": "± 0.01 sd",
     "readout_refit": false,
     "source": "e34c661 public/benchmarks/benchmark.json results[name='untrained readout'] (recomputed)",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "rule",
     "null_type": "no-brain-baseline",
     "label": "hand-written rule",
     "value": 46.49,
     "n": 100,
     "samples": 1,
     "spread": "± 21.08 sd",
     "readout_refit": null,
     "source": "e34c661 public/benchmarks/benchmark.json results[name='rule'] (recomputed)",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "random",
     "null_type": "no-brain-baseline",
     "label": "random keys",
     "value": 4.66,
     "n": 100,
     "samples": 1,
     "spread": "± 0.33 sd",
     "readout_refit": null,
     "source": "e34c661 public/benchmarks/benchmark.json results[name='random'] (recomputed)",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": null,
    "best_null_arm": null,
    "retention": null,
    "retention_method": "none",
    "baseline_arm": "direct-8-20-3",
    "baseline_retention": 1.0036,
    "difference": null,
    "ci": null,
    "p": null,
    "stat_note": "No wiring null. Published checkpoint, training seed 20260912. 10 predeclared training seeds per controller on the same 100 courses (public/benchmarks/replicates.json, matches the 20 committed direct seed files): connectome mean 174.90 s, 92.2 courses completed (78-100); direct 8-20-3 171.37 s, 92.2 (56-100); direct 8-12-3 163.64 s, 78.4 (5-100). Zeroed-circuit arm keeps the readout trained on the intact circuit."
   },
   "wiring_effect": "not-tested",
   "baseline_beaten": "no",
   "fairness": {
    "keeps_degree": null,
    "keeps_io_boundary": null,
    "tuned_equally": true,
    "null_samples": 0,
    "has_no_brain_baseline": true,
    "verdict": "partly-fair",
    "reason": "Trained no-connectome readouts (same CEM budget and predeclared seeds) are now included and match the connectome; the zeroed-circuit and untrained arms are not re-trained; still no wiring null."
   },
   "quality": {
    "n_per_arm": 100,
    "held_out": true,
    "uncertainty_reported": true,
    "results_in_repo_files": true,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": true,
    "grade": "moderate",
    "reason": "100 held-out courses per arm in a committed JSON (recomputed), plus 10 training seeds per controller in replicates.json; no wiring null."
   },
   "extraction": "files",
   "one_line": "80-neuron fly circuit plus trained readout survives 179 s on held-out Dino courses; a 243-parameter network without the connectome survives 180 s.",
   "commit": "e34c6614e7d13a1018585f705cdeb59b9f38291d",
   "checked_at": "2026-10-06",
   "sources": [
    {
     "url": "https://github.com/cobanov/flyjump/blob/e34c6614e7d13a1018585f705cdeb59b9f38291d/public/benchmarks/benchmark.json",
     "label": "direct"
    },
    {
     "url": "https://github.com/cobanov/flyjump/blob/e34c6614e7d13a1018585f705cdeb59b9f38291d/public/benchmarks/replicates.json",
     "label": "direct"
    },
    {
     "url": "https://github.com/cobanov/flyjump/blob/e34c6614e7d13a1018585f705cdeb59b9f38291d/src/lib/benchmark.ts",
     "label": "direct"
    }
   ]
  },
  {
   "id": "fly-plays-games",
   "name": "fly-plays-games (Arkanoid / Retroid chapter)",
   "evidence_grade": "C",
   "made_by_us": false,
   "task_family": "game-play",
   "loop": "closed-loop",
   "wiring": {
    "dataset": "MaleCNS",
    "release": "v1.0",
    "scope": "whole brain (loaded via external fly.ai package)"
   },
   "trained_parts": "logistic readout on descending-neuron trace (re-trained for every ablation); connectome weights frozen",
   "metric": {
    "name": "frames to first lost ball, mean over 4 starting scenes (capped at 1500)",
    "unit": "frames",
    "higher_is_better": true
   },
   "real": {
    "value": 1500,
    "n": 4,
    "spread": null,
    "source": "ecb95f7 retroid/README.md:L118 ('>1500 (never lost)', capped)"
   },
   "floor": {
    "label": "weights silenced (all weights zero)",
    "value": 204,
    "source": "ecb95f7 retroid/README.md:L121"
   },
   "arms": [
    {
     "arm_id": "shuffle",
     "null_type": "weight-shuffle",
     "label": "weights shuffled (permuted among existing connections)",
     "value": 1500,
     "n": 4,
     "samples": 2,
     "spread": null,
     "readout_refit": true,
     "source": "ecb95f7 retroid/README.md:L119; retroid/retroidsim/ablate.py:1-14",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "rewire",
     "null_type": "random-graph",
     "label": "topology rewired (each neuron keeps its in-degree, presynaptic partners drawn at random)",
     "value": 204,
     "n": 4,
     "samples": 2,
     "spread": null,
     "readout_refit": true,
     "source": "ecb95f7 retroid/README.md:L120; retroid/retroidsim/ablate.py rewire",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "silence",
     "null_type": "no-edges",
     "label": "weights silenced",
     "value": 204,
     "n": 4,
     "samples": 1,
     "spread": null,
     "readout_refit": true,
     "source": "ecb95f7 retroid/README.md:L121",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "tracker",
     "null_type": "no-brain-baseline",
     "label": "scripted tracker (oracle)",
     "value": 1372,
     "n": 4,
     "samples": 1,
     "spread": null,
     "readout_refit": null,
     "source": "ecb95f7 retroid/README.md:L122",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "random",
     "null_type": "no-brain-baseline",
     "label": "random paddle (two seeds: 111 / 174)",
     "value": 142.5,
     "n": 4,
     "samples": 2,
     "spread": null,
     "readout_refit": null,
     "source": "ecb95f7 retroid/README.md:L123 (mean of 111 and 174 computed by us)",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "no_input",
     "null_type": "stimulus-absent",
     "label": "no input at all",
     "value": 162,
     "n": 4,
     "samples": 1,
     "spread": null,
     "readout_refit": null,
     "source": "ecb95f7 retroid/README.md:L124",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "shuffle",
    "best_null_arm": "shuffle",
    "retention": 1,
    "retention_method": "floor",
    "baseline_arm": "tracker",
    "baseline_retention": 0.9012,
    "difference": 0,
    "ci": null,
    "p": null,
    "stat_note": "Real and weight-shuffled both hit the 1500-frame cap (censored); rewired and silenced fail at 204 frames, about as bad as a random paddle. Uncapped, real loses its first ball at frame 4002 vs 2344 for the scripted tracker (README.md:L134-135). Proxy |dx|: real 2.9 px = shuffle 2.9 px vs rewire/silence 30.6 px (README.md table under the ablate section)."
   },
   "wiring_effect": "mixed",
   "baseline_beaten": "no",
   "fairness": {
    "keeps_degree": false,
    "keeps_io_boundary": false,
    "tuned_equally": true,
    "null_samples": 2,
    "has_no_brain_baseline": true,
    "verdict": "partly-fair",
    "reason": "Readout re-trained on each ablated brain. Weight shuffle keeps topology; rewire keeps in-degree only and flattens the descending trace (same as silence), so it may cut the input pathway rather than test specific wiring. 2 seeds; 4 scenes."
   },
   "quality": {
    "n_per_arm": 4,
    "held_out": null,
    "uncertainty_reported": false,
    "results_in_repo_files": false,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": false,
    "grade": "weak",
    "reason": "Survival numbers only in the README (no committed result files), 4 scenes, capped outcome, no uncertainty."
   },
   "extraction": "files",
   "one_line": "In Arkanoid a weight-shuffled fly brain plays as well as the real one (both hit the 1500-frame cap); rewired or silenced wiring fails by 204.",
   "commit": "ecb95f7b98d273752705942af2405066ff98866b",
   "checked_at": "2026-09-30T09:44:53Z",
   "sources": [
    {
     "url": "https://github.com/blackicon-eth/fly-plays-games/blob/ecb95f7b98d273752705942af2405066ff98866b/retroid/README.md",
     "label": "direct"
    },
    {
     "url": "https://github.com/blackicon-eth/fly-plays-games/blob/ecb95f7b98d273752705942af2405066ff98866b/retroid/retroidsim/ablate.py",
     "label": "direct"
    }
   ]
  },
  {
   "id": "fly-racer",
   "name": "Fly-Racer",
   "evidence_grade": "C",
   "made_by_us": false,
   "task_family": "game-play",
   "loop": "closed-loop",
   "wiring": {
    "dataset": "MaleCNS",
    "release": "v1.0 via neuPrint (male-cns:v1.0)",
    "scope": "Tier S subgraph: 3,350 neurons (800 visual input, 2,430 central brain, 120 descending), 164,773 connections"
   },
   "trained_parts": "PPO trains synapse strengths (topology and signs fixed), gain, time constants, biases, CNN encoder, input projection, Beta readout and critic; same recipe for every core",
   "metric": {
    "name": "mean CarRacing-v3 score of each core's best.pt on 100 full episodes on held-out tracks (deterministic actions)",
    "unit": "score",
    "higher_is_better": true
   },
   "real": {
    "value": 903.2,
    "n": 100,
    "spread": "± 23.2 sd",
    "source": "71ede00 README.md results table: fly '903.2 +/- 23.2' (the +/- is np.std over episodes, flyracer/evaluate.py:57)"
   },
   "floor": null,
   "arms": [
    {
     "arm_id": "random",
     "null_type": "degree-preserving",
     "label": "'random' core: degree-preserving shuffle of the same graph (postsynaptic endpoints permuted; in/out degree and per-neuron signs kept), trained with the same recipe, one seed",
     "value": 883.7,
     "n": 100,
     "samples": 1,
     "spread": "± 82.0 sd",
     "readout_refit": true,
     "source": "71ede00 README.md results table: random '883.7 +/- 82.0'; flyracer/connectome/shuffle.py degree_preserving_shuffle",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "signs",
     "null_type": "sign-shuffle",
     "label": "'signs' core: real wiring with shuffled excitatory/inhibitory labels, same recipe, one seed",
     "value": 896.9,
     "n": 100,
     "samples": 1,
     "spread": "± 28.2 sd",
     "readout_refit": true,
     "source": "71ede00 README.md results table: signs '896.9 +/- 28.2'; flyracer/connectome/shuffle.py shuffle_signs",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "mlp",
     "null_type": "no-graph-model",
     "label": "'mlp' core: plain MLP, no connectome, same recipe, one seed",
     "value": 894,
     "n": 100,
     "samples": 1,
     "spread": "± 75.9 sd",
     "readout_refit": true,
     "source": "71ede00 README.md results table: mlp '894.0 +/- 75.9'",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "random",
    "best_null_arm": "signs",
    "retention": 0.9784,
    "retention_method": "ratio",
    "baseline_arm": "mlp",
    "baseline_retention": 0.9898,
    "difference": 19.5,
    "ci": null,
    "p": null,
    "stat_note": "Numbers only in the README table and docs/core_comparison.png; runs/ (eval.csv, checkpoints) is not committed. One training seed per core; best.pt chosen by evaluating 16 held-out tracks every 25 updates. Medians 912.7 / 904.0 / 911.1 / 909.7 (fly / signs / mlp / random); worst episode 796.3 / 745.0 / 351.2 / 430.9; episodes at 900+ 66 / 57 / 63 / 63. At the end of training all four score 813-865 with mlp slightly ahead (README)."
   },
   "fairness": {
    "keeps_degree": true,
    "keeps_io_boundary": null,
    "tuned_equally": true,
    "null_samples": 1,
    "has_no_brain_baseline": false,
    "verdict": "partly-fair",
    "reason": "Degree-preserving and sign shuffles go through the same PPO recipe as the real wiring; one seed and one shuffle per core; trained encoder, weights and readout can absorb much of the skill."
   },
   "quality": {
    "n_per_arm": 100,
    "held_out": true,
    "uncertainty_reported": true,
    "results_in_repo_files": false,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": false,
    "grade": "weak",
    "reason": "100 held-out episodes per core with sd, but one training seed per core and results only in the README (no run files committed)."
   },
   "extraction": "files",
   "one_line": "PPO car racer on a 3,350-neuron MaleCNS core: 903 mean score vs 884 shuffled, 897 sign-shuffled, 894 MLP; one seed each, README only.",
   "commit": "71ede00f02c29f844188360829d23cb586c0ae90",
   "checked_at": "2026-10-06",
   "sources": [
    {
     "url": "https://github.com/supat-roong/fly-racer/blob/71ede00f02c29f844188360829d23cb586c0ae90/README.md",
     "label": "direct"
    },
    {
     "url": "https://github.com/supat-roong/fly-racer/blob/71ede00f02c29f844188360829d23cb586c0ae90/flyracer/evaluate.py",
     "label": "direct"
    },
    {
     "url": "https://github.com/supat-roong/fly-racer/blob/71ede00f02c29f844188360829d23cb586c0ae90/flyracer/connectome/shuffle.py",
     "label": "direct"
    }
   ],
   "wiring_effect": "mixed",
   "baseline_beaten": "no"
  },
  {
   "id": "fly-self-driving",
   "name": "Fly Self Driving",
   "evidence_grade": "A",
   "made_by_us": false,
   "task_family": "game-play",
   "loop": "closed-loop",
   "wiring": {
    "dataset": "MaleCNS",
    "release": "v1.0",
    "scope": "whole traced graph (165,122 neurons, 25,563,197 edges)"
   },
   "trained_parts": "all 25.7 M per-synapse gains and per-neuron leaks (behaviour cloning + 3 DAgger rounds); frozen random input/output maps",
   "metric": {
    "name": "held-out streets completed, street task v5 with traffic",
    "unit": "% of 20 streets per set",
    "higher_is_better": true
   },
   "real": {
    "value": 96.67,
    "n": 60,
    "spread": null,
    "source": "3516a09 docs/results.md:L18 (20 · 19 · 19 of 20 on three held-out sets, same checkpoint)"
   },
   "floor": {
    "label": "steer straight",
    "value": 0,
    "source": "3516a09 docs/results.md:L12"
   },
   "arms": [
    {
     "arm_id": "rewired",
     "null_type": "target-permutation",
     "label": "same recipe, randomly rewired graph (column indices permuted, fixed seed 20260910)",
     "value": 80,
     "n": 20,
     "samples": 1,
     "spread": null,
     "readout_refit": true,
     "source": "3516a09 docs/results.md:L19; tasks/street/train_street.py:40-41",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "mlp40",
     "null_type": "no-graph-model",
     "label": "MLP, one hidden layer of 40 (46k parameters)",
     "value": 90,
     "n": 20,
     "samples": 1,
     "spread": null,
     "readout_refit": true,
     "source": "3516a09 docs/results.md:L14",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "linear",
     "null_type": "no-graph-model",
     "label": "linear map, pixels to steering (1,153 parameters)",
     "value": 60,
     "n": 20,
     "samples": 1,
     "spread": null,
     "readout_refit": true,
     "source": "3516a09 docs/results.md:L13",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "rewired",
    "best_null_arm": "rewired",
    "retention": 0.8276,
    "retention_method": "floor",
    "baseline_arm": "mlp40",
    "baseline_retention": 0.931,
    "difference": 16.67,
    "ci": null,
    "p": null,
    "stat_note": "One training seed per condition, no intervals. Speed task v6 (docs/results.md:L27-33): connectome 18·19·16/20 vs rewired 5/20, but a second training seed of the connectome gave 13/20. Flat road (L37-48): rewired graph 'did as well or better' than measured wiring; MLP-40 beat both. Scripted expert 20/20 (teacher)."
   },
   "wiring_effect": "mixed",
   "baseline_beaten": "no",
   "fairness": {
    "keeps_degree": false,
    "keeps_io_boundary": false,
    "tuned_equally": true,
    "null_samples": 1,
    "has_no_brain_baseline": true,
    "verdict": "partly-fair",
    "reason": "Column relabelling keeps out-degree and the in-degree multiset but not per-neuron in-degree; same training recipe for real and rewired; one rewiring and one training seed."
   },
   "quality": {
    "n_per_arm": 20,
    "held_out": true,
    "uncertainty_reported": false,
    "results_in_repo_files": false,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": false,
    "grade": "weak",
    "reason": "Best-model and rewired street numbers only in docs/results.md (not in results/ JSON); single seed per condition, no uncertainty; held-out streets."
   },
   "extraction": "files",
   "one_line": "Trained fly graph drives streets 97% vs 80% rewired and 90% for a small MLP; one seed, and the flat-road task showed no wiring advantage.",
   "commit": "3516a094a6462e2a4ab7cc517286d5c3bc3ff63e",
   "checked_at": "2026-09-30T09:44:53Z",
   "sources": [
    {
     "url": "https://github.com/suanmiao/fly-self-driving/blob/3516a094a6462e2a4ab7cc517286d5c3bc3ff63e/docs/results.md",
     "label": "direct"
    },
    {
     "url": "https://github.com/suanmiao/fly-self-driving/blob/3516a094a6462e2a4ab7cc517286d5c3bc3ff63e/tasks/street/train_street.py",
     "label": "direct"
    },
    {
     "url": "https://github.com/suanmiao/fly-self-driving/tree/3516a094a6462e2a4ab7cc517286d5c3bc3ff63e/results/street",
     "label": "direct"
    }
   ]
  },
  {
   "id": "fly-tennis",
   "name": "Connectome ping pong (fly tennis)",
   "evidence_grade": "A",
   "made_by_us": false,
   "task_family": "game-play",
   "loop": "closed-loop",
   "wiring": {
    "dataset": "MaleCNS",
    "release": "v1.0 flat connectome (minconf 0.5, traced only), edges with at least 5 synapses",
    "scope": "whole brain and nerve cord: 163,903 neurons, 6,235,682 edges; input medulla columns (721-ommatidium eye model), readout Mi1/Tm3/T2/T3 bearing decoder"
   },
   "trained_parts": "none (--ckpt \"\" loads no checkpoint); hand-made bump decoder with no trained parameters",
   "metric": {
    "name": "racket hits by both flies in one 40 s match (same seed and ball physics per condition)",
    "unit": "hits",
    "higher_is_better": true
   },
   "real": {
    "value": 54,
    "n": 1,
    "spread": null,
    "source": "c0b5e17 cache/v2/logs/bump2.log: '[bump] 40s: points 1, hits 54, mean rally 54.00, max rally 54' (README table: 54, longest rally 54, no miss)"
   },
   "floor": null,
   "arms": [
    {
     "arm_id": "shuffle",
     "null_type": "degree-preserving",
     "label": "shuffled wiring, same decoder: postsynaptic endpoints permuted across edges (in- and out-degree, weights and signs kept), one shuffle (net.shuffle_edges_(0))",
     "value": 10,
     "n": 1,
     "samples": 1,
     "spread": null,
     "readout_refit": null,
     "source": "c0b5e17 cache/v2/logs/shuffle.log: '[shuffle] 40s: points 12, hits 10, mean rally 0.83, max rally 4' (README table: 10, longest rally 4)",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "parked",
     "null_type": "no-brain-baseline",
     "label": "no steering, fly parked at the centre of its baseline",
     "value": 14,
     "n": 1,
     "samples": 1,
     "spread": null,
     "readout_refit": null,
     "source": "c0b5e17 cache/v2/logs/none.log: '[none] 40s: points 10, hits 14, mean rally 1.40, max rally 6' (README table: 14, 6)",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "dna-steer",
     "null_type": "other",
     "label": "real wiring, steering read from descending neurons DNa01/02 instead of the medulla bearing decoder",
     "value": 12,
     "n": 1,
     "samples": 1,
     "spread": null,
     "readout_refit": null,
     "source": "c0b5e17 cache/v2/logs/contrast.log: '[contrast] 40s: points 10, hits 12, mean rally 1.20, max rally 6' (README table: 12, 6)",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "shuffle",
    "best_null_arm": "shuffle",
    "retention": 0.1852,
    "retention_method": "ratio",
    "baseline_arm": "parked",
    "baseline_retention": 0.2593,
    "difference": 44,
    "ci": null,
    "p": null,
    "stat_note": "One 40 s match per condition, one seed, one shuffle: no spread. Open loop (before each match), decoded bearing over 9 ball positions: real 13, 10, 8, 3, 1, -4, -8, -9, -16 deg (monotonic); shuffled 4, 6, -16, -15, -12, -9, 2, 26, 18 (bump2.log, shuffle.log). In the 54-hit match the logged closed-loop steer check is corr A = -0.230, B = +0.020 (the log says it should be positive), so the long rally is not shown to come from tracking. DNa01/02 arm uses the real wiring with a different readout (entered as 'other')."
   },
   "fairness": {
    "keeps_degree": true,
    "keeps_io_boundary": null,
    "tuned_equally": true,
    "null_samples": 1,
    "has_no_brain_baseline": true,
    "verdict": "partly-fair",
    "reason": "Shuffle keeps in- and out-degree, weights and signs; nothing is trained in any arm and the same decoder rule is used; one shuffle; whether the eye-input and medulla-readout boundary is kept was not checked."
   },
   "quality": {
    "n_per_arm": 1,
    "held_out": null,
    "uncertainty_reported": false,
    "results_in_repo_files": true,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": false,
    "grade": "weak",
    "reason": "Committed match logs, but one 40 s match per arm, one seed, one shuffle, no spread; closed-loop steer check near zero."
   },
   "extraction": "files",
   "one_line": "Untrained MaleCNS flies play ping pong: 54 hits in one 40 s match vs 10 with shuffled wiring and 14 parked; one match each.",
   "commit": "c0b5e17cde94291bd5ef5c87897c2b4e9b8d28c7",
   "checked_at": "2026-10-06",
   "sources": [
    {
     "url": "https://github.com/castor639/fly-tennis/blob/c0b5e17cde94291bd5ef5c87897c2b4e9b8d28c7/README.md",
     "label": "direct"
    },
    {
     "url": "https://github.com/castor639/fly-tennis/blob/c0b5e17cde94291bd5ef5c87897c2b4e9b8d28c7/cache/v2/logs/bump2.log",
     "label": "direct"
    },
    {
     "url": "https://github.com/castor639/fly-tennis/blob/c0b5e17cde94291bd5ef5c87897c2b4e9b8d28c7/cache/v2/logs/shuffle.log",
     "label": "direct"
    },
    {
     "url": "https://github.com/castor639/fly-tennis/blob/c0b5e17cde94291bd5ef5c87897c2b4e9b8d28c7/cache/v2/logs/none.log",
     "label": "direct"
    },
    {
     "url": "https://github.com/castor639/fly-tennis/blob/c0b5e17cde94291bd5ef5c87897c2b4e9b8d28c7/cache/v2/logs/contrast.log",
     "label": "direct"
    },
    {
     "url": "https://github.com/castor639/fly-tennis/blob/c0b5e17cde94291bd5ef5c87897c2b4e9b8d28c7/flypp/network.py",
     "label": "direct"
    }
   ],
   "wiring_effect": "helps",
   "baseline_beaten": "yes"
  },
  {
   "id": "fly-worker",
   "name": "Fly Worker",
   "evidence_grade": "A",
   "made_by_us": false,
   "task_family": "game-play",
   "loop": "closed-loop",
   "wiring": {
    "dataset": "FlyWire",
    "release": "v783",
    "scope": "whole brain (138,639 neurons, 2,700,513 connections >= 5 synapses)"
   },
   "trained_parts": "none (1.2 s start-up calibration picks and normalises up to 40 DN channels)",
   "metric": {
    "name": "map coverage in the QA game benchmark (24,000 frames per run)",
    "unit": "% of map",
    "higher_is_better": true
   },
   "real": {
    "value": 42,
    "n": 3,
    "spread": "range 35-50% over 3 runs",
    "source": "83b9104 LIMITATIONS.md:L192; README.md:L48"
   },
   "floor": {
    "label": "straight ahead only",
    "value": 5,
    "source": "83b9104 LIMITATIONS.md:L195"
   },
   "arms": [
    {
     "arm_id": "smooth_noise",
     "null_type": "no-brain-baseline",
     "label": "smoothed noise",
     "value": 61,
     "n": 5,
     "samples": 1,
     "spread": "range 57-66% over 5 runs",
     "readout_refit": null,
     "source": "83b9104 LIMITATIONS.md:L193",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "uniform_noise",
     "null_type": "no-brain-baseline",
     "label": "uniform noise",
     "value": 50,
     "n": 5,
     "samples": 1,
     "spread": "range 46-52% over 5 runs",
     "readout_refit": null,
     "source": "83b9104 LIMITATIONS.md:L194",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "paired_noise_alone",
     "null_type": "no-brain-baseline",
     "label": "paired control: same noise sequence without the connectome term (vs fly+noise hybrid 53%, range 49-56)",
     "value": 56,
     "n": null,
     "samples": 1,
     "spread": "range 52-63%",
     "readout_refit": null,
     "source": "83b9104 LIMITATIONS.md:L218-219",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": null,
    "best_null_arm": null,
    "retention": null,
    "retention_method": "none",
    "baseline_arm": "smooth_noise",
    "baseline_retention": 1.5135,
    "difference": null,
    "ci": null,
    "p": null,
    "stat_note": "No wiring null. Fly loses to smoothed noise (ranges do not overlap). Paired test: fly+noise 53% vs same noise alone 56% (connectome adds -3 pp, within run-to-run range). Steering vs light direction correlation ~0 over 8 seeds (0.046 ± 0.272, LIMITATIONS.md:15-24). Lesion: static left/right separation 0.947 -> 0.000 after cutting 40 steering channels -> 1.121 restored (README.md:L69-72)."
   },
   "wiring_effect": "not-tested",
   "baseline_beaten": "no",
   "fairness": {
    "keeps_degree": null,
    "keeps_io_boundary": null,
    "tuned_equally": null,
    "null_samples": 0,
    "has_no_brain_baseline": true,
    "verdict": "unclear",
    "reason": "Only no-brain baselines (noise policies, straight) and a paired noise-alone control; no shuffled or rewired connectome."
   },
   "quality": {
    "n_per_arm": 3,
    "held_out": null,
    "uncertainty_reported": false,
    "results_in_repo_files": false,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": false,
    "grade": "weak",
    "reason": "3-5 seeded runs per policy with ranges, but numbers only in README/LIMITATIONS (docs/policy.js hard-codes the table); policy_bench.mjs is committed but no output file."
   },
   "extraction": "files",
   "one_line": "Whole-brain fly explores 42% of a game map vs 61% for smoothed noise; adding the connectome to noise changes nothing (-3 points).",
   "commit": "83b910466ed845a826cb291cb044eeb82dc8b956",
   "checked_at": "2026-09-30T09:44:53Z",
   "sources": [
    {
     "url": "https://github.com/hwkim3330/flyworker/blob/83b910466ed845a826cb291cb044eeb82dc8b956/LIMITATIONS.md",
     "label": "direct"
    },
    {
     "url": "https://github.com/hwkim3330/flyworker/blob/83b910466ed845a826cb291cb044eeb82dc8b956/README.md",
     "label": "direct"
    },
    {
     "url": "https://github.com/hwkim3330/flyworker/blob/83b910466ed845a826cb291cb044eeb82dc8b956/policy_bench.mjs",
     "label": "direct"
    }
   ]
  },
  {
   "id": "flyaim",
   "name": "FlyAim",
   "evidence_grade": "A",
   "made_by_us": false,
   "task_family": "game-play",
   "loop": "closed-loop",
   "wiring": {
    "dataset": "MaleCNS",
    "release": "v1.0",
    "scope": "whole CNS: 166,700 neurons, 25,582,938 edges (124,177,616 synapses); input 6,098 ol_sensory photoreceptors, readout 1,360 descending neurons"
   },
   "trained_parts": "ridge-regression readout from the 1,360 descending neurons to crosshair velocity (lambda 10, 2 training seeds x 500 frames); connectome weights frozen (assert_connectome_frozen)",
   "metric": {
    "name": "mean crosshair-to-target distance in a 640x480 aiming arena (10 seeds x 900 frames)",
    "unit": "px",
    "higher_is_better": false
   },
   "real": {
    "value": 412.51,
    "n": 10,
    "spread": "± 133.64 sd",
    "source": "15d800f flyaim/runs/20261003-195410-phase3/raw_results.json: results.fly[*].summary.mean_target_dist_px, mean and sd over seeds 0-9 (recomputed; REPORT.md table 412.5)"
   },
   "floor": null,
   "arms": [
    {
     "arm_id": "shuffle",
     "null_type": "degree-preserving",
     "label": "shuffled wiring: CSR column indices permuted, per-row nnz and weight multiset kept (one shuffle)",
     "value": 430.88,
     "n": 10,
     "samples": 1,
     "spread": "± 136.67 sd",
     "readout_refit": null,
     "source": "15d800f flyaim/runs/20261003-195410-phase3/raw_results.json: results.shuffle (recomputed; REPORT.md 430.9); flyaim/baselines/shuffle.py",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "random",
     "null_type": "no-brain-baseline",
     "label": "uniform random walker",
     "value": 251.69,
     "n": 10,
     "samples": 1,
     "spread": "± 73.29 sd",
     "readout_refit": null,
     "source": "15d800f flyaim/runs/20261003-195410-phase3/raw_results.json: results.random (recomputed; REPORT.md 251.7)",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "pid",
     "null_type": "no-brain-baseline",
     "label": "PID controller that reads the target position (upper reference)",
     "value": 155.83,
     "n": 10,
     "samples": 1,
     "spread": "± 12.82 sd",
     "readout_refit": null,
     "source": "15d800f flyaim/runs/20261003-195410-phase3/raw_results.json: results.pid (recomputed; REPORT.md 155.8)",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "shuffle",
    "best_null_arm": "shuffle",
    "retention": null,
    "retention_method": "none",
    "baseline_arm": "pid",
    "baseline_retention": null,
    "difference": -18.37,
    "ci": [
     -109.8,
     73
    ],
    "p": 0.66,
    "stat_note": "Paired fly minus shuffle over seeds 0-9: mean -18.4 px, 95% t CI computed by us from per-seed values; p = 0.66 is the author's paired test on distance (README table; dz -0.14). The REPORT's pre-registered rule used hit rate (0 for both network arms, p = 1.0). Fly vs random +160.8 px (p 0.0069) and shuffle vs random +179.2 px (p 0.0080): both network arms are worse than random. Plasticity arm (flyaim/runs/plastic/verdict.json, per catalogue grade basis, not re-read): real plastic 432.9 vs shuffle plastic 432.2 px, p 0.68; not entered as an arm because its real value differs. Readout refit for the shuffle arm not confirmed in code; this run's config uses dt 1 ms x 33 steps per frame."
   },
   "fairness": {
    "keeps_degree": true,
    "keeps_io_boundary": true,
    "tuned_equally": null,
    "null_samples": 1,
    "has_no_brain_baseline": true,
    "verdict": "partly-fair",
    "reason": "Shuffle keeps each row's nnz (one side of the degree) and the weight multiset, and the photoreceptor input and DN readout sets; one shuffle; whether the readout was refit on the shuffled network was not confirmed."
   },
   "quality": {
    "n_per_arm": 10,
    "held_out": true,
    "uncertainty_reported": true,
    "results_in_repo_files": true,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": true,
    "grade": "moderate",
    "reason": "Pre-registered, 10 seeds per arm in a committed raw_results.json, recomputed; one shuffle; readout trained on separate seeds 9000+."
   },
   "extraction": "files",
   "one_line": "Whole MaleCNS LIF aiming a crosshair: 412.5 px mean target distance vs 430.9 shuffled (p 0.66); both worse than a random walker (251.7).",
   "commit": "15d800f40c1cf8a69a5fb3d1aed18ddc2120c081",
   "checked_at": "2026-10-05",
   "sources": [
    {
     "url": "https://github.com/0Sakura721/flyaim/blob/15d800f40c1cf8a69a5fb3d1aed18ddc2120c081/flyaim/runs/20261003-195410-phase3/raw_results.json",
     "label": "direct"
    },
    {
     "url": "https://github.com/0Sakura721/flyaim/blob/15d800f40c1cf8a69a5fb3d1aed18ddc2120c081/flyaim/runs/20261003-195410-phase3/REPORT.md",
     "label": "direct"
    },
    {
     "url": "https://github.com/0Sakura721/flyaim/blob/15d800f40c1cf8a69a5fb3d1aed18ddc2120c081/README.md",
     "label": "direct"
    },
    {
     "url": "https://github.com/0Sakura721/flyaim/blob/15d800f40c1cf8a69a5fb3d1aed18ddc2120c081/flyaim/baselines/shuffle.py",
     "label": "direct"
    }
   ],
   "wiring_effect": "no-difference",
   "baseline_beaten": "no"
  },
  {
   "id": "flybrain-flappy",
   "name": "FlyBrain · Flappy (escape reflex)",
   "evidence_grade": "B",
   "made_by_us": false,
   "task_family": "game-play",
   "loop": "closed-loop",
   "wiring": {
    "dataset": "FlyWire FAFB",
    "release": "v783 (fafb_783_simple_edgelist and fafb_783_meta)",
    "scope": "whole brain: 144,837 neurons, 15,023,799 synapses; input LC4/LPLC2 looming drive, readout DNp01 (with DNp04) flap"
   },
   "trained_parts": "none in the escape-reflex demo; two hand-set gains (dors_scale 0.35, vent_gain 2.0) chosen by sweeps on the real wiring",
   "metric": {
    "name": "mean Flappy Bird score (pipes passed) over 40 games, offline bench scripts/flappy_bench.py, same parameters per graph",
    "unit": "score",
    "higher_is_better": true
   },
   "real": {
    "value": 100.8,
    "n": 40,
    "spread": null,
    "source": "7a8b601 README.md conclusion table: real '100.8' (old noise stream), 256 flaps"
   },
   "floor": {
    "label": "passive, no control (bird falls freely); 200 games, max_ticks 5000",
    "value": 0,
    "source": "7a8b601 README.md score-progression table: passive 0.00"
   },
   "arms": [
    {
     "arm_id": "shuffled",
     "null_type": "degree-preserving",
     "label": "targets of all edges shuffled globally (out-degree, in-degree, edge count and weight distribution kept; looming.variant shuffle_seed=7), same parameters",
     "value": 0,
     "n": 40,
     "samples": 1,
     "spread": null,
     "readout_refit": null,
     "source": "7a8b601 README.md conclusion table: shuffled '0.00', 0 flaps; src/fpv/looming.py variant(); scripts/flappy_bench.py shuffle_seed=7",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "cut",
     "null_type": "lesion",
     "label": "direct LC4/LPLC2 -> DNp01 edges cut",
     "value": 0,
     "n": 40,
     "samples": 1,
     "spread": null,
     "readout_refit": null,
     "source": "7a8b601 README.md conclusion table: cut '0.00', 0 flaps",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "passive",
     "null_type": "no-brain-baseline",
     "label": "no control, free fall (200 games, max_ticks 5000)",
     "value": 0,
     "n": 200,
     "samples": 1,
     "spread": null,
     "readout_refit": null,
     "source": "7a8b601 README.md score-progression table: passive 0.00",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "shuffled",
    "best_null_arm": "shuffled",
    "retention": 0,
    "retention_method": "floor",
    "baseline_arm": "passive",
    "baseline_retention": 0,
    "difference": 100.8,
    "ci": null,
    "p": null,
    "stat_note": "Control numbers only in README.md and ESCAPE.md; the bench writes output/*.json, which is not committed. Measured under an earlier shared noise stream; the current code gives statistically similar but not identical numbers (README). ESCAPE.md: all 472 descending neurons silent under the shuffled graph (truncated graph); DNp01 peaks near 4.1 Hz at full drive and needs about 29 Hz per LC4 cell."
   },
   "fairness": {
    "keeps_degree": true,
    "keeps_io_boundary": null,
    "tuned_equally": false,
    "null_samples": 1,
    "has_no_brain_baseline": true,
    "verdict": "unfair",
    "reason": "Shuffle keeps out- and in-degree and weights; one shuffle; the two input gains were swept on the real wiring and applied unchanged to the cut and shuffled graphs, so a gain re-tuned for them was not tried."
   },
   "quality": {
    "n_per_arm": 40,
    "held_out": null,
    "uncertainty_reported": false,
    "results_in_repo_files": false,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": false,
    "grade": "weak",
    "reason": "40 games per arm but one shuffle, no spread, numbers only in the README under an older noise stream; result files not committed."
   },
   "extraction": "files",
   "one_line": "Untrained FlyWire whole brain flaps from DNp01: mean score 100.8 over 40 games vs 0.00 and no flaps for shuffled or cut wiring; README only.",
   "commit": "7a8b6017e48d0485e6a032294c0591f8ede69722",
   "checked_at": "2026-10-06",
   "sources": [
    {
     "url": "https://github.com/programmingWTF/FlyBrain/blob/7a8b6017e48d0485e6a032294c0591f8ede69722/README.md",
     "label": "direct"
    },
    {
     "url": "https://github.com/programmingWTF/FlyBrain/blob/7a8b6017e48d0485e6a032294c0591f8ede69722/ESCAPE.md",
     "label": "direct"
    },
    {
     "url": "https://github.com/programmingWTF/FlyBrain/blob/7a8b6017e48d0485e6a032294c0591f8ede69722/src/fpv/looming.py",
     "label": "direct"
    },
    {
     "url": "https://github.com/programmingWTF/FlyBrain/blob/7a8b6017e48d0485e6a032294c0591f8ede69722/scripts/flappy_bench.py",
     "label": "direct"
    }
   ],
   "wiring_effect": "helps",
   "baseline_beaten": "yes"
  },
  {
   "id": "haltere",
   "name": "Haltere",
   "evidence_grade": "C",
   "made_by_us": false,
   "task_family": "game-play",
   "loop": "closed-loop",
   "wiring": {
    "dataset": "MaleCNS",
    "release": "v1.0",
    "scope": "30,000-neuron flight subgraph (2,767,698 edges)"
   },
   "trained_parts": "per-edge weight magnitudes, unknown signs, neuron parameters, encoders and readout (gradient descent + imitation of an MLP teacher)",
   "metric": {
    "name": "full three-lap race finishes in a frozen matched comparison",
    "unit": "fraction of races finished",
    "higher_is_better": true
   },
   "real": {
    "value": 0,
    "n": 2,
    "spread": null,
    "source": "3e3e7b5 docs/flight_cards/2026-09-23_matched_full_races.md:L3 \"Brain: 0/2 finishes\" and results table"
   },
   "floor": null,
   "arms": [
    {
     "arm_id": "pd-controller",
     "null_type": "no-brain-baseline",
     "label": "PD controller (brain in shadow), same race-cue pilot and settings",
     "value": 0.5,
     "n": 2,
     "samples": null,
     "spread": null,
     "readout_refit": false,
     "source": "3e3e7b5 docs/flight_cards/2026-09-23_matched_full_races.md:L3 \"PD: 1/2 finishes\"; table row Straw Bale 13:04.047",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": null,
    "best_null_arm": null,
    "retention": null,
    "retention_method": "none",
    "baseline_arm": "pd-controller",
    "baseline_retention": null,
    "difference": null,
    "ci": null,
    "p": null,
    "stat_note": "2 races per arm; author: \"They do not establish a statistical ranking\". Both crashed on Minus Two; PD alone finished Straw Bale."
   },
   "wiring_effect": "not-tested",
   "baseline_beaten": "no",
   "fairness": {
    "keeps_degree": null,
    "keeps_io_boundary": null,
    "tuned_equally": null,
    "null_samples": null,
    "has_no_brain_baseline": true,
    "verdict": "unclear",
    "reason": "No wiring null; PD baseline shares the race-cue pilot; n = 2 per arm. An MLP baseline exists (artifacts/mlp_baseline.pt, the imitation teacher) but no matched score was found."
   },
   "quality": {
    "n_per_arm": 2,
    "held_out": null,
    "uncertainty_reported": false,
    "results_in_repo_files": false,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": false,
    "grade": "weak",
    "reason": "Two races per arm, docs-only flight card, no uncertainty, not peer reviewed."
   },
   "extraction": "files",
   "one_line": "Connectome brain finished 0 of 2 frozen races; a plain PD controller finished 1 of 2. No wiring null.",
   "commit": "3e3e7b5e3aa0056caa6401abd89f7ec59508c57e",
   "checked_at": "2026-09-30T09:45:34Z",
   "sources": [
    {
     "url": "https://raw.githubusercontent.com/skulitom/haltere/3e3e7b5e3aa0056caa6401abd89f7ec59508c57e/docs/flight_cards/2026-09-23_matched_full_races.md",
     "label": "direct"
    }
   ]
  },
  {
   "id": "making-fly-play-chess",
   "name": "making-fly-play-chess",
   "evidence_grade": "C",
   "made_by_us": false,
   "task_family": "game-play",
   "loop": "closed-loop",
   "wiring": {
    "dataset": "FlyWire FAFB",
    "release": "v783 (Shiu et al. Connectivity_783)",
    "scope": "8,192-neuron breadth-first patch, 168,930 edges"
   },
   "trained_parts": "linear readout only (513 parameters; self-play TD-style updates)",
   "metric": {
    "name": "head-to-head match score (win=1, draw=0.5) against the paired opponent",
    "unit": "score (0-1)",
    "higher_is_better": true
   },
   "real": {
    "value": 0.506,
    "n": 5,
    "spread": "± 0.081 sd",
    "source": "4ddc9b5 notebooks/connectome_vs_classical_architectures.ipynb cell 22 output: vs rewired match_score 0.506 (80 games; per-seed rewired_score 0.469/0.500/0.406/0.625/0.531)"
   },
   "floor": null,
   "arms": [
    {
     "arm_id": "rewired",
     "null_type": "degree-preserving",
     "label": "rewired connections (rewire_reservoir: shuffled postsynaptic rows)",
     "value": 0.494,
     "n": 5,
     "samples": 5,
     "spread": "± 0.081 sd",
     "readout_refit": true,
     "source": "4ddc9b5 notebooks/connectome_vs_classical_architectures.ipynb cell 22 output: 1 - vs rewired match_score (16W/49D/15L of 80 from fly side)",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "classical",
     "null_type": "no-brain-baseline",
     "label": "classical features (same linear readout)",
     "value": 0.469,
     "n": 5,
     "samples": 5,
     "spread": null,
     "readout_refit": true,
     "source": "4ddc9b5 notebooks/connectome_vs_classical_architectures.ipynb cell 22 output: 1 - vs classical match_score 0.531",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "material",
     "null_type": "no-brain-baseline",
     "label": "material-count player (untrained heuristic)",
     "value": 0.694,
     "n": 5,
     "samples": 5,
     "spread": null,
     "readout_refit": null,
     "source": "4ddc9b5 notebooks/connectome_vs_classical_architectures.ipynb cell 22 output: 1 - vs material match_score 0.306",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "rewired",
    "best_null_arm": "rewired",
    "retention": 0.9763,
    "retention_method": "ratio",
    "baseline_arm": "material",
    "baseline_retention": 1.3715,
    "difference": 0.012,
    "ci": null,
    "p": null,
    "stat_note": "CI is the notebook cell 24 t-interval for fly score vs rewired (0.506, seed-level, 5 seeds; bootstrap 0.444-0.575), spanning 0.5. Arm values are opponents' head-to-head scores against the fly (1 - fly score), so real+arm=1 per pairing. Quick mode flagged by author as code check only. CI [0.4058, 0.6067] is for the real arm's score, not for the difference."
   },
   "wiring_effect": "no-difference",
   "baseline_beaten": "no",
   "fairness": {
    "keeps_degree": true,
    "keeps_io_boundary": true,
    "tuned_equally": true,
    "null_samples": 5,
    "has_no_brain_baseline": true,
    "verdict": "fair",
    "reason": "Same patch, same readout, same self-play budget per seed; rewire permutes edge targets (keeps in/out degree counts and weights). Quick mode: only 30 self-play games and 16 games per pairing per seed."
   },
   "quality": {
    "n_per_arm": 5,
    "held_out": true,
    "uncertainty_reported": true,
    "results_in_repo_files": true,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": true,
    "grade": "strong",
    "reason": "5 seeds with CI in stored notebook output, but author-labelled quick code-check run with tiny game counts; results only as notebook output."
   },
   "extraction": "files",
   "one_line": "FlyWire-patch chess agent ties its rewired twin (0.506, CI spans 0.5) and loses to a material counter; author-labelled code-check run.",
   "commit": "4ddc9b5df19bbe6703b20d80abfd279f973f9350",
   "checked_at": "2026-09-30T09:42:55Z",
   "sources": [
    {
     "url": "https://github.com/mncrftfrcnm/making-fly-play-chess",
     "label": "direct"
    },
    {
     "url": "https://raw.githubusercontent.com/mncrftfrcnm/making-fly-play-chess/4ddc9b5df19bbe6703b20d80abfd279f973f9350/notebooks/connectome_vs_classical_architectures.ipynb",
     "label": "direct"
    }
   ]
  },
  {
   "id": "fly-cartpole",
   "name": "fly-cartpole",
   "evidence_grade": "A",
   "made_by_us": false,
   "task_family": "ml-benchmark",
   "loop": "closed-loop",
   "wiring": {
    "dataset": "MaleCNS",
    "release": "v1.0",
    "scope": "flight-stabilisation circuit: 5,459 neurons (ocellar, haltere and HS sensory groups to wing-amplitude motor neurons), linear rate units"
   },
   "trained_parts": "Three sensor gains (five with the landmark) by reward-modulated random search; the circuit's couplings never change. Learning rates chosen by the author's reflex-tune search.",
   "metric": {
    "name": "CartPole mean episode length over the final 100 episodes (fresh seeds 160-179)",
    "unit": "steps",
    "higher_is_better": true
   },
   "real": {
    "value": 499.9,
    "n": 20,
    "spread": "± 0.3 sd",
    "source": "a8c6257 results/linear-retuned/summary.md: fly-reflex-adaptive row (same numbers in results/reflex/summary.md, 930ab65)"
   },
   "floor": {
    "label": "random pushes (chance)",
    "value": 22.1,
    "source": "a8c6257 results/linear-retuned/summary.md: random row"
   },
   "arms": [
    {
     "arm_id": "shuffled",
     "null_type": "degree-preserving",
     "label": "fly-reflex-adaptive-shuffled (directed double-edge swaps; couplings travel with their edges; sensor and motor identities kept)",
     "value": 165.9,
     "n": 20,
     "samples": 20,
     "spread": "± 182.8 sd",
     "readout_refit": true,
     "source": "a8c6257 results/linear-retuned/summary.md: fly-reflex-adaptive-shuffled row; reflex.py shuffle_flight, one shuffle per evaluation seed (reflex_report.py line 196)",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "untuned",
     "null_type": "untrained",
     "label": "fly-reflex (wiring only, no tuning)",
     "value": 275.6,
     "n": 20,
     "samples": 1,
     "spread": "± 9.4 sd",
     "readout_refit": false,
     "source": "a8c6257 results/linear-retuned/summary.md: fly-reflex row",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "linear",
     "null_type": "no-graph-model",
     "label": "the circuit replaced by its own linear map, tuned the same way (seeds 280-299)",
     "value": 499.6,
     "n": 20,
     "samples": 1,
     "spread": "± 0.7 sd",
     "readout_refit": true,
     "source": "a8c6257 results/linear-retuned/summary.md: linear-adaptive row; author's two-sided permutation p = 0.1797 against the circuit",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "lqr",
     "null_type": "no-graph-model",
     "label": "LQR designed from CartPole's equations, no learning (seeds 300-319)",
     "value": 500,
     "n": 20,
     "samples": 1,
     "spread": "± 0.0 sd",
     "readout_refit": null,
     "source": "a8c6257 results/linear-retuned/summary.md: lqr row",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "random",
     "null_type": "no-brain-baseline",
     "label": "uniform random pushes",
     "value": 22.1,
     "n": 20,
     "samples": 1,
     "spread": "± 0.8 sd",
     "readout_refit": null,
     "source": "a8c6257 results/linear-retuned/summary.md: random row",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "shuffled",
    "best_null_arm": "shuffled",
    "retention": 0.301,
    "retention_method": "floor",
    "baseline_arm": "lqr",
    "baseline_retention": 1.0002,
    "difference": 334,
    "ci": null,
    "p": 0.0001,
    "stat_note": "Re-read on 2026-10-03 at a8c6257 (re-grade, material): the author's headline experiment is now the MaleCNS flight-stabilisation circuit (README 'Results'). Author's permutation test, fly-reflex-adaptive > fly-reflex-adaptive-shuffled, p = 0.0001 (fresh seeds 160-179). The circuit's own linear map tuned the same way scores 499.6 (p = 0.18, two-sided) and LQR 500, so balancing alone needs no more than a small linear controller (author's words); the circuit beats its linear map only on holding station past the 500-step cap (results/lqr-station/summary.md). The real arm sits at the 500 ceiling. The previous study (both mushroom bodies, fly-bilateral 392.5 vs degree-preserving shuffled 390.3, p = 0.4771, results/summary.md) is unchanged in the repository but is no longer the headline."
   },
   "wiring_effect": "helps",
   "baseline_beaten": "no",
   "fairness": {
    "keeps_degree": true,
    "keeps_io_boundary": true,
    "tuned_equally": true,
    "null_samples": 20,
    "has_no_brain_baseline": true,
    "verdict": "fair",
    "reason": "Degree-preserving directed double-edge swaps with couplings travelling, sensor and motor identities kept, the same self-tuning of sensor gains and rates, 20 fresh seeds with one shuffle each; a random-push baseline and an equally tuned linear-map baseline."
   },
   "quality": {
    "n_per_arm": 20,
    "held_out": true,
    "uncertainty_reported": true,
    "results_in_repo_files": true,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": false,
    "grade": "strong",
    "reason": "Fresh seeds, the same tuning for every arm, author's permutation tests and per-seed values committed; not recomputed by us this run; the real arm is at the 500-step ceiling."
   },
   "extraction": "files",
   "one_line": "Fly flight circuit balances CartPole (499.9 steps) and shuffled wiring drops it to 165.9 (p = 0.0001), but its own linear map also reaches 499.6.",
   "commit": "a8c6257",
   "checked_at": "2026-10-03",
   "sources": [
    {
     "url": "https://github.com/Curt-Park/fly-cartpole/blob/a8c62579bb4566baa517e55c63d53f85344018d4/results/linear-retuned/summary.md",
     "label": "direct"
    },
    {
     "url": "https://github.com/Curt-Park/fly-cartpole/blob/a8c62579bb4566baa517e55c63d53f85344018d4/results/reflex/summary.md",
     "label": "direct"
    },
    {
     "url": "https://github.com/Curt-Park/fly-cartpole",
     "label": "direct"
    }
   ]
  },
  {
   "id": "flybench",
   "name": "flybench",
   "evidence_grade": "A",
   "made_by_us": false,
   "task_family": "ml-benchmark",
   "loop": "open-loop",
   "wiring": {
    "dataset": "FlyWire FAFB",
    "release": "v783",
    "scope": "whole brain (139,255 neurons)"
   },
   "trained_parts": "none in the brain (reference LIF, global gain 0.45)",
   "metric": {
    "name": "mean task score over 28 circuit tasks (3 seeds per task)",
    "unit": "score 0-1",
    "higher_is_better": true
   },
   "real": {
    "value": 0.5991,
    "n": 28,
    "spread": "± 0.3264 sd across tasks",
    "source": "3052ce5 results/flywire783_gain0.45.json:score (= mean of tasks[].score)"
   },
   "floor": null,
   "arms": [
    {
     "arm_id": "rewired",
     "null_type": "degree-preserving",
     "label": "rewired: postsynaptic endpoints of all edges permuted (out-degree, weights, signs kept; in-degree multiset kept up to merged duplicates)",
     "value": 0.3993,
     "n": 28,
     "samples": 1,
     "spread": "± 0.3006 sd across tasks",
     "readout_refit": false,
     "source": "3052ce5 results/flywire783_gain0.45.json:tasks[].controls.rewired.score (mean recomputed by us); specificity key 0.1998",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "rewired",
    "best_null_arm": "rewired",
    "retention": 0.6665,
    "retention_method": "ratio",
    "baseline_arm": null,
    "baseline_retention": null,
    "difference": 0.1998,
    "ci": null,
    "p": null,
    "stat_note": "Difference = author \"specificity\" 0.1998 = mean over tasks of score(real) - score(rewired); per-task difference sd 0.328 (recomputed). One rewired connectome (seed = params.seed). Other result files: shiu2024.json 0.5547 vs 0.3746 (spec 0.180); malecns-gain-0.65.json 0.5779 vs 0.3585 over 34 tasks (spec 0.219)."
   },
   "wiring_effect": "mixed",
   "baseline_beaten": "not-tested",
   "fairness": {
    "keeps_degree": true,
    "keeps_io_boundary": false,
    "tuned_equally": true,
    "null_samples": 1,
    "has_no_brain_baseline": false,
    "verdict": "partly-fair",
    "reason": "Same untrained LIF and parameters on the rewired graph; degree-preserving, but only one rewired realisation and the sensory/motor boundary is rewired too. Gain was chosen on the real graph."
   },
   "quality": {
    "n_per_arm": 28,
    "held_out": null,
    "uncertainty_reported": false,
    "results_in_repo_files": true,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": true,
    "grade": "weak",
    "reason": "Committed JSON with per-task control scores, recomputed by us; bootstrap CI only for the graded score; single shuffle realisation; not peer reviewed."
   },
   "extraction": "files",
   "one_line": "Benchmark of 28 circuit tasks: real FlyWire scores 0.60 vs 0.40 on one degree-preserving rewiring (specificity +0.20).",
   "commit": "3052ce5fa9ab8cc0a9e101e8cd8f5bb520a80ff5",
   "checked_at": "2026-09-30T09:45:34Z",
   "sources": [
    {
     "url": "https://github.com/brandoncho369/flybench/blob/3052ce5fa9ab8cc0a9e101e8cd8f5bb520a80ff5/results/flywire783_gain0.45.json",
     "label": "direct"
    },
    {
     "url": "https://github.com/brandoncho369/flybench/blob/3052ce5fa9ab8cc0a9e101e8cd8f5bb520a80ff5/flybench/controls.py#L59-L78",
     "label": "direct"
    }
   ]
  },
  {
   "id": "flys-hash-function",
   "name": "The Fly's Hash Function",
   "evidence_grade": "A",
   "made_by_us": false,
   "task_family": "ml-benchmark",
   "loop": "offline",
   "wiring": {
    "dataset": "hemibrain",
    "release": "v1.2",
    "scope": "130 PNs x 1,745 KCs (PN->KC projection)"
   },
   "trained_parts": "none (fixed projection + top-5% winner-take-all)",
   "metric": {
    "name": "MNIST nearest-neighbour precision@16",
    "unit": "fraction (0-1)",
    "higher_is_better": true
   },
   "real": {
    "value": 0.4993,
    "n": 5,
    "spread": "± 0.0039 sem",
    "source": "271cc85 out/benchmark.json: mnist.real[\"16\"].mean/sem"
   },
   "floor": null,
   "arms": [
    {
     "arm_id": "degree_preserving",
     "null_type": "degree-preserving",
     "label": "degree_preserving (Maslov-Sneppen)",
     "value": 0.552,
     "n": 5,
     "samples": 5,
     "spread": "± 0.0020 sem",
     "readout_refit": null,
     "source": "271cc85 out/benchmark.json: mnist.degree_preserving[\"16\"].mean/sem",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "glomerulus_matched",
     "null_type": "boundary-preserving",
     "label": "glomerulus_matched",
     "value": 0.555,
     "n": 5,
     "samples": 5,
     "spread": "± 0.0022 sem",
     "readout_refit": null,
     "source": "271cc85 out/benchmark.json: mnist.glomerulus_matched[\"16\"].mean/sem",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "weight_shuffled",
     "null_type": "weight-shuffle",
     "label": "weight_shuffled",
     "value": 0.517,
     "n": 5,
     "samples": 5,
     "spread": "± 0.0034 sem",
     "readout_refit": null,
     "source": "271cc85 out/benchmark.json: mnist.weight_shuffled[\"16\"].mean/sem",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "random_flyhash",
     "null_type": "random-graph",
     "label": "random_flyhash",
     "value": 0.6212,
     "n": 5,
     "samples": 5,
     "spread": "± 0.0010 sem",
     "readout_refit": null,
     "source": "271cc85 out/benchmark.json: mnist.random_flyhash[\"16\"].mean/sem",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "real_binary",
     "null_type": "other",
     "label": "real_binary (real edges, weights set to 1)",
     "value": 0.5472,
     "n": 5,
     "samples": 5,
     "spread": "± 0.0028 sem",
     "readout_refit": null,
     "source": "271cc85 out/benchmark.json: mnist.real_binary[\"16\"].mean/sem",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "simhash",
     "null_type": "no-graph-model",
     "label": "simhash",
     "value": 0.2414,
     "n": 5,
     "samples": 5,
     "spread": "± 0.0024 sem",
     "readout_refit": null,
     "source": "271cc85 out/benchmark.json: mnist.simhash[\"16\"].mean/sem",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "degree_preserving",
    "best_null_arm": "random_flyhash",
    "retention": 1.1055,
    "retention_method": "ratio",
    "baseline_arm": "simhash",
    "baseline_retention": 0.4835,
    "difference": -0.0527,
    "ci": [
     -0.061,
     -0.044
    ],
    "p": null,
    "stat_note": "CI recomputed by us from stored sems (independent-arm normal approx, z=-12); seeds are paired so this is conservative. Fashion-MNIST 0.480 vs 0.524, odour-correlated 0.368 vs 0.427 same direction. Only in the connectome-own synthetic world (out/specialization.json) does real (0.219) beat random (0.155)."
   },
   "wiring_effect": "worse",
   "baseline_beaten": "yes",
   "fairness": {
    "keeps_degree": true,
    "keeps_io_boundary": true,
    "tuned_equally": true,
    "null_samples": 5,
    "has_no_brain_baseline": true,
    "verdict": "fair",
    "reason": "Shared encoder, same top-5% KC code and tie-break seed for every condition; stochastic nulls re-drawn per seed (benchmark.py:171). Nothing is trained, so tuning is equal."
   },
   "quality": {
    "n_per_arm": 5,
    "held_out": true,
    "uncertainty_reported": true,
    "results_in_repo_files": true,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": true,
    "grade": "strong",
    "reason": "5 seeds, sem stored in committed out/benchmark.json, values match README; untrained so no tuning bias."
   },
   "extraction": "files",
   "one_line": "Real hemibrain PN-to-KC wiring hashes worse than degree-preserving and random FlyHash nulls on MNIST, Fashion-MNIST and correlated odours; beats SimHash.",
   "commit": "271cc854dcd2c517a41074c493a504681c12df8a",
   "checked_at": "2026-09-30T09:39:26Z",
   "sources": [
    {
     "url": "https://github.com/realgauravvyas/flys-hash-function",
     "label": "direct"
    },
    {
     "url": "https://raw.githubusercontent.com/realgauravvyas/flys-hash-function/271cc854dcd2c517a41074c493a504681c12df8a/out/benchmark.json",
     "label": "direct"
    }
   ]
  },
  {
   "id": "larva-vs-shuffles",
   "name": "Does the larval connectome beat its own shuffles? (connectome-null-models)",
   "evidence_grade": "A",
   "made_by_us": false,
   "task_family": "ml-benchmark",
   "loop": "offline",
   "wiring": {
    "dataset": "Larval connectome (Winding et al. 2023)",
    "release": "Winding 2023 ad edges",
    "scope": "whole larval brain, 2,956 neurons / 63,545 edges"
   },
   "trained_parts": "input projection and linear readout only; recurrent matrix frozen",
   "metric": {
    "name": "CIFAR-10 test accuracy (final epoch)",
    "unit": "percent",
    "higher_is_better": true
   },
   "real": {
    "value": 43.04,
    "n": 15,
    "spread": "± 0.40 sd",
    "source": "be301d9 results/results_cifar10.json: condition=connectome final_acc, 15 seeds"
   },
   "floor": {
    "label": "no recurrence (chance)",
    "value": 10,
    "source": "be301d9 results/results_cifar10.json: condition=no_recurrence final_acc, 5 seeds"
   },
   "arms": [
    {
     "arm_id": "degree_preserving",
     "null_type": "degree-preserving",
     "label": "degree_preserving",
     "value": 42.95,
     "n": 15,
     "samples": 15,
     "spread": "± 0.37 sd",
     "readout_refit": true,
     "source": "be301d9 results/results_cifar10.json: condition=degree_preserving final_acc",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "weight_shuffle",
     "null_type": "weight-shuffle",
     "label": "weight_shuffle",
     "value": 42.96,
     "n": 5,
     "samples": 5,
     "spread": "± 0.62 sd",
     "readout_refit": true,
     "source": "be301d9 results/results_cifar10.json: condition=weight_shuffle final_acc",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "erdos_renyi",
     "null_type": "random-graph",
     "label": "erdos_renyi",
     "value": 42.98,
     "n": 5,
     "samples": 5,
     "spread": "± 0.12 sd",
     "readout_refit": true,
     "source": "be301d9 results/results_cifar10.json: condition=erdos_renyi final_acc",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "no_recurrence",
     "null_type": "no-edges",
     "label": "no_recurrence",
     "value": 10,
     "n": 5,
     "samples": 5,
     "spread": "± 0.00 sd",
     "readout_refit": true,
     "source": "be301d9 results/results_cifar10.json: condition=no_recurrence final_acc",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "degree_preserving",
    "best_null_arm": "erdos_renyi",
    "retention": 0.9973,
    "retention_method": "floor",
    "baseline_arm": null,
    "baseline_retention": null,
    "difference": 0.09,
    "ci": [
     -0.2,
     0.38
    ],
    "p": 0.51,
    "stat_note": "Welch t=0.66 (df~28) recomputed by us from per-seed final_acc; CI approximate (t~2.05); p from catalogue text, consistent with t. MNIST (results/results.json, n=5): real 97.30 vs degree-preserving 97.28, ER 97.53, weight shuffle 97.38, no recurrence 11.35."
   },
   "wiring_effect": "no-difference",
   "baseline_beaten": "not-tested",
   "fairness": {
    "keeps_degree": true,
    "keeps_io_boundary": true,
    "tuned_equally": true,
    "null_samples": 15,
    "has_no_brain_baseline": false,
    "verdict": "fair",
    "reason": "Same Dale signs, spectral radius, torch seed and training for every condition; degree-preserving swap with 15 distinct shuffles; sensory/descending I/O kept. No wiring-free model baseline beyond no-recurrence."
   },
   "quality": {
    "n_per_arm": 15,
    "held_out": true,
    "uncertainty_reported": true,
    "results_in_repo_files": true,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": true,
    "grade": "strong",
    "reason": "15 seeds per main arm, per-seed results committed, spread recomputed by us; not peer reviewed."
   },
   "extraction": "files",
   "one_line": "Larval connectome reservoir ties its degree-preserving shuffles, Erdos-Renyi and weight shuffles on CIFAR-10 and MNIST; only removing recurrence hurts.",
   "commit": "be301d9c2c5fe481c68db174cbed40f1fb2881c9",
   "checked_at": "2026-09-30T09:38:58Z",
   "sources": [
    {
     "url": "https://github.com/cqw-acq/connectome-null-models",
     "label": "direct"
    },
    {
     "url": "https://raw.githubusercontent.com/cqw-acq/connectome-null-models/be301d9c2c5fe481c68db174cbed40f1fb2881c9/results/results_cifar10.json",
     "label": "direct"
    }
   ]
  },
  {
   "id": "neuroweave",
   "name": "NeuroWeave",
   "evidence_grade": "C",
   "made_by_us": false,
   "task_family": "ml-benchmark",
   "loop": "offline",
   "wiring": {
    "dataset": "MaleCNS",
    "release": "v1.0 (author parquet)",
    "scope": "150 neurons (first 150 rows of the neuron table, not a named circuit)"
   },
   "trained_parts": "edge scales, biases, input projection and readout (mask fixes which edges exist)",
   "metric": {
    "name": "T-001 static classification test accuracy",
    "unit": "percent",
    "higher_is_better": true
   },
   "real": {
    "value": 98,
    "n": 1,
    "spread": null,
    "source": "518f162 artifacts/t001_topology_results.json: A1-BIO.test_acc"
   },
   "floor": null,
   "arms": [
    {
     "arm_id": "A3-CONFIG",
     "null_type": "degree-preserving",
     "label": "A3-CONFIG (configuration model)",
     "value": 99,
     "n": 1,
     "samples": 1,
     "spread": null,
     "readout_refit": true,
     "source": "518f162 artifacts/t001_topology_results.json: A3-CONFIG.test_acc",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "A3-ER",
     "null_type": "random-graph",
     "label": "A3-ER (same-density Erdos-Renyi)",
     "value": 100,
     "n": 1,
     "samples": 1,
     "spread": null,
     "readout_refit": true,
     "source": "518f162 artifacts/t001_topology_results.json: A3-ER.test_acc",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "A3-DENSE",
     "null_type": "other",
     "label": "A3-DENSE (all 22,500 edges)",
     "value": 100,
     "n": 1,
     "samples": 1,
     "spread": null,
     "readout_refit": true,
     "source": "518f162 artifacts/t001_topology_results.json: A3-DENSE.test_acc",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "A3-CONFIG",
    "best_null_arm": "A3-ER",
    "retention": 1.0102,
    "retention_method": "ratio",
    "baseline_arm": null,
    "baseline_retention": null,
    "difference": -1,
    "ci": null,
    "p": null,
    "stat_note": "Single seed, near ceiling. T-002 delayed recall (artifacts/confirmation_t002/*.json, 5 seeds): A1-BIO 94.8 ± 4.4 sd, A1-FROZEN 92.4 ± 3.9, LSTM 57.6 ± 19.3, A0-Random 50.6 ± 4.8; no random-graph arm on T-002."
   },
   "wiring_effect": "no-difference",
   "baseline_beaten": "not-tested",
   "fairness": {
    "keeps_degree": true,
    "keeps_io_boundary": null,
    "tuned_equally": true,
    "null_samples": 1,
    "has_no_brain_baseline": false,
    "verdict": "partly-fair",
    "reason": "Same training script for all masks, but one seed and one null graph each; stored edge counts differ (A1-BIO 3,029 vs A3-CONFIG 1,681), so the degree match is doubtful; ER has more params (16,410 vs 13,714)."
   },
   "quality": {
    "n_per_arm": 1,
    "held_out": true,
    "uncertainty_reported": false,
    "results_in_repo_files": true,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": false,
    "grade": "weak",
    "reason": "Single seed, ceiling task, no uncertainty; README figures disagree with stored files; data package missing from repo."
   },
   "extraction": "files",
   "one_line": "150-neuron MaleCNS mask scores 98% versus 99-100% for configuration-model, random and dense masks on a ceiling task; single seed.",
   "commit": "518f162ef7a68022d3dfa9466d3b8b73dd92652c",
   "checked_at": "2026-09-30T09:41:09Z",
   "sources": [
    {
     "url": "https://github.com/Titanium-xd/neuroweave",
     "label": "direct"
    },
    {
     "url": "https://raw.githubusercontent.com/Titanium-xd/neuroweave/518f162ef7a68022d3dfa9466d3b8b73dd92652c/artifacts/t001_topology_results.json",
     "label": "direct"
    }
   ]
  },
  {
   "id": "flybrain-reservoir",
   "name": "flybrain-reservoir",
   "evidence_grade": "A",
   "made_by_us": false,
   "task_family": "reservoir",
   "loop": "offline",
   "wiring": {
    "dataset": "MaleCNS",
    "release": "v1.0 (also FlyWire 783 replication)",
    "scope": "whole CNS, 166,700 neurons"
   },
   "trained_parts": "linear ridge readout only; recurrent weights fixed, spectral radius 0.9",
   "metric": {
    "name": "memory capacity with readout noise, standard gain",
    "unit": "capacity (summed r^2 over delays)",
    "higher_is_better": true
   },
   "real": {
    "value": 2.2,
    "n": 10,
    "spread": null,
    "source": "6da6325 README.md:L96 (whole CNS, readout noise, connectome)"
   },
   "floor": null,
   "arms": [
    {
     "arm_id": "degree_preserving",
     "null_type": "degree-preserving",
     "label": "degree-preserving",
     "value": 15.5,
     "n": 10,
     "samples": 10,
     "spread": null,
     "readout_refit": true,
     "source": "6da6325 README.md:L96",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "weight_shuffle",
     "null_type": "weight-shuffle",
     "label": "weight shuffle",
     "value": 2.4,
     "n": 10,
     "samples": 10,
     "spread": null,
     "readout_refit": true,
     "source": "6da6325 README.md:L96",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "sign_shuffle",
     "null_type": "sign-shuffle",
     "label": "sign shuffle",
     "value": 2.9,
     "n": 10,
     "samples": 10,
     "spread": null,
     "readout_refit": true,
     "source": "6da6325 README.md:L96",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "erdos_renyi",
     "null_type": "random-graph",
     "label": "Erdős–Rényi",
     "value": 7.4,
     "n": 10,
     "samples": 10,
     "spread": null,
     "readout_refit": true,
     "source": "6da6325 README.md:L96",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "degree_preserving",
    "best_null_arm": "degree_preserving",
    "retention": 7.0455,
    "retention_method": "ratio",
    "baseline_arm": null,
    "baseline_retention": null,
    "difference": -13.3,
    "ci": null,
    "p": null,
    "stat_note": "No per-seed spread in README for this table. Other tasks agree: best-gain whole CNS 2.9 vs 22.8 deg-pres / 30.8 ER (README.md:L145); NARMA-10 NRMSE 0.77 vs 0.42 / 0.48 (L276); FlyWire 783 whole brain 3.0 vs 21.8 / 72.4 (L354); 3,000-neuron circuit best gain 6.7 vs 6.0-7.4 tie (L144); market Sharpe differences all CIs include zero (L79-85)."
   },
   "wiring_effect": "worse",
   "baseline_beaten": "not-tested",
   "fairness": {
    "keeps_degree": true,
    "keeps_io_boundary": true,
    "tuned_equally": true,
    "null_samples": 10,
    "has_no_brain_baseline": false,
    "verdict": "fair",
    "reason": "All wirings rescaled to spectral radius 0.9, same input weights, bias and readout neurons per seed (README.md:L518-519), plus a per-wiring gain sweep; 10 seeds for whole CNS. Linear HAR baseline exists only for volatility task."
   },
   "quality": {
    "n_per_arm": 10,
    "held_out": true,
    "uncertainty_reported": false,
    "results_in_repo_files": false,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": false,
    "grade": "weak",
    "reason": "Numbers only in README tables; result folders are not committed; no spread for headline table. Design itself is careful (paired seeds, gain sweep)."
   },
   "extraction": "files",
   "one_line": "Whole fly CNS reservoir remembers far less than its degree-preserving shuffle (2.2 vs 15.5); spectral-radius scaling around an antennal-lobe hot spot explains it.",
   "commit": "6da63257154a334006579c692a31f671e572f83e",
   "checked_at": "2026-09-30T09:39:26Z",
   "sources": [
    {
     "url": "https://github.com/MilanKalajdzic/flybrain-reservoir",
     "label": "direct"
    },
    {
     "url": "https://raw.githubusercontent.com/MilanKalajdzic/flybrain-reservoir/6da63257154a334006579c692a31f671e572f83e/README.md",
     "label": "direct"
    }
   ]
  },
  {
   "id": "flm",
   "name": "FLM - Fly Language Model",
   "evidence_grade": "A",
   "made_by_us": false,
   "task_family": "language-model",
   "loop": "offline",
   "wiring": {
    "dataset": "MaleCNS",
    "release": "v1.0",
    "scope": "whole CNS, 166,700 nodes / 25,582,938 edges"
   },
   "trained_parts": "278,528-parameter readout adapter only; LLM backbone and graph frozen",
   "metric": {
    "name": "confirmation-set next-token NLL (32 conversations, 1,236 target tokens)",
    "unit": "nats/token",
    "higher_is_better": false
   },
   "real": {
    "value": 1.359816,
    "n": 3,
    "spread": "± 0.000110 sd",
    "source": "paper (artificialscientific.com/papers/flies-are-all-you-need) Table 1: Fly readout"
   },
   "floor": {
    "label": "no edges (= frozen backbone)",
    "value": 1.381995,
    "source": "paper (artificialscientific.com/papers/flies-are-all-you-need) Table 1: No edges / Frozen backbone"
   },
   "arms": [
    {
     "arm_id": "relabeled_no_refit",
     "null_type": "other",
     "label": "Relabeled, without refitting (node relabelling of the graph)",
     "value": 1.381265,
     "n": 3,
     "samples": 3,
     "spread": "± 0.000802 sd",
     "readout_refit": false,
     "source": "paper (artificialscientific.com/papers/flies-are-all-you-need) Table 1: Relabeled, without refitting",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "no_edges",
     "null_type": "no-edges",
     "label": "No edges",
     "value": 1.381995,
     "n": 3,
     "samples": 1,
     "spread": "± 0 sd",
     "readout_refit": false,
     "source": "paper (artificialscientific.com/papers/flies-are-all-you-need) Table 1: No edges",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "direct_input",
     "null_type": "no-graph-model",
     "label": "parameter-matched direct-input readout",
     "value": 1.359328,
     "n": 3,
     "samples": 3,
     "spread": "± 0.000108 sd",
     "readout_refit": true,
     "source": "paper (artificialscientific.com/papers/flies-are-all-you-need) Table 1: Direct-input readout",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": null,
    "best_null_arm": null,
    "retention": null,
    "retention_method": "none",
    "baseline_arm": "direct_input",
    "baseline_retention": 1.022,
    "difference": null,
    "ci": [
     0.00000502,
     0.00104
    ],
    "p": null,
    "stat_note": "CI is paper descriptive 95% conversation-bootstrap for fly minus direct-input (+0.000488 nats/token; all 3 paired fits favour direct-input). Fly minus frozen -0.0222, CI -0.0264 to -0.0184. Relabelled graph without refit ~ floor (retention 0.03) but readout was not refit, so it is not a fair wiring null."
   },
   "wiring_effect": "not-tested",
   "baseline_beaten": "no",
   "fairness": {
    "keeps_degree": null,
    "keeps_io_boundary": false,
    "tuned_equally": false,
    "null_samples": 3,
    "has_no_brain_baseline": true,
    "verdict": "partly-fair",
    "reason": "Direct-input control is parameter-matched and fit the same way (fair baseline). The only graph null is a node relabelling evaluated without refitting the readout, not a random-graph control (flm/graph.py docstring)."
   },
   "quality": {
    "n_per_arm": 3,
    "held_out": true,
    "uncertainty_reported": true,
    "results_in_repo_files": false,
    "independent_test": false,
    "peer_review": "preprint",
    "recomputed_by_us": false,
    "grade": "weak",
    "reason": "3 fit seeds, frozen confirmation holdout, bootstrap CI; numbers only in the preprint web page, not repo files."
   },
   "extraction": "paper",
   "one_line": "Fly-graph residual lowers LLM loss slightly, but a parameter-matched direct-input adapter does as well or better; no refit wiring null.",
   "commit": "7251a8921db4f891c39bd75ee5ad827f7031a24b",
   "checked_at": "2026-09-30T09:40:47Z",
   "sources": [
    {
     "url": "https://artificialscientific.com/papers/flies-are-all-you-need",
     "label": "direct"
    },
    {
     "url": "https://github.com/nftechie/flm",
     "label": "direct"
    }
   ]
  },
  {
   "id": "bioreservoir",
   "name": "BioReservoir",
   "evidence_grade": "B",
   "made_by_us": false,
   "task_family": "forecasting",
   "loop": "offline",
   "wiring": {
    "dataset": "MaleCNS",
    "release": "v1.0 (min_syn 5); BANC 626 secondary",
    "scope": "whole CNS, 165,122 neurons / 6,235,682 connections"
   },
   "trained_parts": "none in the brain; pretrained MiniLM text encoder with fixed seeded random projection",
   "metric": {
    "name": "mean forecast p(yes) over 33 questions x 3 phrasings (no outcomes yet, so no accuracy)",
    "unit": "probability",
    "higher_is_better": null
   },
   "real": {
    "value": 0.4996,
    "n": 99,
    "spread": "± 0.0028 sd",
    "source": "ac0253e experiments/001-fly-oracle/results.csv: condition=real, p_yes over runs"
   },
   "floor": {
    "label": "uninformative answer (0.5)",
    "value": 0.5,
    "source": "definition; no outcome-based chance level exists yet"
   },
   "arms": [
    {
     "arm_id": "rewired",
     "null_type": "degree-preserving",
     "label": "rewired (degree-preserving rewire)",
     "value": 0.5009,
     "n": 99,
     "samples": null,
     "spread": "± 0.0087 sd",
     "readout_refit": null,
     "source": "ac0253e experiments/001-fly-oracle/results.csv: condition=rewired, p_yes over runs",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "er",
     "null_type": "random-graph",
     "label": "er (Erdős–Rényi, same size and density)",
     "value": 0.5027,
     "n": 99,
     "samples": null,
     "spread": "± 0.0557 sd",
     "readout_refit": null,
     "source": "ac0253e experiments/001-fly-oracle/results.csv: condition=er, p_yes over runs",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "no_brain",
     "null_type": "no-brain-baseline",
     "label": "no_brain (same encoder and readout, no connectome)",
     "value": 0.4999,
     "n": 99,
     "samples": 1,
     "spread": "± 0.0039 sd",
     "readout_refit": null,
     "source": "ac0253e experiments/001-fly-oracle/results.csv: condition=no_brain, p_yes over runs",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "coin",
     "null_type": "other",
     "label": "coin (seeded pseudo-random draw)",
     "value": 0.5758,
     "n": 33,
     "samples": 1,
     "spread": "± 0.5019 sd",
     "readout_refit": null,
     "source": "ac0253e experiments/001-fly-oracle/results.csv: condition=coin, p_yes over runs",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "rewired",
    "best_null_arm": null,
    "retention": null,
    "retention_method": "none",
    "baseline_arm": null,
    "baseline_retention": null,
    "difference": -0.0013,
    "ci": null,
    "p": null,
    "stat_note": "No performance metric yet: questions resolve by 2026-12-15 (RESULTS.md:L49-52). Mean |p-0.5|: real 0.0022, no_brain 0.0031, rewired 0.0073, ER 0.0470; the real brain is the least decisive. Recomputed by us."
   },
   "wiring_effect": "not-tested",
   "baseline_beaten": "not-tested",
   "fairness": {
    "keeps_degree": true,
    "keeps_io_boundary": true,
    "tuned_equally": true,
    "null_samples": null,
    "has_no_brain_baseline": true,
    "verdict": "partly-fair",
    "reason": "Rewire keeps neurons, degrees, weights, signs and cell types, with same encoder/readout; nothing is trained. But no outcome score exists, so fairness of the comparison cannot yet be judged."
   },
   "quality": {
    "n_per_arm": 99,
    "held_out": true,
    "uncertainty_reported": true,
    "results_in_repo_files": true,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": true,
    "grade": "moderate",
    "reason": "Per-run values committed, but no accuracy or Brier score until questions resolve; only process numbers."
   },
   "extraction": "files",
   "one_line": "Fly-brain oracle answers sit at p(yes)~0.50 like the no-brain baseline; controls exist but no forecast outcomes are scored yet.",
   "commit": "ac0253e20b41182b03a7e280ac06d1f747bf137f",
   "checked_at": "2026-09-30T09:41:37Z",
   "sources": [
    {
     "url": "https://github.com/igdigitallab/bioreservoir",
     "label": "direct"
    },
    {
     "url": "https://raw.githubusercontent.com/igdigitallab/bioreservoir/ac0253e20b41182b03a7e280ac06d1f747bf137f/experiments/001-fly-oracle/results.csv",
     "label": "direct"
    }
   ]
  },
  {
   "id": "wired-different",
   "name": "Wired Different (ConnectomeLens)",
   "evidence_grade": "n/a",
   "made_by_us": false,
   "task_family": "graph-analysis",
   "loop": "offline",
   "wiring": {
    "dataset": "MaleCNS",
    "release": "v1.0 (neuPrint male-cns:v1.0)",
    "scope": "cell-type graph, 11,751 types"
   },
   "trained_parts": "LightGBM classifier on graph/neuropil/transmitter features (retrained per rewiring, same folds)",
   "metric": {
    "name": "out-of-fold AUC-PR for sex-related cell types (all 89 features)",
    "unit": "AUC-PR (0-1)",
    "higher_is_better": true
   },
   "real": {
    "value": 0.7589,
    "n": 1,
    "spread": null,
    "source": "03d9c5c results/null_model_summary.json: full.real"
   },
   "floor": {
    "label": "chance (prevalence of 478/11,751 sex-related types)",
    "value": 0.041,
    "source": "03d9c5c README.md:L29"
   },
   "arms": [
    {
     "arm_id": "degree_preserving_full",
     "null_type": "degree-preserving",
     "label": "randomized wirings (igraph degree-preserving swaps, 10x|E|, out-strength kept)",
     "value": 0.7116,
     "n": 500,
     "samples": 500,
     "spread": "± 0.0049 sd",
     "readout_refit": true,
     "source": "03d9c5c results/null_model_scores.csv: auc_pr_full, 500 trials (summary json full.null_mean)",
     "recomputed_by_us": true,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "neuropil_only",
     "null_type": "no-graph-model",
     "label": "neuropil only (exploratory)",
     "value": 0.722,
     "n": 1,
     "samples": 1,
     "spread": null,
     "readout_refit": true,
     "source": "03d9c5c README.md:L104",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "degree_preserving_full",
    "best_null_arm": "degree_preserving_full",
    "retention": 0.9341,
    "retention_method": "floor",
    "baseline_arm": "neuropil_only",
    "baseline_retention": 0.9486,
    "difference": 0.0473,
    "ci": null,
    "p": 0.002,
    "stat_note": "Empirical p=(0+1)/(500+1) with 0/500 nulls >= real (recomputed by us; z=9.6). Topology-only features: real 0.4888 vs null 0.0696 ± 0.0038 (retention 0.06, p=0.002). Most signal is neuropil location (0.722 alone); nulls keep non-topology features, so full-feature gap is small."
   },
   "wiring_effect": "helps",
   "baseline_beaten": "no",
   "fairness": {
    "keeps_degree": true,
    "keeps_io_boundary": null,
    "tuned_equally": true,
    "null_samples": 500,
    "has_no_brain_baseline": true,
    "verdict": "fair",
    "reason": "Topology features recomputed on each of 500 degree-preserving rewirings, same model and folds retrained each time; non-topology features kept. I/O boundary not applicable (no dynamics)."
   },
   "quality": {
    "n_per_arm": 500,
    "held_out": true,
    "uncertainty_reported": true,
    "results_in_repo_files": true,
    "independent_test": false,
    "peer_review": "none",
    "recomputed_by_us": true,
    "grade": "strong",
    "reason": "500 null rewirings with per-trial scores committed, cross-validated metric, empirical p recomputed; correlational annotation task, not peer reviewed."
   },
   "extraction": "files",
   "one_line": "Real MaleCNS topology predicts sex-related cell types better than all 500 degree-preserving rewirings, though neuropil location carries most of the signal.",
   "commit": "03d9c5c2d194532b661141cf3c60f8e58f0707d7",
   "checked_at": "2026-09-30T09:42:09Z",
   "sources": [
    {
     "url": "https://github.com/dhruvin-sarkar/ConnectomeLens",
     "label": "direct"
    },
    {
     "url": "https://raw.githubusercontent.com/dhruvin-sarkar/ConnectomeLens/03d9c5c2d194532b661141cf3c60f8e58f0707d7/results/null_model_summary.json",
     "label": "direct"
    },
    {
     "url": "https://raw.githubusercontent.com/dhruvin-sarkar/ConnectomeLens/03d9c5c2d194532b661141cf3c60f8e58f0707d7/results/null_model_scores.csv",
     "label": "direct"
    }
   ]
  },
  {
   "id": "flyconnectome-nulls",
   "name": "Null-model treatment of the sensory-motor boundary changes an evolutionary connectome comparison",
   "evidence_grade": "A",
   "made_by_us": false,
   "task_family": "evolution",
   "loop": "closed-loop",
   "wiring": {
    "dataset": "FlyWire FAFB",
    "release": "v783",
    "scope": "512 cell-type groups + 1,000 Kenyon cells (no optic lobes)"
   },
   "trained_parts": "evolved: all group weights, time constants, biases, input gains, readout weights, plasticity rates",
   "metric": {
    "name": "fitness at generation 600, pooled over four ecologies (median over 10 seeds)",
    "unit": "fitness (a.u.)",
    "higher_is_better": true
   },
   "real": {
    "value": 1.57,
    "n": 10,
    "spread": null,
    "source": "2f5683d results/fix_analysis.txt \"final\" block, row ALL \"A 1.57\""
   },
   "floor": null,
   "arms": [
    {
     "arm_id": "N1-column-shuffle",
     "null_type": "target-permutation",
     "label": "N1 column shuffle (each source keeps out-degree and weights, targets random non-sensory rows)",
     "value": 1.78,
     "n": 10,
     "samples": 10,
     "spread": "95% CI of connectome minus N1 [-0.34, -0.06]",
     "readout_refit": false,
     "source": "2f5683d results/fix_analysis.txt \"final\" block, row ALL \"N1 1.78\"; diff -0.224, p_holm 0.029",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "N2-degree-swap",
     "null_type": "degree-preserving",
     "label": "N2 degree-preserving (Maslov-Sneppen) swaps",
     "value": 1.77,
     "n": 10,
     "samples": 10,
     "spread": "95% CI of connectome minus N2 [-0.41, -0.11]",
     "readout_refit": false,
     "source": "2f5683d results/fix_analysis.txt \"final\" block, row ALL \"N2 1.77\"; diff -0.199, p_holm 0.016",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "N4-interior-swap",
     "null_type": "boundary-preserving",
     "label": "N4 interior-only degree swap (sensory and descending interface edges untouched)",
     "value": 1.58,
     "n": 10,
     "samples": 10,
     "spread": "95% CI of connectome minus N4 [-0.04, +0.05]",
     "readout_refit": false,
     "source": "2f5683d results/fix_analysis.txt \"final\" block, row ALL \"N4 1.58\"; diff +0.002, p_holm 0.750",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    },
    {
     "arm_id": "N5-interior-shuffle",
     "null_type": "boundary-preserving",
     "label": "N5 interior-only column shuffle (interface edges untouched)",
     "value": 1.62,
     "n": 10,
     "samples": 10,
     "spread": "95% CI of connectome minus N5 [-0.10, +0.05]",
     "readout_refit": false,
     "source": "2f5683d results/fix_analysis.txt \"final\" block, row ALL \"N5 1.62\"; diff -0.074, p_holm 0.750",
     "recomputed_by_us": false,
     "activity_ratio": null,
     "activity_class": "not-reported"
    }
   ],
   "effect": {
    "headline_arm": "N2-degree-swap",
    "best_null_arm": "N1-column-shuffle",
    "retention": 1.1274,
    "retention_method": "ratio",
    "baseline_arm": null,
    "baseline_retention": null,
    "difference": -0.2,
    "ci": [
     -0.41,
     -0.11
    ],
    "p": 0.016,
    "stat_note": "Author statistics: median of per-seed connectome-minus-null differences with 95% CI and Holm p over 4 controls (10 seeds per ecology). Headline difference is the reported median paired difference (-0.199), not 1.57-1.77. Standard nulls beat the connectome; boundary-preserving nulls do not differ (+0.002, -0.074, n.s.)."
   },
   "wiring_effect": "mixed",
   "baseline_beaten": "not-tested",
   "fairness": {
    "keeps_degree": true,
    "keeps_io_boundary": false,
    "tuned_equally": true,
    "null_samples": 10,
    "has_no_brain_baseline": false,
    "verdict": "fair",
    "reason": "All arms evolved with the same budget and 10 seeds, a new null realisation per seed (evo_fix.py:631). Headline N2 keeps degree but not the IO boundary; the boundary-preserving N4/N5 remove the null advantage (standard nulls open sensory-to-descending shortcuts: 10.7% vs 0.01%)."
   },
   "quality": {
    "n_per_arm": 10,
    "held_out": null,
    "uncertainty_reported": true,
    "results_in_repo_files": true,
    "independent_test": false,
    "peer_review": "preprint",
    "recomputed_by_us": false,
    "grade": "strong",
    "reason": "Committed analysis text with CIs and Holm-corrected p, 10 seeds and 10 null realisations per arm, pre-registration files; preprint only; not recomputed (per-seed table not opened)."
   },
   "extraction": "files",
   "one_line": "Standard nulls beat the connectome after evolution, but boundary-preserving nulls tie it: the choice of null flips the verdict.",
   "commit": "2f5683daa7a9619c228c05949d3208f3d32d3926",
   "checked_at": "2026-09-30T09:45:34Z",
   "sources": [
    {
     "url": "https://github.com/gyujeongion/flyconnectome-nulls/blob/2f5683daa7a9619c228c05949d3208f3d32d3926/results/fix_analysis.txt",
     "label": "direct"
    },
    {
     "url": "https://github.com/gyujeongion/flyconnectome-nulls/blob/2f5683daa7a9619c228c05949d3208f3d32d3926/evo_fix.py#L242-L342",
     "label": "direct"
    }
   ]
  }
 ],
 "chart_rows": [
  {
   "study_id": "byo-shiu-shuffle",
   "label": "Build your own: sugar to MN9 against four scrambled-wiring nulls (our run)",
   "task_family": "reflex-circuit",
   "headline_null_type": "degree-preserving",
   "retention": 0,
   "retention_method": "floor",
   "baseline_retention": null,
   "wiring_effect": "mixed",
   "quality": "weak",
   "fairness": "unfair"
  },
  {
   "study_id": "drosophila-brain-mlx",
   "label": "drosophila-brain-mlx",
   "task_family": "reflex-circuit",
   "headline_null_type": "degree-preserving",
   "retention": 0,
   "retention_method": "ratio",
   "baseline_retention": null,
   "wiring_effect": "helps",
   "quality": "weak",
   "fairness": "fair"
  },
  {
   "study_id": "fly-brain-lulzx",
   "label": "fly-brain",
   "task_family": "reflex-circuit",
   "headline_null_type": "weight-shuffle",
   "retention": 0.7023,
   "retention_method": "ratio",
   "baseline_retention": null,
   "wiring_effect": "helps",
   "quality": "moderate",
   "fairness": "partly-fair"
  },
  {
   "study_id": "shiu-drosophila-brain-model",
   "label": "Drosophila_brain_model (Shiu et al. 2024)",
   "task_family": "reflex-circuit",
   "headline_null_type": "weight-shuffle",
   "retention": 0.01,
   "retention_method": "ratio",
   "baseline_retention": null,
   "wiring_effect": "helps",
   "quality": "weak",
   "fairness": "partly-fair"
  },
  {
   "study_id": "fly-ocr",
   "label": "Fly OCR",
   "task_family": "sensory-model",
   "headline_null_type": "degree-preserving",
   "retention": 0.7083,
   "retention_method": "floor",
   "baseline_retention": null,
   "wiring_effect": "helps",
   "quality": "strong",
   "fairness": "fair"
  },
  {
   "study_id": "flydoom-mutkuoz",
   "label": "flydoom",
   "task_family": "sensory-model",
   "headline_null_type": "degree-preserving",
   "retention": 0.0006,
   "retention_method": "ratio",
   "baseline_retention": null,
   "wiring_effect": "helps",
   "quality": "weak",
   "fairness": "unfair"
  },
  {
   "study_id": "byo-flygym-body",
   "label": "Build your own, Add a body: FlyWire brain drives FlyGym walking (our run)",
   "task_family": "body-steering",
   "headline_null_type": "degree-preserving",
   "retention": 1.0161,
   "retention_method": "floor",
   "baseline_retention": 1.0253,
   "wiring_effect": "no-difference",
   "quality": "strong",
   "fairness": "fair"
  },
  {
   "study_id": "byo-flygym-loom",
   "label": "Build your own, Add a sense: a looming shadow turns the FlyGym fly (our run)",
   "task_family": "body-steering",
   "headline_null_type": "degree-preserving",
   "retention": 0,
   "retention_method": "floor",
   "baseline_retention": 0.1206,
   "wiring_effect": "helps",
   "quality": "strong",
   "fairness": "fair"
  },
  {
   "study_id": "flight-test-the-fly",
   "label": "Flight-test the fly",
   "task_family": "body-steering",
   "headline_null_type": "degree-preserving",
   "retention": 0.0014,
   "retention_method": "floor",
   "baseline_retention": 3.0687,
   "wiring_effect": "helps",
   "quality": "strong",
   "fairness": "fair"
  },
  {
   "study_id": "fly-brain-zero-shot",
   "label": "Are fruit flies zero-shot adapters?",
   "task_family": "body-steering",
   "headline_null_type": "degree-preserving",
   "retention": 0.0773,
   "retention_method": "floor",
   "baseline_retention": null,
   "wiring_effect": "helps",
   "quality": "moderate",
   "fairness": "partly-fair"
  },
  {
   "study_id": "fly-connectome-adds-to-body",
   "label": "FLY-lab: What a fly connectome adds to controlling a body",
   "task_family": "body-steering",
   "headline_null_type": "degree-preserving",
   "retention": 0.5267,
   "retention_method": "floor",
   "baseline_retention": 1,
   "wiring_effect": "helps",
   "quality": "strong",
   "fairness": "fair"
  },
  {
   "study_id": "fly-exe",
   "label": "Fly.exe (MaleCNS Virtual Fly)",
   "task_family": "body-steering",
   "headline_null_type": "degree-preserving",
   "retention": null,
   "retention_method": "none",
   "baseline_retention": null,
   "wiring_effect": "mixed",
   "quality": "weak",
   "fairness": "partly-fair"
  },
  {
   "study_id": "flyarm",
   "label": "FlyArm",
   "task_family": "body-steering",
   "headline_null_type": "degree-preserving",
   "retention": 0.7742,
   "retention_method": "ratio",
   "baseline_retention": 0.7881,
   "wiring_effect": "no-difference",
   "quality": "strong",
   "fairness": "fair"
  },
  {
   "study_id": "brain-runners",
   "label": "Brain Runners: untrained FlyWire brain (fly2) vs the same rule with no brain and with shuffled wiring",
   "task_family": "game-play",
   "headline_null_type": "degree-preserving",
   "retention": 0.3381,
   "retention_method": "ratio",
   "baseline_retention": 0.8993,
   "wiring_effect": "helps",
   "quality": "weak",
   "fairness": "unfair"
  },
  {
   "study_id": "chessfly",
   "label": "ChessFly",
   "task_family": "game-play",
   "headline_null_type": "degree-preserving",
   "retention": 0.9951,
   "retention_method": "floor",
   "baseline_retention": null,
   "wiring_effect": "no-difference",
   "quality": "weak",
   "fairness": "partly-fair"
  },
  {
   "study_id": "doom-fly-control",
   "label": "Is the fly brain actually playing DOOM? (control experiments)",
   "task_family": "game-play",
   "headline_null_type": "degree-preserving",
   "retention": 0.0261,
   "retention_method": "floor",
   "baseline_retention": 0.9705,
   "wiring_effect": "helps",
   "quality": "strong",
   "fairness": "fair"
  },
  {
   "study_id": "doomfly-rl",
   "label": "doomfly-rl",
   "task_family": "game-play",
   "headline_null_type": "degree-preserving",
   "retention": 1.0218,
   "retention_method": "ratio",
   "baseline_retention": 1.1263,
   "wiring_effect": "no-difference",
   "quality": "moderate",
   "fairness": "partly-fair"
  },
  {
   "study_id": "fly-dino",
   "label": "Fly Dino (flyjump)",
   "task_family": "game-play",
   "headline_null_type": null,
   "retention": null,
   "retention_method": "none",
   "baseline_retention": 1.0036,
   "wiring_effect": "not-tested",
   "quality": "moderate",
   "fairness": "partly-fair"
  },
  {
   "study_id": "fly-plays-games",
   "label": "fly-plays-games (Arkanoid / Retroid chapter)",
   "task_family": "game-play",
   "headline_null_type": "weight-shuffle",
   "retention": 1,
   "retention_method": "floor",
   "baseline_retention": 0.9012,
   "wiring_effect": "mixed",
   "quality": "weak",
   "fairness": "partly-fair"
  },
  {
   "study_id": "fly-racer",
   "label": "Fly-Racer",
   "task_family": "game-play",
   "headline_null_type": "degree-preserving",
   "retention": 0.9784,
   "retention_method": "ratio",
   "baseline_retention": 0.9898,
   "wiring_effect": "mixed",
   "quality": "weak",
   "fairness": "partly-fair"
  },
  {
   "study_id": "fly-self-driving",
   "label": "Fly Self Driving",
   "task_family": "game-play",
   "headline_null_type": "target-permutation",
   "retention": 0.8276,
   "retention_method": "floor",
   "baseline_retention": 0.931,
   "wiring_effect": "mixed",
   "quality": "weak",
   "fairness": "partly-fair"
  },
  {
   "study_id": "fly-tennis",
   "label": "Connectome ping pong (fly tennis)",
   "task_family": "game-play",
   "headline_null_type": "degree-preserving",
   "retention": 0.1852,
   "retention_method": "ratio",
   "baseline_retention": 0.2593,
   "wiring_effect": "helps",
   "quality": "weak",
   "fairness": "partly-fair"
  },
  {
   "study_id": "fly-worker",
   "label": "Fly Worker",
   "task_family": "game-play",
   "headline_null_type": null,
   "retention": null,
   "retention_method": "none",
   "baseline_retention": 1.5135,
   "wiring_effect": "not-tested",
   "quality": "weak",
   "fairness": "unclear"
  },
  {
   "study_id": "flyaim",
   "label": "FlyAim",
   "task_family": "game-play",
   "headline_null_type": "degree-preserving",
   "retention": null,
   "retention_method": "none",
   "baseline_retention": null,
   "wiring_effect": "no-difference",
   "quality": "moderate",
   "fairness": "partly-fair"
  },
  {
   "study_id": "flybrain-flappy",
   "label": "FlyBrain · Flappy (escape reflex)",
   "task_family": "game-play",
   "headline_null_type": "degree-preserving",
   "retention": 0,
   "retention_method": "floor",
   "baseline_retention": 0,
   "wiring_effect": "helps",
   "quality": "weak",
   "fairness": "unfair"
  },
  {
   "study_id": "haltere",
   "label": "Haltere",
   "task_family": "game-play",
   "headline_null_type": null,
   "retention": null,
   "retention_method": "none",
   "baseline_retention": null,
   "wiring_effect": "not-tested",
   "quality": "weak",
   "fairness": "unclear"
  },
  {
   "study_id": "making-fly-play-chess",
   "label": "making-fly-play-chess",
   "task_family": "game-play",
   "headline_null_type": "degree-preserving",
   "retention": 0.9763,
   "retention_method": "ratio",
   "baseline_retention": 1.3715,
   "wiring_effect": "no-difference",
   "quality": "strong",
   "fairness": "fair"
  },
  {
   "study_id": "fly-cartpole",
   "label": "fly-cartpole",
   "task_family": "ml-benchmark",
   "headline_null_type": "degree-preserving",
   "retention": 0.301,
   "retention_method": "floor",
   "baseline_retention": 1.0002,
   "wiring_effect": "helps",
   "quality": "strong",
   "fairness": "fair"
  },
  {
   "study_id": "flybench",
   "label": "flybench",
   "task_family": "ml-benchmark",
   "headline_null_type": "degree-preserving",
   "retention": 0.6665,
   "retention_method": "ratio",
   "baseline_retention": null,
   "wiring_effect": "mixed",
   "quality": "weak",
   "fairness": "partly-fair"
  },
  {
   "study_id": "flys-hash-function",
   "label": "The Fly's Hash Function",
   "task_family": "ml-benchmark",
   "headline_null_type": "degree-preserving",
   "retention": 1.1055,
   "retention_method": "ratio",
   "baseline_retention": 0.4835,
   "wiring_effect": "worse",
   "quality": "strong",
   "fairness": "fair"
  },
  {
   "study_id": "larva-vs-shuffles",
   "label": "Does the larval connectome beat its own shuffles? (connectome-null-models)",
   "task_family": "ml-benchmark",
   "headline_null_type": "degree-preserving",
   "retention": 0.9973,
   "retention_method": "floor",
   "baseline_retention": null,
   "wiring_effect": "no-difference",
   "quality": "strong",
   "fairness": "fair"
  },
  {
   "study_id": "neuroweave",
   "label": "NeuroWeave",
   "task_family": "ml-benchmark",
   "headline_null_type": "degree-preserving",
   "retention": 1.0102,
   "retention_method": "ratio",
   "baseline_retention": null,
   "wiring_effect": "no-difference",
   "quality": "weak",
   "fairness": "partly-fair"
  },
  {
   "study_id": "flybrain-reservoir",
   "label": "flybrain-reservoir",
   "task_family": "reservoir",
   "headline_null_type": "degree-preserving",
   "retention": 7.0455,
   "retention_method": "ratio",
   "baseline_retention": null,
   "wiring_effect": "worse",
   "quality": "weak",
   "fairness": "fair"
  },
  {
   "study_id": "flm",
   "label": "FLM - Fly Language Model",
   "task_family": "language-model",
   "headline_null_type": null,
   "retention": null,
   "retention_method": "none",
   "baseline_retention": 1.022,
   "wiring_effect": "not-tested",
   "quality": "weak",
   "fairness": "partly-fair"
  },
  {
   "study_id": "bioreservoir",
   "label": "BioReservoir",
   "task_family": "forecasting",
   "headline_null_type": "degree-preserving",
   "retention": null,
   "retention_method": "none",
   "baseline_retention": null,
   "wiring_effect": "not-tested",
   "quality": "moderate",
   "fairness": "partly-fair"
  },
  {
   "study_id": "wired-different",
   "label": "Wired Different (ConnectomeLens)",
   "task_family": "graph-analysis",
   "headline_null_type": "degree-preserving",
   "retention": 0.9341,
   "retention_method": "floor",
   "baseline_retention": 0.9486,
   "wiring_effect": "helps",
   "quality": "strong",
   "fairness": "fair"
  },
  {
   "study_id": "flyconnectome-nulls",
   "label": "Null-model treatment of the sensory-motor boundary changes an evolutionary connectome comparison",
   "task_family": "evolution",
   "headline_null_type": "degree-preserving",
   "retention": 1.1274,
   "retention_method": "ratio",
   "baseline_retention": null,
   "wiring_effect": "mixed",
   "quality": "strong",
   "fairness": "fair"
  }
 ]
}