Thirteenth check: the controls ledger catches up, and 4 new entries

What changed on the site

• Does fly wiring help?: chart and counts rebuilt from the new ledger; a "Changed on 11 Oct" list with the five plain sentences and their quality labels.
• New and updated: the 11 Oct entry with the three flag changes and the 4 new entries.
• Ways to try: one more "Does it actually run?" row.
• Is it real?: the plain answer about suffering is now titled the way people ask it, "Is the fly being tortured? Can it suffer?" (same sourced answer). Every view count on the site now comes from the gallery's data file, with "approximately" and its date.
• For AI agents: release date, release id and current counts first, and a one-line description of Digital Fly Lab.

Five new ledger rows, with every number and source line

Authors' own numbers · not recomputed by us · scored only by our fixed rules, no override

Each row was built by our ledger script from the source lines below; a second build gave the same result. The plain sentences, verbatim:

• Mixed fly-distill (grade A; catalogue flag was "no difference" until 10 Oct): fly-distill: the authors compared the real wiring with a degree-preserving shuffle on bits per byte (lower is better). Real 3.9572, null 3.9171 (4 null samples). Our ledger rules read this as mixed. Quality: fair, strong (authors' numbers, not recomputed by us; source STATUS.md lines 9-18 at ad5aef6).
• Worse flywire-connectome (mushroom-body model) (grade A; catalogue flag was "mixed" until 10 Oct; only the task metric is scored): flywire-connectome (the author's mushroom-body rate model): the authors compared the real wiring with strength-preserving null wirings (median of 100) on odour discrimination d'. Real 18.32, null 19.45 (100 null samples). Our ledger rules read this as worse. Quality: partly fair, strong (authors' numbers, not recomputed by us; source HALLEM_COMMUNITY_RESULTS.md lines 9 and 52-57 and PREREGISTRATION_16_18.md line 55 at 8aea8e8).
• No difference brainlab (grade A; catalogue flag was "not tested" until 10 Oct; a task metric was found in its committed results): brainlab: the authors compared the real wiring with a target-permuted copy on held-out decision accuracy (%). Real 92.91, null 92.20 (1 null sample). Our ledger rules read this as no-difference. Quality: partly fair, weak (authors' numbers, not recomputed by us; source reports/2026-10-05-scorer-polis-s2-screen.md lines 39-40 at 1fe1aa4).
• Helps MelanoGraph (grade A; new entry): MelanoGraph: the authors compared the real wiring with a degree-preserving rewired copy on the receiver's gain over a resampled message. Real 0.4941, null 0 (1 null sample). Our ledger rules read this as helps. Quality: partly fair, weak (authors' numbers, not recomputed by us; source docs/NEURAL_C2B_REPORT.md lines 35-42 at 45ad05d).
• Worse fly-arithmetic (grade A; new entry): fly-arithmetic: the authors compared the real wiring with a random network of the same size on held-out addition accuracy (%). Real 48, null 98 (10 null samples). Our ledger rules read this as worse. Quality: partly fair, strong (authors' numbers, not recomputed by us; source docs/RESULTS.md lines 4 and 11-15 at e36dd2b).

• fly-distill: The Erdos-Renyi random graph (3.5234 ± 0.0294) beats the real wiring, which is why the rules read mixed.
• flywire-connectome (mushroom-body model): Only the task metric is scored; the structure statistic (convergent projection-neuron pairs) is not. The author's own rule also reads worse. Per-null recalibration is not stated.
• brainlab: One seed and one permutation, no uncertainty reported; the authors read the gap as not load-bearing. A fine-tuned 2B language-model decider scores 100% in its own files. The spiking-activity campaign (about 3.7 times the spikes) is not scored.
• MelanoGraph: 95% CI of the difference 0.4938-0.4944. Per-seed outputs are not committed; numbers come from the committed report. A two-choice task, so any scrambling drops it to chance.
• fly-arithmetic: The random network does not keep each neuron's degree. A linear model on the raw numbers scored 94.

Choices made under the fixed rules

• brainlab gets a row. The plan said it gets a row if its committed results contain a task metric; they do (a frozen 1,536-neuron FlyWire olfactory circuit inside a trained decision scorer, compared with a target-permuted copy). The fallback sentence for "no task" was therefore not used. Its "Is it AI?" class also moved to "whole network or synapses trained", because the scorer trains per-connection gains, an encoder, a readout and heads.
• The mushroom-body odour model: only the task metric (odour d') is scored, not the structure statistic (59 convergent pairs against 5.6 expected). That is why its flag moved from mixed to worse.
• MelanoGraph has two pre-registered studies; we scored C2b, the task, whose report names the gain as primary. C2a would also read "helps" (decoding 8 patterns: real 0.7285, degree-preserving 0.4590, random targets 0.2500).
• Task families are our choice: brainlab is a game-play study, MelanoGraph a sensory model, fly-arithmetic a reservoir. Sensory models now read helps 4, worse 1; reservoir studies worse 2.

Catalogue flag changes

4 new entries, each graded from its code

Each repository was read at a fixed commit; the basis for each grade is in its catalogue record. Project claims below are quoted from their READMEs and are not our findings.

• A MelanoGraph (akashacorporation): a Python and Rust re-implementation of the Shiu et al. whole-brain model with sealed pre-registrations and a rewiring control that keeps degree, weight and excitation/inhibition balance. The project says: "An executable, reproducible simulation substrate for the Drosophila whole-brain connectome, and a preregistered programme asking what circuit organisation adds to social learning." (README)
• A fly-arithmetic (krezio): about 2,000 FlyWire neurons as a rate network with a trained readout, against a random network of the same size over 10 seeds. The project asks: "Can a Fruit Fly's Brain Learn Maths?" and writes: "The real fly wiring did worse than random wiring." (README)
• B FlyRoyale (Zaib111): a Clash Royale-style game whose policy is a FlyWire rate model with gains, encoder and decoder trained by self-play. The project says: "A real fruit-fly brain learns to play Clash Royale, just from winning and losing against itself." (README) No results are committed, so there is nothing to compare with its MLP baseline.
• B connectome-control-exploration (sakikoniko314-blip): the fly.ai connectome model steers a simulated cart toward heat; lesion and input controls only, no shuffled wiring, and no licence file for its own code. Its README says, in Chinese, that it "does not show that fly control is better than conventional algorithms" (our translation).

Lead scan and triage

• Backlog triage: all 35 waiting names got a dated decision on 11 Oct (6 priority 1, including the two admitted with a ledger row; 10 priority 2; 16 priority 3; 3 excluded: one repository now gone, one with no README, one duplicate). Four name a wiring control in their README and are queued for the next checks.
• New leads: 17 new GitHub repositories since 10 Oct (all with 0-1 stars) and 16 new videos; none met the bar for a verdict. The directory pages we track list the same totals as on 10 Oct.
• Attention: videos that frame the simulated fly as "tortured" keep spreading in English, French and Russian; this is new attention on an existing theme, not a new claim. Our plain answer now uses that wording.

Tracking, freshness and view counts

• Code: 111 repositories: 109 unchanged, 1 changed (fly.ai, an upstream dependency), 1 still gone (flydoom, record kept). The mushroom-body model's repository moved to a new commit today; its grade stays.
• Links: 273 checked: no new failure of a main link. Two paper PDFs on GitHub answered HTTP 503 at check time, probably for a moment; not confirmed.
• Verdicts: all 21 re-checked: 13 with code unchanged since graded, 8 with no code, 0 changed.
• Python packages: no pinned package was patched since 27 Sep; FlyGym still requires Python 3.12.
• View counts: 18 gallery cards re-read on 11 Oct 2026 (10 from video pages, 8 from search pages). For example, the "learns to hack" video had approximately 70,445 views on 11 Oct 2026, the Gorilla Tag video approximately 995,029 views on 11 Oct 2026 and SWOT's farming video approximately 231,851 views on 11 Oct 2026. 25 counts are still dated 28 Sep (mostly X, TikTok and project-site cards), and 10 records have no count.

Browser check: fly-arithmetic

krezio.github.io/fly-arithmetic ran in headless Chromium 154 on Linux with no GPU and no action from us: first activity after about 11.2 s, about 2.05 MB transferred, 0 console errors and 0 page errors. The page used one worker and WebGL canvases, so by our fixed rule where it computes is "unclear". Peak memory about 1.02 GB. We did not type a sum or train a readout. The row on Ways to try

Method and limits

• All ledger numbers are the authors' own; none was recomputed by us. Three sources are author summaries (fly-distill, MelanoGraph and the mushroom-body model); the mushroom-body p restates the author's "98% of nulls at or above the real value", and its null value is a median.
• MelanoGraph and brainlab each rest on one rewired graph, so their quality is weak.
• Each new entry was graded in a few minutes from targeted file reads; the fly.ai kernel behind connectome-control-exploration was not inspected.
• Our ledger script limits a study's one-line summary to 25 words, so the full plain sentence is stored at the start of the row's quality reason and a shorter form with the same numbers in its one-line field.
• The "Is it AI?" counts were recomputed (59 of 98 graded entries have a trained part); the other plain answers are unchanged.
• No simulation of our own ran. A memory probe for our planned flyvis retry found no writable disk outside the workspace, so run 14 must fit within the shared memory budget.