Research report · 2 November 2026 · twenty-ninth check
Twenty-ninth check: the beginner guide re-tested from an empty folder, with memory measured per process
We re-ran the beginner path of Build your own from an empty folder twice, once with our exact package lock and once with the short install command, measured the memory of the simulation process itself and recorded why the machine counts as having no GPU. We also checked the pinned packages against the Python package index, scanned for new claims and re-checked the tracked repositories, links and videos.
Short answer
Re-test, 2 November 2026: 6 of 6 tests, 82.0 Hz bit-exact. The beginner path re-run from an empty folder on 2 Nov 2026 (Linux x86_64, Python 3.11.2, 2 CPUs, no GPU) with the pinned packages, and again with the short requirements.txt command (6 of 6 tests, 82.0 Hz). No page fixes were needed.
Device: Python 3.11; on 2 Nov 2026 we used 2 CPUs and no GPU (no NVIDIA device or driver present; Brian2 used its RuntimeDevice with the numpy code target, which the demo picks because the host has no C++ compiler). The simulation process itself peaked at 0.59 GB of memory for one 1-second trial, and the tests at 0.61 GB. These figures are the largest single process (measured with getrusage), not the whole test container, which peaked at about 1.7 GB including file cache and the parent process.
Pins: unchanged; checked against PyPI on 2 Nov 2026 (latest: Brian2 2.11.0, numpy 2.5.4, pandas 3.1.0, mujoco 3.17.0; our pins stay at 2.9.0, 2.3.5, 2.3.3 and 3.9.0), and the Brian2, numpy and pandas pins installed cleanly in the re-test of 2 Nov (mujoco, the body pin, was not re-installed); next re-test due 16 Nov 2026.
The re-test: two empty folders, two ways to install
Linux x86_64 (Debian 12), Python 3.11.2, 2 CPUs (AMD EPYC 7571), 3 GiB memory limit · model at commit 91bdd1e7 · the pass rule and the stop rules were written down before the first install step
Path A follows the guide with requirements.lock (21 exact versions). Path B uses the short requirements.txt command (6 pinned main packages; their helpers are whatever the package index offers that day). The first environment was deleted before the second was made, so each path started from nothing.
| Step | A: pinned lock | B: short command |
|---|---|---|
| Environment without pip, then get-pip | Worked as written; pip 26.2.2 (9.6 s) | Same (11.5 s) |
| Install | Exit 0 in 58.2 s; pip freeze equals the lock line for line | Exit 0 in 50.7 s; 7 helper packages newer than the lock, six not installed; the six main packages unchanged |
| Partial clone and data checksums | 4.8 s; both files match the guide's table; the upstream main branch still points to 91bdd1e7 | |
pytest -q | 6 passed in 87.2 s | 6 passed in 77.3 s |
| First sugar trial (real wiring, seed 0, 150 Hz, 1 s) | MN9 82.0 Hz, 13,372 spikes → extend proboscis; build 3.0 s, simulation 98.1 s, 111.5 s in all | MN9 82.0 Hz, 13,372 spikes → extend proboscis; build 2.7 s, simulation 86.1 s, 96.1 s in all |
| Peak memory, largest single process | Trial 0.59 GB; tests 0.60 GB | Trial 0.59 GB; tests 0.61 GB |
The trial's result is bit-exact with our runs of 28 Sep, 1, 5 and 19 Oct 2026 and inside the guide's 75–95 Hz pass band. The network had 138,639 neurons and 15,091,983 synapses, as before. The 20 Hz threshold that turns the rate into "extend proboscis" is our own rule.
On 2 Nov 2026 the short requirements.txt command installed 7 newer helper packages than our exact lock (Cython 3.3.0 → 3.3.1, iniconfig 2.3.0 → 2.3.1, MarkupSafe 3.0.3 → 3.0.4, pluggy 1.6.0 → 1.7.0, python-dateutil 2.9.0.post0 → 4.0.0, pytz 2026.4 → 2026.6, tzdata 2026.4 → 2026.6) and gave 82.0 Hz; the six main packages are pinned and did not change. It also did not install the package six, which is in our lock (python-dateutil 4.0.0 appears not to need it; we did not check its metadata). Use requirements.lock to get exactly our environment.
What the memory numbers measure
- 0.59 GB The simulation process, one 1-second sugar trial
- 0.61 GB The tests (pytest)
- about 1.7 GB The whole test container, with file cache and the parent process
The guide's code runs each trial in a fresh worker process. We read the peak memory of the largest finished child process from the operating system (getrusage), which here is that worker, the simulation itself: 564–566 MiB for a trial (0.59 GB) and 574–584 MiB for the tests (0.60–0.61 GB). The parent process, sampled every 2 s, stayed at about 0.2 GiB. GB values are decimal; in GiB the trial peak is 0.55.
The whole test container peaked at about 1.7 GB (sampled 1.66–1.69 GB during the four runs; its own lifetime record 1.70 GB). That figure also counts file cache from the 101 MB wiring file and from the environment, which sat on in-memory storage, so it is not the memory the project needs. Before each run the container held 0.98–1.01 GB, and no other simulation was running.
Why we say "no GPU"
- No NVIDIA device files and no NVIDIA driver directory on the machine; no graphics device directory either.
- Brian2 reported version 2.9.0 with its RuntimeDevice; each trial ran with the numpy code target, which the guide's code picks when no C++ compiler is found (none was installed).
- This describes one machine. It says nothing about how the project behaves on a machine with a GPU or a compiler; we did not test those.
Pins and package versions
Pins: unchanged; checked against PyPI on 2 Nov 2026 (latest: Brian2 2.11.0, numpy 2.5.4, pandas 3.1.0, mujoco 3.17.0; our pins stay at 2.9.0, 2.3.5, 2.3.3 and 3.9.0), and the Brian2, numpy and pandas pins installed cleanly in the re-test of 2 Nov (mujoco, the body pin, was not re-installed); next re-test due 16 Nov 2026.
- All 21 versions in our lock still have files for CPython 3.11 on the package index.
- Newer than our lock on 2 Nov: Brian2 2.11.0, Cython 3.3.1, iniconfig 2.3.1, MarkupSafe 3.0.4, mpmath 1.4.1, numpy 2.5.4, pandas 3.1.0, pluggy 1.7.0, pyarrow 26.0.0, pytest 9.2.0, python-dateutil 4.0.0, pytz 2026.6 and tzdata 2026.6. Cython 3.3.1 and python-dateutil 4.0.0 appeared after 31 Oct. Brian2 2.11.0 needs Python 3.13 or newer, so 2.9.0 stays the right pin for Python 3.11.
- get-pip.py is a newer file than on 19 Oct (sha256
f4448a01…f424; it installs pip 26.2.2 instead of 26.2.1). The guide's commands did not need to change.
Run times: slower than on 19 Oct
| Step | 19 Oct | 2 Nov |
|---|---|---|
| Pinned install | 35 s | 58 s |
| Tests | 58 s | 77–87 s |
| First sugar trial, start to finish | 68 s | 96–112 s |
On 2 Nov the first sugar trial took 96 to 112 s from start to finish and the tests 77 to 87 s, longer than on 19 Oct (68 s and 58 s); we do not know why. Load from other work on the shared processors is a plausible reason, but we did not check it. A light, low-priority search task overlapped the first 25 s of one test run; the results are deterministic, so only the time could be affected.
Disk
The working folder took about 660 MB of disk: a 464 MB Python environment and a 196 MB model folder with its data. The largest use of our scratch space during the run was 692 MB. Both environments and the model folder were deleted at the end.
Tracking and leads
- 2 Nov: we scanned news, videos and new code from 1 to 2 Nov; nothing met the bar for a new claim check. Both YouTube searches (by upload date and by views this week) answered in full; Hacker News had no fly stories; Reddit could not be searched (access refused). Dataset pages answered as on 31 Oct, and the Janelia male CNS page still answers 404.
- Coverage gap: the news and code searches started at 1 Nov 00:00 UTC, so the rest of 31 Oct after our previous scan was covered only by the video and Google News searches.
- 2 Nov 2026: we looked up the latest commit of 142 tracked repositories again: 4 have new commits since 31 Oct, recorded and not judged yet; flydoom is still gone or private. The repositories of 13 of the 25 verdict rows were re-checked, none with new code since 31 Oct. Links as on 31 Oct (fruitfly.live still fails its security certificate; the ascend demo and the FlyGym browser viewer still answer 404). 57 of 57 videos available.
What we could not resolve
- The memory figures are the largest single process. If a future version of the code ran trials side by side, the total would be larger.
- One machine, one operating system (Linux) and one Python version (3.11.2). The quick look, the 20-trial run and the "change one thing" experiments were not re-run.
- Why the run times were longer than on 19 Oct is not known.
- That python-dateutil 4.0.0 no longer needs six is inferred from what was installed; we did not read its package metadata.
- 4 tracked repositories with new commits since 31 Oct were recorded but not judged.
Sources: Shiu et al. brain model code (commit 91bdd1e7); get-pip.py (sha256 above, read 2 Nov 2026); Python package index (versions read 2 Nov 2026); the guide: Build your own; the plain answer: What do I need to run it?