# Felisyn laptop move: the plaids pilot's intact run repeated spike for spike on a laptop

Evidence for journal note 23, dated 7 November 2026. The research server was lost, so the 2 mm model moved to an 8 GB laptop. The test was the plaids pilot's intact run from protocol 022, repeated on the laptop with the same arguments, seed and stimulus sequence, under a 3.7 GiB memory cap.

- **Result.** All 18,600 neurons' spike trains equal the server pilot's export over the 35,133 ms sequence, at the export's 0.01 ms precision: 7,948,521 spikes. The frozen analysis gave the same XOS, Dc and ds. Memory peaked at 3.208 GiB, and the run took 7,216 s (2.00 h).
- **Run.** 7 November 2026, 06:08:32 to 08:08:48 UTC. Exit 0, not stopped by the cap.
- **Numerical agreement only.** This is not new biology, not a confirmation of the pilot and not a statement about cat cortex. It covers one network instance and one stimulus sequence, between one server build and one laptop build. Protocol 023, the confirmatory test, was declared on 7 November 2026 at 20:25:30 UTC and has not been run.

## What is here

| File | What it holds |
| --- | --- |
| `g4-gate.md` | The four pass conditions, written before the run. See "The gate" below |
| `g4-verdict.json` | The gate's result, condition by condition, plus the cortical trains and peaks |
| `compare.json` | Spike train comparison with the pilot, per sheet, over 0 to 35,133 ms |
| `results.json` | The frozen analysis on the laptop's export. See "Reading results.json" below |
| `g4.mem.json` | The memory cap's summary of the run |
| `g4.mem.csv` | The run's memory curve, thinned. See "Thinning" below |
| `g2-n.txt` | The small regression before and after the noise fix: 15 of 1,656 trains equal, then 1,656 of 1,656 |
| `proof-bytes.txt` | Byte-identity tests of the memory changes on a 0.6 mm sheet, with the control run that must differ |
| `run_controller_mem.py` | The wrapper that ran the model. Its docstring explains each memory change |
| `memcap.py` | The memory cap |
| `patches/0001-lgn-noise-pair-order.patch` | The LGN noise patch |
| `patches/0002-block-size-shift.patch` | The NEST block size patch |
| `deviations-public.md` | Every known difference from the server's recipe |
| `figures/` | Figure 1 as PNG and SVG, its plotting code, and its caption, alt text and attribution in `captions.md` |
| `manifest.json` | SHA-256 and size of every file here except itself |

## Reading these files

- **Names.** "Mac" means the laptop. "G2" is the small regression and "G4" the full pilot rerun. "Build N" and "build NM" are explained in `deviations-public.md`. "R0" to "R10" are the memory changes in `run_controller_mem.py`.
- **`compare.json`.** Fields named `short_*` describe the laptop run. The tool was first written for a short run, and the name stayed.
- **Comparison window.** `compare.json` covers the sequence segments only, 0 to 35,133 ms. The closing blank, which runs to 35,583.1 ms of model time, is not compared.
- **Export files.** The laptop's export has SHA-256 `472eaf9d7d58828f2fbec71ed7490cdb6369ddda1e38df0c2e94b6bc6c6da841` and 25,552,688 bytes. The pilot's has SHA-256 `952b469b9a8e74a482f704854a0b5462f6c68798ba3466e1a15aa0e893eae281` and 25,552,690 bytes, as `data-sha256.txt` in the plaids 022 package lists. The files differ. Each records its own `run_seconds`, and no whole-file comparison was made. The comparison is over spike times, neuron tables, resolved parameters, steps and the connection count.
- **Peaks in the CSV and JSON.** The cap samples every 2 s and logs a row once at least 15 s have passed since the last, about every 16 s in this run, so the log can miss the top. `g4.mem.json` gives 3.208 GiB at 6,952.6 s from the 2 s samples, and the highest logged row reads 3.184 GiB at 6,944 s. Peak resident size is 1.906 GiB in the JSON and 1.823 GiB in the log.
- **Swap.** It stood at 423 MiB when the run started, and that is also its highest value during the run. It fell to 399 MiB by the end.

## Reading results.json

The laptop rerun repeated only the intact run. The frozen `analyze_plaids.py` needs a status for the lesion run as well, so it was called with `lesion=FAILED`. The file therefore says `"lesion": "failed twice"`, and `connection_readings` says the lesion run failed twice. **No lesion run was attempted on the laptop.** Those strings come from that call, and they do not describe an event. The file's `outcome` and `text` fields repeat the pilot's, because the same spikes went through the same frozen code. They are not a new outcome.

A check of this file against the published `plaids-022/results/results.json`:

- The two files share 6,284 numeric values, and 6,283 are exactly equal.
- The one that differs is `runs/intact/run_seconds`: 10,378.49 on the server and 7,179.57 on the laptop.
- The other shared fields that differ are the lesion status, the two connection readings, the package path and the export hash.
- The pilot's lesion values are absent here.

## The gate

`g4-gate.md` was written before the run. Its file time is 06:08:21 UTC and the run started at 06:08:32 UTC. That timing comes from the laptop's own clock and file system. No hash of the gate was published before the run, so the order rests on those records alone. Its tolerances are descriptive and size nothing in protocol 023.

## Scrubs

Published copies differ from the local originals only as listed here. `manifest.json` gives each original's SHA-256 under `source_sha256`. No frozen package was changed, and none of these files is part of one. The changes:

| File | Lines of the original | Change |
| --- | --- | --- |
| `compare.json` | 2, 4 | Local folder prefixes removed from two path fields, leaving paths relative to the project folder |
| `proof-bytes.txt` | 13, 17, 18, 22, 26, 30, 33, 36, 39, 43, 47 | Local folder prefix removed from ten path lines, leaving `runs/proof/...`. In line 13, a first-person phrase about a tool time limit now reads "by a tool time limit". The meaning did not change |
| `results.json` | 289 | Local folder prefix replaced by `<local>/` in one path field |
| `memcap.py` | 12 | One docstring line reworded to remove internal working notes. No code changed |
| `g4-gate.md` | 18 | One sentence shortened to remove an internal working note. The conditions did not change |
| `patches/0001-lgn-noise-pair-order.patch` | 4 | One header line reworded. The diff did not change |
| `patches/0002-block-size-shift.patch` | 4 | One header line reworded. The diff did not change |

`run_controller_mem.py`, `g4-verdict.json`, `g4.mem.json` and `g2-n.txt` are the originals, byte for byte. The wrapper's SHA-256, `c7c19469b8bb87cf83866da04c2d207ff34388433957640d942f01ca2a248c3e`, is the one the run's start script checked.

`memcap.py`'s original SHA-256, `28d619e3b2a05414cfda28a81e06a6ca1b8d609cd1eb28210ddbd7199fb832ab`, is the one the run recorded at its start.

## Thinning

`g4.mem.csv` keeps 140 of the log's 442 rows: every 4th row of the log, plus the first row, the last row, the row with the highest footprint, and both rows of any jump larger than 0.05 GiB between neighbouring rows. The values are unchanged. The full log's SHA-256 is in `manifest.json`. Figure 1 is drawn from this thinned file, so anyone can redraw it from this folder.

## Not here

- The laptop's export, 25.6 MB, stays off the site. Its hash is above. The pilot's export is too large for the site as well, as note 22 says.
- The run logs and proof run folders that `proof-bytes.txt` and `g2-n.txt` mention are not published, nor are the memory study and working logs that the patches and the wrapper cite. They contain local paths and working notes. Their numbers that matter are in `deviations-public.md`, in these files or in note 23.
- The larger files of run 011 no longer exist. See note 23. Their hashes and sizes are in `run-011/data-on-server-sha256.txt` and `run-011/data-on-server-bytes.txt`.

## Credits and licences

- **Model.** Antolík J, Cagnol R, Rózsa T, Monier C, Frégnac Y, Davison AP (2024), *A comprehensive data-driven model of cat primary visual cortex*, [PLOS Computational Biology 20(8): e1012342](https://doi.org/10.1371/journal.pcbi.1012342), published under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
- **Software.**
  - The LSV1M model code v1.0 is under the MIT license.
  - The Mozaik simulation framework (v0.4.0) and PyNN (0.11.0) are under CeCILL.
  - The NEST simulator (3.4) is under the GNU GPL, version 2 or later.
  - The imagen pattern library is under its BSD license.
  - GSL is under the GNU GPL.
- **Patches.** Patch 0002 changes a NEST source file. Patch 0001 changes a file of the NEST step current module, [CSNG-MFF/nest-step-current-module](https://github.com/CSNG-MFF/nest-step-current-module/tree/fe91c0345a602d564a74ebe0ea2c66d9a4157005) at commit `fe91c0345a602d564a74ebe0ea2c66d9a4157005`. That file's header states that it is part of NEST under the GNU GPL, version 2 or later. The module has no licence file at its root. Both patches are ours, and they are offered under the terms of the files they change.
- **Figures.** Set in Latin Modern Roman (GUST Font License). Drawn with matplotlib.
- **This work.** The reduction, the runs, the analysis, the figures, the two patches, the wrapper, the memory cap and the text are our own.
