# Felisyn run 011: the full model under protocol 007

Recorded 22 October 2026 (Europe/Vienna). The run started on 22 October 2026 at 06:12 UTC and saved its data at 21:08 UTC. The analysis ran from 21:21 to 21:34 UTC. Everything ran on the project's existing 4-vCPU research server.

This is the first complete run of the full published cat primary visual cortex model in this project: 108,150 neurons and about 99.4 million connections, 40.32 s of spontaneous activity. It followed the frozen [protocol 007](https://www.felisyn.com/research/protocol-007/README.md) without changes.

- **Execution passed.** Exit 0; raw data saved; no explosion. The four simulation calls ran as declared and ended at 40,770.1 ms of NEST time, with 16 threads and 16 virtual processes on every call.
- **Analysis parity: 48 of 48.** For all six measurements in all four populations, the independent helper code and the numerical part of the authors' Figure 4 code give the same mean and SD to within 8.9 × 10⁻¹⁶. The authors' code is extracted unmodified and stops before any plotting; the paper's figure itself is not redrawn.
- **All 24 printed Figure 4 means are matched at printed precision.** Each population mean lies within half a unit of the last digit printed in the paper's Figure 4.

The comparison is descriptive. It shows agreement with the paper's printed model values. It is not biological validation, and not a paper-reproduction pass, because the full-precision reference values are not available. It is not controller evidence.

## What ran

| Item | Value |
| --- | --- |
| Model | LSV1M v1.0 (unchanged; 37 files match protocol-007's hashes) |
| Recipe | `param_spont/defaults`, `create_experiments_spont` |
| Network | 108,150 neurons; 99,352,247 connections in the 16 projections (NEST's count after construction in feasibility-010's build of the same network; see below) |
| Split | 1 MPI rank × 16 OpenMP threads, passive waiting |
| Host | existing 4 vCPU, ~7.8 GiB, temporary 64 GiB swap file (removed afterwards, `receipts/host-after-run.txt`) |
| Contract / manifest SHA-256 | `54ab49ae…` / `0a49fda9…` (in `receipts/launch.json`) |
| Run directory | `full-20261022T061230Z` |

The run calls, from the receipts written after each PyNN run step (`receipts/simulation.log.gz`, `FELISYN-SIMULATE` lines):

| Call | Requested | NEST time after | Wall time | NEST loop time | Loop per simulated second |
| --- | --- | --- | --- | --- | --- |
| Blank | 150 ms | 150.1 ms | 626 s | 347 s | 2,314 s |
| Spontaneous segment | 40,320 ms | 40,470.1 ms | 37,296 s (10.4 h) | 37,296 s | 925 s |
| Blank | 150 ms | 40,620.1 ms | 181 s | 180 s | 1,200 s |
| Terminal blank | 150 ms | 40,770.1 ms | 122 s | 122 s | 812 s |

**Connections.** This run's receipts give NEST's count during the run calls: 99,385,122. That count includes the connections to recording devices, which are attached after construction. This run did not record a count right after construction. The feasibility-010 probe built the same network with the same seeds and split, and NEST counted 99,352,247 connections at that point. All 16 projection counts in Mozaik's log, which prints six significant digits, are the same in the two builds, and they sum to 99,352,247 within that rounding. The difference of 32,875 equals the 31,258 spike-recorded and 1,597 voltage-recorded cortical neurons plus 20. The 20 fit the five-cell thalamic recorder in each of the two thalamic sheets, with spikes and voltage recorded for each cell; this breakdown is inferred from the counts.

**Seeds.** As documented in protocol 007, PyNN seed 5 is used and the effective Mozaik seed is 513. NEST's own kernel seed was 42. That is PyNN's default, because the framework's `sim.setup` call passes no seed (`receipts/seed-source.txt` quotes the lines, with file hashes, from the pinned environment). It is part of how the released code runs and is not a change to the protocol, but it is recorded here because protocol 007 did not mention it.

## Results

Population means of our run against the values printed in Figure 4 (`analysis/printed-reference-comparison.json`). All 24 are within half a printed unit.

| Population | Rate (Hz) | ISI CV | Correlation | Voltage (mV) | Exc. conductance (nS) | Inh. conductance (nS) |
| --- | --- | --- | --- | --- | --- | --- |
| L4 excitatory | 1.374 (1.4) | 0.950 (0.95) | 0.00163 (0.0016) | −66.95 (−67) | 1.266 (1.3) | 3.057 (3.1) |
| L4 inhibitory | 7.581 (7.6) | 0.979 (0.98) | 0.00352 (0.0035) | −64.56 (−65) | 1.189 (1.2) | 1.832 (1.8) |
| L2/3 excitatory | 2.015 (2.0) | 0.982 (0.98) | 0.00670 (0.0067) | −71.02 (−71) | 1.146 (1.1) | 5.838 (5.8) |
| L2/3 inhibitory | 4.725 (4.7) | 0.972 (0.97) | 0.0206 (0.021) | −69.20 (−69) | 1.010 (1.0) | 3.520 (3.5) |

Printed values are in brackets, transcribed in benchmark-004. The analysis covered the full 40,320 ms window with 4,032 bins of 10 ms:

- **Spike trains:** 12,128 / 3,467 / 12,176 / 3,487 (L4 exc / L4 inh / L2/3 exc / L2/3 inh).
- **Eligible for ISI CV and correlation:** 8,391 / 3,467 / 12,167 / 3,487. These neurons have at least 7 spikes; silent and sparse cells still count in the rates.
- **Voltage and conductance traces:** 400 / 399 / 400 / 398 neurons, 40,320 samples each.

The SDs, eligibility counts and all values are in `analysis/analysis-parity.json`. The open question from protocol 006 remains: the population-weighted rounded inhibitory conductance is about 4.09 nS against 3.56 nS in the manuscript text.

**Declared secondary analysis.** These are spike measures on protocol-006's sample of neurons that the spontaneous recorders also record (`analysis/secondary-subset.json`). The neuron positions matched the frozen hashes.

| Population | Neurons | Rate (Hz) | ISI CV | Correlation |
| --- | --- | --- | --- | --- |
| L4 excitatory | 836 | 1.354 | 0.941 | 0.00744 |
| L4 inhibitory | 152 | 7.410 | 0.973 | 0.0330 |
| L2/3 excitatory | 856 | 1.964 | 0.966 | 0.0294 |
| L2/3 inhibitory | 173 | 4.585 | 0.966 | 0.103 |

Rates and ISI CVs are close to the full sets. Spike-count correlations in this small central sample are 4.4 to 9.4 times higher than in the full spontaneous sets and in the printed values. This fits recipe-check-009's inference that Figure 4 used the larger spontaneous set, but does not prove it. These values are descriptive and not a pass criterion.

## Time and resources

From `run-summary.json` and `receipts/resources.csv`:

- **Construction:** 3.34 h.
- **Main segment:** 10.36 h.
- **Retrieval of the 40 s recordings:** about 20 min.
- **Whole run:** 14 h 57 min.
- **Peak host memory in use:** 7.53 GiB.
- **Peak swap:** 28.4 GiB.
- **Swap traffic:** 888 GiB swapped in and 867 GiB swapped out over the run.

The feasibility probe's projection (about 12.6 h plus unmeasured steps, feasibility-010) underestimated the cost of the main segment: NEST's loop needed 925 s per simulated second rather than 760. The whole run took 14.95 h.

The analysis took 12 min 35 s, with a peak resident size of 6.8 GiB. Its correlation step built 12,000 × 12,000 matrices. The two smaller programs took 0.1 s and 35 s.

## Erratum to protocol 007

Protocol 007's README and contract (`decomposition.evidence`) say that in the tiny-model split check, first recorded voltage samples "differed by up to 10 mV". That is the largest difference between the sorted value sets. Per neuron, the difference reaches 13.0 mV for 16 × 1 versus 1 × 16, the split used here, and 16.3 mV for 16 × 1 versus 4 × 4 (feasibility-010, `receipts/decomp/sample0-groups.json`). The frozen files are not edited. No rule depends on this size.

## Limits

- One run, one seed set, one host. Rounding-level differences between decompositions were not ruled out at full scale. The results are those of the 1 × 16 split.
- Printed precision is coarse: half a unit of 1.4 Hz is ±0.05 Hz. The full-precision reference values are not available, so this is not a reproduction pass.
- Matching the paper's model output says nothing on its own about cat visual cortex. No biological validation was done.
- The datastore (27 files, 1.66 GB) and the voltage and conductance arrays (12 files, 1.44 GB) remain on the research server. `data-on-server-sha256.txt` and `data-on-server-bytes.txt` list their SHA-256 hashes and sizes. The spike data is published with this package as a separate 12 MB file.

## Evidence map

- `run-summary.json` (`analyze_full_run.py`): calls, phases and resources.
- `analysis/`: `analysis-parity.json` (48 comparisons), `printed-reference-comparison.json` (24 means), `secondary-subset.json`, and `exported-spikes.json.gz` (all spike times, 12 MB).
- `receipts/`: the launcher record (`launch.json`, `launch-007.log`), the runtime and parameter-schema checks, `simulation.log.gz` (complete; it decompresses to the hash recorded on the server), the framework's own log (`mozaik-log.gz`), host samples (`resources.csv`), the analysis logs and GNU time receipts, `results-metadata/` (saved parameters, recorders, workflow marker), the source of the NEST seed (`seed-source.txt`) and the host's swap state after the run (`host-after-run.txt`).
- `data-on-server-sha256.txt` and `data-on-server-bytes.txt`: hashes and sizes of the 39 data files kept on the server.
- `draw_results.py` → `figures/results-vs-printed.png`/`.svg`; `draw_run.py` → `figures/run-memory.png`/`.svg`. Both plot retained data only.
- `manifest.json` (SHA-256 and size of every file except itself, the zip and the X drafts) and `run-evidence.zip`. The zip holds every manifest file except `analysis/exported-spikes.json.gz`, which is a separate 12 MB download.

## Attribution

Model and Figure 4: 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](https://doi.org/10.1371/journal.pcbi.1012342), published under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). The printed values are our transcription. LSV1M v1.0 is under its MIT license. Mozaik v0.4.0 (CeCILL), PyNN 0.11.0 and NEST 3.4 are as pinned in protocol 006. The figures are new plots of our own run. No animal recordings were collected.
