# Felisyn recipe check 009: which recording set matches the published spontaneous statistics?

Recorded 22 October 2026 (Europe/Vienna). The recorder probe ran on 22 October 2026 from 00:16:30 to 00:32:24 UTC on the existing four-vCPU research server. The Figure 4G measurement was reproduced locally the same day. No neural simulation was run.

**Result 1, recorder counts.** We built the released model with our pinned stack and seeds and stopped before any connection was created. On that model, the two released recorder recipes select exactly the neuron counts listed in the paper's S5 Table. All **16 of 16** numbers are equal:

| Population | S5 "Spontaneous activity" | Released `param_spont` recorders | S5 "Evoked activity" | Released `param` recorders |
| --- | --- | --- | --- | --- |
| Layer 4 excitatory | 12,128 spikes · 400 V/g | 12,128 · 400 | 838 · 62 | 838 · 62 |
| Layer 4 inhibitory | 3,467 · 399 | 3,467 · 399 | 152 · 22 | 152 · 22 |
| Layer 2/3 excitatory | 12,176 · 400 | 12,176 · 400 | 862 · 80 | 862 · 80 |
| Layer 2/3 inhibitory | 3,487 · 398 | 3,487 · 398 | 173 · 26 | 173 · 26 |

"V/g" means membrane potential, excitatory and inhibitory conductance. All 16 MPI ranks computed the same selections. The four cortical populations contain 37,500, 9,375, 37,500 and 9,375 neurons.

**Result 2, the published figure.** The smallest values plotted in the published Figure 4G require more recorded neurons than the frozen `param` recipe records, in every population. For the inhibitory populations the shortfall is 7-10 fold (table below).

Together: protocol 006 froze `param`, the recording set that S5 lists for *evoked* activity. The published spontaneous statistics were most likely computed from the larger `param_spont` set. This remains an inference, because no record of the authors' Figure 4 run is available.

## Evidence for `param_spont`

- **S5 Table** documents a separate spontaneous recording set. On our build its 8 counts are exactly those selected by the released `param_spont` recorders: a dense 3 mm spike grid and a 2 mm analog grid. Its evoked counts are exactly those selected by the released `param` recorders, which protocol 006 froze: central 0.2-0.6 mm grids, plus a sparse 2 mm spike grid for the excitatory populations.
- **Code path.** In the released model, `Fig4.png` is written only by `perform_analysis_and_visualization_spont`: lines 760-804 of `analysis_and_visualization.py` in the protocol-006 package (`../protocol-006/model/`), with `paper_figures = True` at line 23. Only `run_spont.py` calls that function, and the released README runs it with `param_spont/defaults`.
- **Figure 4G.** The released plotting code bins log10 firing rates with `numpy.histogram(..., range=(-2, 2), bins=20, density=True)`. Every plotted value is therefore `count / (0.2 × N)`, where N is the number of nonzero rates in range and the count is a whole number of at least 1. The smallest value bounds N from below. We measured the published figure from the axis tick marks (calibration residuals ≤ 0.006 decades; one pixel is about 0.033 decades):

| Population | Smallest plotted value | Neurons needed (with 1-pixel error) | `param` records | `param_spont` records |
| --- | --- | --- | --- | --- |
| Layer 4 excitatory | 4.26 × 10⁻³ | ≥ 1,173 (1,089) | 838 | 12,128 |
| Layer 4 inhibitory | 4.34 × 10⁻³ | ≥ 1,151 (1,067) | 152 | 3,467 |
| Layer 2/3 excitatory | 2.49 × 10⁻³ | ≥ 2,005 (1,860) | 862 | 12,176 |
| Layer 2/3 inhibitory | 2.89 × 10⁻³ | ≥ 1,733 (1,607) | 173 | 3,487 |

With N equal to the `param_spont` counts, the three smallest layer 4 inhibitory values correspond to 3.01, 9.07 and 10.05 neurons, close to whole numbers. That fits but does not prove that set: neurons with zero rate or out-of-range rates are excluded from N.

## Evidence that points the other way

- **The authors' archive.** The only public archived run of this model (validation-005) uses the `param` recorders. It contains the same 40,320 ms spontaneous segment and lists a `Fig4` image. Its 12 figure entries, however, include files that the released code writes from three different analysis functions, and their code, caption and parameter fields are blank. The record therefore does not establish which run produced Figure 4. The Figure 4G bounds above show the published panel could not have been computed from `param`-recorded spikes.
- **Methods text (section 4.3).** It says the recorded population "was always restricted to the central circular patch (radius of 1 mm)", except for size tuning. It also says membrane potential and conductances were recorded "only from a subset of 190 neurons confined to the central 0.2 × 0.2 mm patch… and whose orientation preference was within 0.25 radians of horizontal". Read literally, both conflict with S5's spontaneous set. The 190, and the "2025 randomly selected neurons" given for the evoked protocols, equal the S5 evoked totals, so we read that section as describing the evoked recordings. The released files select on fixed grids rather than randomly or by orientation, so the text and files do not match exactly in either case.

## Method

- **Recorder probe.** `count_recorders.py` ran under `mpirun --oversubscribe -np 16`; the four-vCPU server needs `--oversubscribe` for 16 ranks, which the frozen launcher does not use. It used the pinned runtime and the installed protocol-006 model directory. Those 37 files matched the contract hashes in a check just before the probe, which was not retained, and in a dated re-check that is retained (`receipts/model-files-check.json`; contract SHA-256 `9613a2e9…`; newest file modified 7 October). Mozaik's normal workflow preparation loads `param_spont/defaults`, seeds PyNN with 5 and leaves Mozaik on its default 513. Model construction is stopped at the first connector call, when every sheet exists but no projection does. Both recipes' released recorder definitions are then applied with Mozaik's own `Sheet.setup_to_record_list()`. The two parameter trees differ only in their recorders.
- **Failed first attempt.** It failed in Mozaik's command-line parser before any model code ran: the `results_dir` value was not a quoted Python string, and Mozaik evaluates these values. It is kept in `receipts/attempt-01/`.
- **Figure measurement.** `measure_fig4g.py` was written during this investigation, then reviewed and re-run by us without code changes (`receipts/fig4g-measurement.txt`). It reads the unmodified PLOS large rendition (`sources/Figure-4-large.png`). `fig4g_bounds.py` converts the measured minima into the bounds above (`receipts/fig4g-bounds.json`).

## Supporting observations

- **Overlap.** The `param_spont` spike sets contain 836 of 838, 152 of 152, 856 of 862 and 173 of 173 of the `param` spike-recorded neurons. A `param_spont` run therefore supports a secondary comparison of spike-based measures on protocol 006's spike sample. Most of protocol 006's central voltage and conductance sample would not be recorded, so those measures have no comparable subset.
- **Recording and the network.** By our reading of the pinned code, recorders are set up only at the first stimulus presentation (`mozaik/models` lines 138-139), after the network is fully built. The cortical grid selectors, the only recorder settings that differ between the two recipes, draw no random numbers. The LGN selector does draw random numbers, but it is identical in both recipes. Switching recipes should therefore change which neurons are sampled, not the network. This is a source reading; no two full runs have been compared.
- **Data volume.** The spontaneous set records about 15 times as many spike trains (31,258 versus 2,025) and about 8 times as many voltage and conductance traces (1,597 versus 190 neurons, three signals each). Its memory and storage cost has not been measured yet.

## Limits

- The probe covers neuron positions and recorder selection only; no connections, stimuli or dynamics were built.
- Not every count is informative about positions. The two excitatory voltage counts of 400 equal the number of grid points (20 × 20 at 100 µm) and would match under almost any placement. The dense spike counts and the central evoked counts are the position-sensitive ones. That all of them match makes a different neuron placement unlikely, but we have not quantified this.
- The Figure 4 recipe is inferred, not recorded. The bounds assume the published panels used the released plotting code; the panel layout, axis ranges and 0.2-wide bins match that code.

## Decision

Protocol 006 is not edited. A new, versioned protocol will use the released `param_spont/defaults`, with comparison rules otherwise unchanged, after its memory and storage cost has been checked. The full-model run is still pending.

## Evidence map

- `count_recorders.py`: the probe, byte-identical to the copy that ran (SHA-256 `0d99d675…`).
- `receipts/recorder-counts.json` (counts, hashes, overlaps, seeds, S5 comparison), `receipts/probe.log.gz`, `receipts/time.txt` (15 min 54 s wall; 441 MiB maximum resident memory for the largest process), `receipts/model-files-check.json` and `receipts/attempt-01/`.
- `receipts/results-root/`: Mozaik's parameter, recorder and run records and the rank-0 log. These are copied unchanged from a results directory whose original name contains a colon; see `ORIGIN.txt`. The fifteen other per-rank logs repeat rank 0's selections and are retained privately.
- `measure_fig4g.py`, `fig4g_bounds.py`, `receipts/fig4g-measurement.txt`, `receipts/fig4g-bounds.json`.
- `sources/S5-table.png` (SHA-256 `bf022b39…`) with `sources/S5-table.json`, and `sources/Figure-4-large.png` (SHA-256 `d8d0df7f…`) with `sources/Figure-4.json`: unmodified PLOS files and their provenance.
- `figures/recorder-counts.png` and `.svg`, plotted by `draw_counts.py` from the counts receipt.
- `manifest.json` and `recipe-evidence.zip`.

## Attribution

S5 Table, Figure 4 and the 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, [doi:10.1371/journal.pcbi.1012342](https://doi.org/10.1371/journal.pcbi.1012342). The article and supplements are published under [CC BY 4.0](http://creativecommons.org/licenses/by/4.0/); the two images are reproduced unmodified. LSV1M v1.0 is under its MIT license. Mozaik v0.4.0 is under CeCILL. No animal recordings were collected in this work.
