# Felisyn comparison 014: the model's spontaneous membrane potential against independent cat recordings

Recorded 23 October 2026 (Europe/Vienna). This is the result of [protocol 013](https://www.felisyn.com/research/protocol-013/README.md), which was frozen at 01:59:46 UTC with its SHA-256 recorded first. The analysis ran from 02:00:23 to 02:06 UTC on the research server.

It is the project's first comparison of the full model with in vivo cat measurements that the model was not built on. The readings below are the ones the protocol fixed in advance. They describe and do not pass or fail.

## Result in brief

The model's spontaneous membrane potential (Vm) runs on a different time scale from that of cats under anaesthesia:

- **Time scale.** A model neuron's Vm decorrelates with a time constant of about 4.5 ms (all 1,597 neurons: 3.8 to 6.0 ms). Azouz & Gray report 38 ± 26 ms. Declared reading: **shorter than the data**.
- **Synchrony, unfiltered.** For nearby layer 2/3 excitatory pairs (up to 500 µm apart), the median peak Vm correlation is 0.37. Yu & Ferster's pairs have a median of 0.72. Declared reading: **below the data**. In the model, synchrony falls steeply with distance: pairs under 100 µm apart reach 0.72, and pairs 400 to 500 µm apart reach 0.28. The separations of the recorded pairs are not reported.
- **Synchrony above 20 Hz.** The model median is 0.41 and the recorded median 0.33. Declared reading: **within the data**.
- **Combined reading:** "overall synchrony weaker than in the recordings; fast synchrony comparable".

Three further measures point the same way:

- the model's cross-correlations are about 8 ms wide, against 54 ms for Yu & Ferster's example pair;
- its coherence peaks near 38 Hz, whereas the recordings' spontaneous coherence falls with frequency;
- after the 20 Hz high-pass the model's median rises from 0.37 to 0.41, while the recorded median falls from 0.72 to 0.33 (visible symbols).

None of this is a model failure by itself. Anaesthesia commonly produces slow, synchronized Vm fluctuations, while the model runs in an asynchronous state. As the protocol declared, the difference could reflect state, shared input, or both, and these data cannot separate them.

## Provenance

- **Inputs.** The run checked every input against `contract.json` of protocol 013: the four voltage exports and the spike export of run 011, and `positions.json`. The results record `script_matches_contract: true`, the frozen contract SHA-256 and Python 3.9.20, NumPy 1.22.3 and SciPy 1.8.0.
- **Frozen protocol hashes** (recorded at 01:59:46 UTC): contract.json `f82f27ac4ab90b4be15b92bc1becc911ff4731130855d65c78c3f05fe0af157c`, README.md `d30c501b922026d3b4a5ac538b86622dac0563572e6d477cce8b1564cc3b9785`, manifest.json `a1acb8e49a48aebd702b93600929fc5fab00fc45b841ffa413ca01ae25b0d4d5`.
- **Checks.** Before the run, the manifest check and the 13 synthetic tests were run on the server and passed; that console output was not saved. Both were rerun after the analysis on the unchanged frozen files (`receipts/post-run-recheck.txt`).
- **Resources.** 6 min 27 s, with a peak resident size of 1.39 GiB under MemoryMax=3G. Exit 0.
- **Files.** `results/vm-synchrony-results.json.gz` decompresses to the SHA-256 recorded on the server (`receipts/results-sha256.txt`). The run log and the exact unit command are in `receipts/`.

## Results

### Synchrony (Yu & Ferster 2010)

Peak Vm cross-correlation within ±10 ms, block-averaged:

| Group | Pairs | Unfiltered median (IQR) | Above 20 Hz median (IQR) | Width at half height (ms) | Coherence 1-10 Hz / 20-80 Hz |
| --- | --- | --- | --- | --- | --- |
| **Layer 2/3 E with E (primary)** | 12,074 | **0.372** (0.293-0.508) | **0.410** (0.328-0.548) | 8.3 | 0.28 / 0.42 |
| Layer 2/3 E with I | 24,735 | 0.371 (0.297-0.491) | 0.411 (0.337-0.530) | 8.6 | 0.28 / 0.42 |
| Layer 2/3 I with I | 12,146 | 0.402 (0.333-0.515) | 0.456 (0.385-0.565) | 8.6 | 0.29 / 0.46 |
| Layer 4 E with layer 2/3 E | 24,615 | 0.115 (0.090-0.146) | 0.156 (0.128-0.190) | 7.7 | 0.06 / 0.17 |
| *Recordings, Fig. 4A/B (digitized)* | *37 and 32 of 44 visible* | *0.722 (all 44: 0.704 to 0.754)* | *0.332 (all 44: 0.270 to 0.402)* | *54 (one example pair)* | *falls with frequency* |

For the primary pairs, the bootstrap 95% intervals of the medians are 0.369-0.376 (unfiltered) and 0.407-0.413 (above 20 Hz). Pairs share neurons, so these intervals are descriptive.

**Placement.**

- Unfiltered: 42% of model pairs lie inside the visible data range (0.403-0.908), and the recorded median falls at the 98th percentile of the model pairs.
- Above 20 Hz: 91% of model pairs lie inside the range (0.189-0.682), and the recorded median falls at the 27th percentile.

**Distance dependence** (primary pairs, median peak):

| Separation (µm) | 0-100 | 100-200 | 200-300 | 300-400 | 400-500 |
| --- | --- | --- | --- | --- | --- |
| Pairs | 358 | 1,537 | 2,724 | 3,354 | 4,101 |
| Unfiltered | 0.722 | 0.645 | 0.494 | 0.364 | 0.275 |
| Above 20 Hz | 0.758 | 0.682 | 0.533 | 0.402 | 0.309 |

**Simple-complex comparison.** Layer 4 E with layer 2/3 E pairs up to 200 µm apart (4,203 pairs) have a median of 0.168. Yu & Ferster's single simple-complex pair had 0.4.

**Coherence.** The group-mean coherence of the primary pairs averages 0.32 at 1-10 Hz, rises to 0.58 at 38 Hz and falls to 0.29 at 100 Hz. Yu & Ferster's spontaneous coherence "declined as a function of frequency" (their Fig. 1F and 4D; not digitized).

### Time scale (Azouz & Gray 1999)

| Set | Neurons | τ median (IQR), ms | Range, ms | Fits |
| --- | --- | --- | --- | --- |
| **Excitatory, layers 4 and 2/3 (primary)** | 800 | **4.52** (4.14-5.05) | 3.78-5.97 | all 800 succeeded |
| All analog-recorded | 1,597 | 4.49 (4.12-4.98) | 3.76-5.97 | all succeeded |
| *Recordings* | *52* | *mean 38 ± 26* | | |

Per population, the medians are 5.05 ms (layer 4 E), 4.92 (layer 4 I), 4.14 (layer 2/3 E) and 4.11 (layer 2/3 I). The declared sensitivity variants agree within 0.16 ms of these: the 'coeff' normalisation, and spike removal by holding the pre-spike value. The bootstrap 95% interval of the primary median is 4.40-4.62 ms. Every one of the 1,597 model neurons lies below the lower end of the data's one-dispersion range (12 ms).

## Reading, as declared

- **Unfiltered: below the data.** The model median (0.372) lies under the declared limit of 0.556. Read as: the model's spontaneous Vm is less synchronized than in anaesthetized cat layer 2/3 for pairs up to 500 µm apart. This is not by itself a model failure. It could reflect a difference in state, in shared input, or both.
- **Above 20 Hz: within the data.** The model median (0.410) lies between 0.108 and 0.762. On this measure the model is favoured, because recordings contain electrode noise and the model does not. All 358 model pairs under 100 µm apart (median 0.76) exceed every visible recorded pair (maximum 0.68).
- **τ: shorter than the data.** The model median (4.52 ms) lies more than one dispersion below the mean of 38 ms. This compares a model median with a data mean ± dispersion of unknown type, so it is coarse, but the gap is large: no model neuron comes within it.

## Context, not part of the declared comparison

**The short time constant is consistent with the model's own parameters.**

- Capacitances are 32 pF (excitatory) and 30 pF (inhibitory). The leak is 4.0 and 3.3 nS (membrane time constants of 8 and 9 ms). Synaptic time constants are 1.5 ms (excitatory) and 4.2 ms (inhibitory). All come from run 011's saved parameters.
- Adding the mean spontaneous conductances measured in run 011 (`run-011/analysis/analysis-parity.json`) gives an effective membrane time constant `C / (g_leak + g_exc + g_inh)` of about 3.8 ms (layer 4 E), 2.9 ms (layer 2/3 E), 4.7 ms (layer 4 I) and 3.8 ms (layer 2/3 I).
- The measured Vm time constants of 4.1-5.1 ms are of that order, though somewhat longer and not in the same order across populations, so this is a consistency check rather than a derivation. The model's Vm follows fast synaptic input. Part of that input is shared, enough to correlate close pairs at 0.7 or more, but it varies over milliseconds rather than tens of milliseconds.

**Distance.** The recorded pairs may have been closer together than the model's median separation of 333 µm. That would explain part of the unfiltered gap, since model pairs under 100 µm apart reach the recorded median (0.72). It cannot explain all of it: above 20 Hz, all 358 of those close model pairs exceed every visible recorded pair, and their correlations are only about 8 ms wide, against 54 ms in the recorded example.

**The 38 Hz coherence peak** suggests shared input concentrated near 38 Hz. We did not check for a matching peak in the power spectrum or analyse its origin.

## Limits

- One model run, one state, one 40 s window. The recordings come from anaesthetized, paralysed cats, from other neurons and in other cortical states.
- The recorded population was digitized from a figure: 37 and 32 of 44 symbols are visible, at about 0.007 per pixel. The declared limits allow for the hidden symbols.
- Pair separations and cell classes in the recordings are not known exactly. The τ comparison is between a model median and a data mean ± dispersion.
- In the model, the recovery from the −60 mV reset outlasts the 2.5 ms spike-removal window (the refractory periods themselves are 2.0 ms for excitatory and 0.5 ms for inhibitory neurons). This leaves a post-spike sawtooth in the cleaned traces. The 'hold' variant changes τ by at most 0.16 ms.
- No awake or desynchronized-state recordings were compared. This does not show how the model would compare with them.
- The model reproducing the paper's numbers (run 011) and differing from these recordings are separate results; neither is a full biological validation.

## Files

- `results/vm-synchrony-results.json.gz`: all results, including per-pair values for the primary group and τ per neuron.
- `receipts/`: the run log with GNU time, start time, exit code, the unit command, the result hashes, and the post-run recheck of the frozen files.
- `draw_comparison.py` → `figures/synchrony.png`/`.svg` and `figures/time-scales.png`/`.svg`. These plot retained data only: model values, our digitized blank values, and the Azouz & Gray text value. They were rendered locally with Python 3.14.3 and Matplotlib 3.10.8.
- `manifest.json` and `comparison-evidence.zip`.

## Attribution

- **Recordings.** Yu J, Ferster D (2010), *Neuron* 68:1187-1201, doi:10.1016/j.neuron.2010.11.027 (NIH author manuscript PMC3031170). Azouz R, Gray CM (1999), *J Neurosci* 19:2209-2223, doi:10.1523/JNEUROSCI.19-06-02209.1999. The digitized values are our measurements of the published Figure 4; the figure itself is not reproduced.
- **Model.** Antolík J et al. (2024), *PLOS Computational Biology*, CC BY 4.0; LSV1M v1.0 (MIT); Mozaik (CeCILL).
- **Model data.** Our run 011.
