# Felisyn protocol 013: spontaneous Vm synchrony and Vm autocorrelation against independent cat V1 recordings

Declared 23 October 2026 (Europe/Vienna), before any Vm synchrony, coherence or autocorrelation statistic of the model was computed. Population Vm means and SDs and spike-count correlations were already known from run 011's Figure 4 analysis. This is the project's first comparison of the full model with in vivo cat measurements that the model was not built on. It is descriptive, with no pass or fail threshold.

## Question

Run 011 reproduced the paper's own printed spontaneous-activity statistics (journal note 15). That is reproduction. Here we ask how the model's spontaneous membrane potential (Vm) compares with independent intracellular recordings from cat primary visual cortex:

1. **Synchrony.** How strongly do the Vm fluctuations of nearby layer 2/3 neurons co-vary? The target is paired whole-cell recordings in Yu & Ferster (2010).
2. **Time scale.** How quickly does a neuron's own Vm decorrelate? The target is Azouz & Gray (1999).

## Why these data

The model paper (Antolík et al. 2024) cites neither publication and shares no author with them. These are independent publications, though not independent laboratories: the model paper does cite other papers from the Ferster laboratory and one co-authored by Azouz and Gray. The model's own Vm and conductance calibration came from other recordings (its reference 44). The literature search behind this choice compared 22 sources by independence and kept quotes and hashes; it is a private working note in `work/bio-validation-013/`.

Both sources record from anaesthetized, paralysed cats:

- **Yu & Ferster:** ketamine and acepromazine for induction, then thiopental or propofol with sufentanil, and vecuronium. "Spontaneous" means blank-screen periods.
- **Azouz & Gray:** ketamine and acepromazine, then halothane in nitrous oxide and oxygen, and pancuronium. "Spontaneous" means pre-stimulus epochs.

The model's spontaneous condition corresponds to a gray blank screen.

## Data targets

### Yu & Ferster (2010)

Yu J, Ferster D (2010). Membrane potential synchrony in primary visual cortex during sensory stimulation. *Neuron* 68:1187-1201. doi:10.1016/j.neuron.2010.11.027. NIH author manuscript, PMC3031170.

**Recordings.** Pairs of complex cells in layer 2/3 (200-600 µm depth). The electrodes were placed 500-700 µm apart at the surface and angled 25° so that their tips converged; the cells were "separated by no more than 500 µm". The exact separations are not given.

**Method, which we follow:**

1. Take non-overlapping 1 s blocks of blank-screen data, remove spikes by interpolation and subtract the mean of each block.
2. Compute `R12(τ) = Σ Vm1(t+τ) Vm2(t) / sqrt(Σ Vm1² Σ Vm2²)` per block and average over blocks (Matlab `xcorr`).
3. Take the peak "within 10 ms of zero time lag" and the width at half height.
4. Repeat after a "20-Hz Butterworth" high-pass.
5. Compute coherence with 7 Slepian tapers on 1 s blocks.

**Population target.** Figure 4A (unfiltered) and 4B (above 20 Hz) plot, for each of 44 pairs, the peak correlation during visual stimulation (y) against the blank period (x). We digitized the x values with `digitize_fig4.py`:

1. The figure image is checked by SHA-256. It is fetched from PMC and not redistributed.
2. Each panel is calibrated from its own tick marks; the largest residual is 0.0024.
3. Filled symbols are found with a disk template. Open symbols are found with an annulus template, accepted greedily on ring pixels not yet explained by other symbols.
4. Every detection was checked by eye against zoomed overlays, which are kept privately. Each one lies on a visible ring or ring fragment. Some in the dense clusters are ambiguous:
   - three green detections 3.2 to 4.4 px apart in panel A's lower cluster;
   - a double red outline with one detection, and a possible duplicate between two filled disks in panel B.
5. Detections are not edited by hand.
6. **Symbols not found** are hidden behind found ones, so each hidden centre lies within 11.5 px of some detection (a symbol diameter plus half a pixel, about 0.081). Quantile bounds for all 44 pairs therefore place every hidden symbol at the lowest detected x value minus that margin, or at the highest plus it. Candidate centres are local maxima over a window; ties on a coverage plateau are resolved in raster order, which can shift a centre by up to 0.75 px (about 0.005).
7. A rerun is byte-identical with the recorded Python, NumPy, SciPy and Pillow versions.

| Blank (spontaneous) peak correlation | Found | Of which clustered | Median of found (range) | Median of all 44, bounds | 25th percentile, lower bound | 75th percentile, upper bound |
| --- | --- | --- | --- | --- | --- | --- |
| Fig. 4A, unfiltered | 37 of 44 | 30 | 0.722 (0.403-0.908) | 0.704 to 0.754 | 0.556 | 0.862 |
| Fig. 4B, above 20 Hz | 32 of 44 | 29 | 0.332 (0.189-0.682) | 0.270 to 0.402 | 0.108 | 0.762 |

**Example values from the text:**

- An example complex pair had a peak of 0.66 and a width of 54 ms, and 0.30 above 20 Hz (Figure 1C).
- A layer 4 simple cell paired with a layer 2/3 complex cell, with the electrode tips "close to one another in the horizontal direction", had "a low correlation (0.4 ...)" (Figure 7C).
- Spontaneous coherence "declined as a function of frequency" (Figure 1F; all pairs in Figure 4D), and it did not depend on the pairs' relative orientation preference.

### Azouz & Gray (1999)

Azouz R, Gray CM (1999). Cellular mechanisms contributing to response variability of cortical neurons in vivo. *J Neurosci* 19:2209-2223. doi:10.1523/JNEUROSCI.19-06-02209.1999.

**Method:**

- Spikes are replaced by the threshold voltage, the trace is resampled at 1 kHz and smoothed with a 3-point median filter.
- The autocorrelation is computed over ±100 ms at 1 ms on 500 ms spontaneous epochs.
- The fit is `y = a + b·e^(−x/c)`.

**Target.** "mean τ_spontaneous = 38 ± 26 msec" (Fig. 4D). There are 52 cells of mixed classes, the layers are not stated, and the type of dispersion is not stated.

## Model data and selection

**Run 011** (`full-20261022T061230Z`) supplies the 40.32 s spontaneous recording:

- Voltage at 1 ms for 400, 399, 400 and 398 neurons (layer 4 E, layer 4 I, layer 2/3 E, layer 2/3 I). These exports stay on the research server, with their SHA-256 in run 011's `data-on-server-sha256.txt` and in `contract.json`.
- Spike times come from run 011's published `exported-spikes.json.gz`. Every analog-recorded neuron has a spike record.

**Positions.** `export_positions.py` read the positions read-only from the saved datastore before this declaration was frozen (`positions.json`, geometry only). Mozaik stores positions in degrees of visual field, and every cortical sheet uses 1000 µm per degree. As a unit check, the median nearest-neighbour distance among recorded neurons is 74-87 µm, consistent in scale with the recorder's 100 µm grid.

**Primary pairs.** All pairs of layer 2/3 excitatory neurons at most 500 µm apart: 12,074 pairs, with a median separation of 333 µm. Yu & Ferster recorded complex cells, which are mostly excitatory, at most 500 µm apart. Results are also given in 100 µm distance bins.

**Secondary pairs:**

- layer 2/3 E with layer 2/3 I;
- layer 2/3 I with layer 2/3 I;
- layer 4 E with layer 2/3 E, also restricted to 200 µm or less for the simple-complex example.

**Primary τ set.** The 800 excitatory analog-recorded neurons (layers 4 and 2/3) pooled, because sharp intracellular samples like Azouz & Gray's are dominated by excitatory cells. All 1,597 neurons pooled, and each population, are also reported.

## Computation (`vm_synchrony.py`)

**1. Inputs.** The SHA-256 of every input is checked against `contract.json`.

**2. Spike removal**, using the model paper's own rule. Samples within 2.5 ms of each spike time are replaced by linear interpolation between the samples just outside.

**3. Pairs, following Yu & Ferster:**

- Use the first 40 non-overlapping 1 s blocks, starting at the beginning of the spontaneous window; the last 0.32 s is unused. No block is dropped, whatever its content.
- Remove the mean per block and compute `R12(τ)` exactly as above, averaged over blocks.
- Take the peak within ±10 ms. Take the full width at half that peak by linear interpolation; it is "not reached" if the half height is not crossed within ±500 ms or the peak is not positive.
- Repeat after a high-pass of the whole trace. The cutoff is Yu & Ferster's 20 Hz Butterworth; order 4 and zero phase are our assumptions.
- Coherence: 7 Slepian tapers (NW = 4) per 1 s block, spectra averaged over blocks and tapers, then `|Sxy| / sqrt(Sxx Syy)`. This is raw coherence, not the Fisher-transformed values Yu & Ferster used for changes in their Fig. 4F-I. Band means are taken at 1-10 Hz and 20-80 Hz. Bins below 4 Hz lie within the taper half-bandwidth, and 0 Hz is excluded because block means are removed.

**4. τ, following Azouz & Gray:**

- After spike removal, apply a 3-point median filter.
- Use 80 non-overlapping 500 ms epochs and remove the mean per epoch.
- Compute the autocorrelation at 0-100 ms lags, scaled to 1 at lag 0 and averaged over epochs.
- Fit `a + b·e^(−x/c)` by least squares on lags 0-100 ms; τ = c.
- Fits that fail, or end within 1% of a bound on b or c, are counted as failures and excluded.

**5. Summaries:**

- median, IQR, 10th and 90th percentiles and range;
- a bootstrap 95% interval for each median (2,000 resamples, seed 13). Pairs share neurons, so this interval is descriptive;
- medians by distance bin and the group-mean coherence spectrum;
- for the primary group, the fraction of model pairs inside the found data range, and the percentile of the data median within the model distribution;
- the script and contract SHA-256 and the software versions.

**Declared deviations from the sources:**

| Step | Source | Here | Why |
| --- | --- | --- | --- |
| Sampling | 20 kHz resampled to 4,096 Hz (Yu & Ferster) | 1 kHz model samples | the model was recorded at 1 ms |
| Line noise and smoothing | Chronux line-noise removal, Savitzky-Golay smoothing | none | the model has no line noise or recording noise, and the smoothing parameters are not given |
| High-pass | "20-Hz Butterworth" | 4th order, zero phase | order and phase not given |
| Spike removal for τ | replaced by the threshold voltage (Azouz & Gray) | interpolation over ±2.5 ms (primary); hold at the last pre-spike sample (sensitivity) | one spike rule for both measures, following the model paper |
| Autocorrelation normalisation | not stated | unbiased per lag (primary); 'coeff' (sensitivity) | unbiased is less tilted by the 500 ms epochs |

**Estimator bias** on synthetic Ornstein-Uhlenbeck traces. Both τ estimators come out low for long time scales, because the mean is removed from each short epoch:

| True τ | Unbiased estimate | 'coeff' estimate |
| --- | --- | --- |
| 10 ms | 10.4 | 9.7 |
| 30 ms | 27.5 | 24.5 |
| 60 ms | 50.1 | 42.6 |

**Tests.** `test_vm_synchrony.py` (13 tests) checks:

- spike interpolation and hold;
- the FFT cross-correlation against the direct formula;
- the expected peak (0.5) and width for a shared exponential component;
- the high-pass removing a shared 2 Hz component;
- coherence of identical and independent signals;
- τ recovery and fits stuck at a bound;
- distance selection and the outcome readings;
- a full run of `main()` on fake files, including rejection of a modified input.

**Resources.** Peak memory is about 1.5 GB. It runs as a systemd unit with MemoryMax=3G on the 7.8 GB server.

## Declared reading of outcomes

These categories describe; they are not pass or fail. The bounds already allow for the hidden symbols.

**Synchrony, primary pairs.** Each measure gets a category from the model median:

| Measure | "Below the data" | "Within the data" | "Above the data" |
| --- | --- | --- | --- |
| Unfiltered peak correlation | under 0.556 | 0.556 to 0.862 | over 0.862 |
| Above-20 Hz peak correlation | under 0.108 | 0.108 to 0.762 | over 0.762 |

The two categories together give one of nine fixed readings, in `contract.json`. For example:

- below and below: "less synchronized than the recordings at slow and at fast time scales";
- below and within: "overall synchrony weaker than in the recordings; fast synchrony comparable".

**Interpretation, fixed in advance:**

- Yu & Ferster's spontaneous coherence is highest below 10 Hz. Anaesthesia commonly produces slow, synchronized Vm fluctuations, while the model runs in an asynchronous state. A model below the data on the unfiltered measure is therefore not, by itself, a model failure: it could reflect a difference in state, in shared input, or both, and these data cannot separate them.
- The above-20 Hz measure is less affected by slow fluctuations, but it carries a bias of its own. Recordings contain uncorrelated electrode noise that lowers their high-frequency correlation, and the model traces contain none. So on this measure the model is favoured: a model below the data here points to weaker fast shared input, while a model above it may partly reflect the missing noise.
- A model within the data on the unfiltered measure is reported as consistent, and a model above it as more synchronized than the recordings. In both cases the state caveat applies in reverse: an asynchronous model would be expected to sit below anaesthetized data.

**τ.** The primary median is compared with 38 ± 26 ms. Within one dispersion of the mean reads as "consistent"; outside it reads "shorter" or "longer than the data". This compares a model median with a data mean ± dispersion of unknown type, so it is coarse. The 'coeff' and hold variants show how much the estimator matters.

**Also reported, without a category:** widths, coherence band means and spectrum shape, the secondary groups, the distance dependence, the example pairs, and the layer 4 E to layer 2/3 E pairs within 200 µm against the 0.4 simple-complex pair.

## Changes from the approved spec, made before freezing

- The category limits use the 25th-percentile lower bound and the 75th-percentile upper bound of the data, not its 10th percentile, because the hidden symbols make the 10th percentile uninformative.
- The low coherence band starts at 1 Hz, not 0 Hz (block means are removed).
- The τ fit uses lags 0-100 ms.
- These were added after the pre-freeze review:
  - the primary τ set and its coded outcome;
  - the nine combined readings;
  - the placement outputs, fit-bound failures, the hold and 'coeff' sensitivities, and provenance in the results;
  - hidden-symbol bounds, first from the clustered detections and, after the second review, from all detections, because a symbol can hide exactly under an isolated one;
  - neutral wording for the readings ("could reflect", "weaker").
- The second review also corrected the opening claim: population Vm means and SDs were already known from run 011.

## What this cannot show

- One model run, one state, one 40 s window. The data come from anaesthetized, paralysed animals, in other cells and other cortical states.
- No biological "validation" in the sense of a pass. Agreement would show consistency on these measures only, and disagreement has the competing explanations above.
- The digitized population is incomplete (37 and 32 of 44) and pixel-limited (about 0.007 per pixel). The bounds carry this.
- Yu & Ferster's pair separations and cell classes do not map exactly onto the model's selection.
- In the model, the −60 mV reset and the refractory hold last longer than the 2.5 ms removal window, so a post-spike sawtooth remains in the cleaned traces. A biological afterhyperpolarization looks different. This mainly affects neurons that fire often.

## Order of work

1. Literature and data search (23 Oct, private).
2. Digitization of Figure 4.
3. Position export (geometry only).
4. Two pre-freeze reviews and the changes above.
5. This declaration and `contract.json`, frozen with their SHA-256 recorded in the project handoff.
6. Tests on the server.
7. Only then `vm_synchrony.py` on the voltage exports.

The results go into a separate package that records this protocol's hash.

## Files

- `README.md`: this declaration.
- `contract.json`: input, target and script hashes, all parameters and the declared readings.
- `digitize_fig4.py` and `targets/`: the digitized blank values (`yu-ferster-fig4-digitized.csv`) and their summary. The figure is not redistributed; the script fetches it from PMC and checks its hash.
- `export_positions.py`, `positions.json` and `export-positions.log`.
- `vm_synchrony.py` and `test_vm_synchrony.py`.
- `manifest.json` and `protocol-evidence.zip`.

## Attribution

**Data targets.** Yu & Ferster (2010), *Neuron* (NIH author manuscript), and Azouz & Gray (1999), *J Neurosci*. Short quotations are for scholarly comparison. The numbers digitized from Figure 4 are our measurements of the published figure.

**Model.** Antolík et al. (2024), *PLOS Computational Biology*, CC BY 4.0; LSV1M v1.0 (MIT); Mozaik (CeCILL).

**Model data.** Our run 011.
