# Felisyn comparison 019: shared spontaneous fluctuations are far weaker in the full model than in cat area 17

Results of [protocol 018](https://www.felisyn.com/research/protocol-018/README.md). The protocol was frozen at 2026-10-23T21:32:20Z: README SHA-256 `9948e3f15b62f57d9e256c4f1c0891d31131181c6ff865fa8d9268e66886b321`, contract `73ff40842105dd436295e4bed52d39c10397d8c8f7667a21bf9cfa358169f2a1`. Its unchanged `compare.py` ran from the frozen package from 21:32:45 to 21:39:11 UTC (`results/compare.log`), after checking the SHA-256 of all three inputs. Nothing was changed after the freeze.

## Question

How strongly do separate groups of neurons in the full model (run 011, 40 s of spontaneous activity on a grey screen) fluctuate together within a second, compared with spontaneous multiunit activity in cat area 17 (Hahn et al. 2017; four isoflurane-anaesthetized cats in darkness)? The measure is described in the protocol. In short:

- It is the normalised covariance of firing between different recording sites, or between model neuron pools, in 100 ms bins.
- It is measured around each second's own mean and keeps only cross-unit terms.
- It does not change when spikes are deleted at random, so it does not depend on the detection threshold.

## Declared result

**Primary reading: "less shared fluctuation than desynchronized (1-2) cat cortex".**

| | Shared fluctuation at 100 ms | 95 % interval |
| --- | --- | --- |
| **Model**, primary pools (18 pools, 400 µm apart, both layers) | **0.041** | 0.031 to 0.053 |
| Cat 1, desynchronized | 2.62 | 2.43 to 2.82 |
| Cat 2, desynchronized | 0.341 | 0.318 to 0.364 |
| Cat 3, desynchronized | 0.404 | 0.385 to 0.424 |
| Cat 4, desynchronized | 0.834 | 0.791 to 0.878 |

The model's interval lies entirely below the range of the four cats' intervals (0.318 to 2.82). It is 8 times below the cat with the lowest value (cat 2) and 64 times below the cat with the highest (cat 1).

**The same reading holds everywhere it was declared to be checked:**

- **For every state.** The model is below the range for each of the five cat states separately, and for the synchronized states pooled (`results/results.json`, `readings`).
- **For every model variant** (`robust_to_model_units`: true):

| Model variant | Value | 95 % interval |
| --- | --- | --- |
| pools, excitatory neurons only | 0.059 | 0.041 to 0.082 |
| pools, inhibitory neurons only | 0.028 | 0.020 to 0.037 |
| pools 800 µm apart | 0.020 | 0.015 to 0.026 |
| single neurons within 500 µm, all pairs | 0.071 | 0.062 to 0.080 |
| primary pools, first 2 s dropped | 0.041 | 0.030 to 0.054 |

- **With the artifact candidates removed** (`robust_to_coincidence_removal`: true; see below).

**The model's value is real.** With its spikes redistributed within each second (the counting-noise null, reported but not read), it falls to 0.0000 (95 % range ±0.0006). So the model's neuron groups do fluctuate together, but far more weakly than the cats' recording sites.

## Artifact check (data only)

Near-simultaneous threshold crossings on several contacts would add shared fluctuation to the data. The protocol therefore declared two checks, in ticks of the data's 0.1 ms resolution.

| Cat | Pair density at ±0.2 ms relative to 1-5 ms | Spikes in events with 3 or more sites | Desynchronized value with those events removed |
| --- | --- | --- | --- |
| 1 | 1.76 | 5.5 % (1,292 events) | 2.33 (2.16 to 2.52), against 2.62 |
| 2 | 1.16 | 0.04 % | 0.340, against 0.341 |
| 3 | 1.90 | 0.37 % | 0.398, against 0.404 |
| 4 | 1.11 | 0.11 % | 0.831, against 0.834 |

- **Cat 1** has clearly more near-simultaneous multi-site events than the others. They are more frequent in its synchronized states (7.3 % of spikes) than in its desynchronized ones (3.9 %). These data cannot say whether they are artifacts or genuinely precise synchrony.
- **Removal does not change the reading.** It lowers cat 1's desynchronized value by 11 % and the others by less than 2 %.

## Secondary measures (reported only)

**Time scale.** The primary measure computed in bins of 10, 50, 100 and 250 ms (`figures/shared-fluctuations-timescale.png`):

| Bin | Model | Cats, desynchronized (range of the four) |
| --- | --- | --- |
| 10 ms | 0.54 | 0.65 to 6.2 |
| 50 ms | 0.075 | 0.46 to 3.9 |
| 100 ms | 0.041 | 0.34 to 2.6 |
| 250 ms | 0.014 | 0.25 to 1.4 |

Both fall as the bins get longer, but the model falls much faster.

- At 10 ms the model's value (0.54) is close to cats 2 and 3 (0.65 and 0.73).
- At 250 ms it is 18 to 99 times lower than the cats.

In other words, what the model's groups share is concentrated at short time scales, while the cats' shared fluctuations extend over hundreds of milliseconds. This was not a declared reading. It matches what note 17 found in the membrane potential: the model's spontaneous voltage decorrelates about eight times faster than in the recordings, and its pair correlations are narrow.

**Other secondary values:**

- **Pooled excess variance** of the summed count, which includes each unit's own variance: model 0.067, cats 0.42 to 2.5.
- **Whole-session excess variance** on stable sites, at 100 ms: model 0.069, cats 1.1 to 6.3.
- **40-second draws:** the model lies at the 0.04th percentile of the cats' 40-second draws of the desynchronized states.

All of these point the same way.

## Interpretation, as fixed in advance

- **Agreement with note 17.** A model below the desynchronized range agrees with note 17. It is not by itself a model failure: anaesthesia, darkness and multiunit sampling all differ from the model's condition, and these data cannot separate them.
- **A reference selected to sit low.** The desynchronized seconds were chosen by the authors' LFP clustering for low synchronization, so that reference sits low by construction. The model is still below it, by a factor of at least 8 at 100 ms.

## What this cannot show

The protocol's limits apply.

- **One model run in one state:** 40 s of a grey screen against hours of anaesthetized cat cortex in darkness, with states defined by the LFP.
- **Not validation:** this is a consistency check on one family of measures.
- **Detection:** spike detection is only approximately random deletion. Contamination with independent noise crossings would lower the data, and common-mode events would raise it.
- **Unknown geometry:** contact depths and shank spacing are not known. The model variants span the declared choices, and the reading did not depend on them.

## Files

- `README.md`: this summary.
- `results/results.json`: everything `compare.py` computed:
  - per cat and state group: values, bootstrap intervals, nulls, curves, secondary measures, the artifact check and cleaned estimates;
  - the reference ranges;
  - every model variant;
  - all readings.
- `results/compare.log`, `results/start-utc.txt` and `results/end-utc.txt`: the run receipt.
- `figures/shared-fluctuations.png` and `.svg`: the primary measure per state, cats and model.
- `figures/shared-fluctuations-timescale.png` and `.svg`: the time-scale curves (secondary).
- `draw_results.py`: plotting code, which reads only `results/results.json`.
- `manifest.json` and `comparison-evidence.zip`.

## Attribution

**Data.** Hahn G, Ponce-Alvarez A, Monier C, Benvenuti G, Kumar A, Chavane F, Deco G, Frégnac Y (2017), *Spontaneous cortical activity is transiently poised close to criticality*, PLoS Comput Biol 13: e1005543, S5 Dataset, CC BY 4.0.

**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](https://doi.org/10.1371/journal.pcbi.1012342), published under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). LSV1M v1.0 is under its MIT license, and the Mozaik simulation framework under CeCILL.

**Model data and figures.** Our run 011. The figures are new plots of the retained results.
