# Felisyn reference check 012: two ways to average the conductances

Recorded 23 October 2026 (Europe/Vienna). Descriptive check of published numbers; no new simulation and no new analysis of recordings.

**Result.** The paper's text gives average conductances of 1.15 nS (excitatory) and 3.56 nS (inhibitory). Both values are reproduced from our run when every population counts about equally. That can be done by pooling over the recorded neurons (about 400 per population, the paper's own recording counts) or by taking the plain mean of the four population means. The paper's Figure 4 code weights the populations 1 : 0.25 : 1 : 0.25, matching their sizes, which gives 1.185 and 4.093 nS. A difference in averaging is therefore the most likely explanation of the discrepancy recorded in benchmark-004 (about 4.09 nS against 3.56 nS). This is an inference: the code that produced the text values is not public, and the comparison rests on one run.

## Question

The paper reports the conductance averages in two places:

- Its Figure 4 code logs population-weighted means, `(L4E + 0.25 L4I + L23E + 0.25 L23I) / 2.5`.
- Its text says:

> As shown in Figs 3B, 4E and 4F, the excitatory synaptic conductances recorded in model neurons during spontaneous activity (1.15 nS, averaged across all layers) are consistent with levels observed in cat V1 (∼ 1 nS; layer origin not known) [44].

> The inhibitory conductance during spontaneous activity was 3.56 nS when averaged across all model neurons, slightly lower than the value of 4.9 nS reported in cat [44].

Benchmark-004 applied the code's weights to the inhibitory conductances printed in Figure 4 and got about 4.09 nS. The gap exceeds the rounding of the printed values. Protocols 006 and 007 carried the question as `open_reference_issue`. Which averaging produces the numbers in the text?

## Sources

| Input | Identity |
| --- | --- |
| Article text | Europe PMC full-text XML retrieved on 7 October 2026 (audit-001), SHA-256 `0ffddca8…`. A copy is in `sources/antolik-2024.xml`. |
| Figure 4 code | LSV1M v1.0 `visualization_functions.py`, lines 538 and 549, SHA-256 `1f1bc47d…` (protocol 007's pinned model). |
| Our population means | run-011 `analysis/analysis-parity.json`, SHA-256 `affb604c…`. Our analysis code and the authors' Figure 4 code agree on these means to within 9 × 10⁻¹⁶. The voltage and conductance recordings cover 400, 399, 400 and 398 neurons. |
| Printed Figure 4 values | benchmark-004 `reference-means.csv`, SHA-256 `fb460c31…`. |

## Method

`average_conductances.py` checks every input against its recorded SHA-256 and stops on a mismatch. It extracts the two sentences and the two code lines verbatim. It then combines the four population means of each conductance in three ways:

1. **Weighted by population size** (37,500 : 9,375 : 37,500 : 9,375 neurons, as built in run 011 and recipe-check-009). This is the same number as the code's formula, which the program confirms.
2. **Pooled over the recorded neurons**, weighting each population by its recorded neuron count. The counts, 400/399/400/398, equal the paper's S5 Table spontaneous counts (protocol 007 contract, from recipe-check-009).
3. **Unweighted mean** of the four population means.

A result matches a value in the text if it lies within half a unit of the value's last digit (±0.005 nS). The same rules are also applied to the printed Figure 4 values, which carry ±0.05 nS of rounding.

For scale, the program also estimates how much each average would move if a different set of neurons of the same size had been recorded. This is one standard error, computed from the across-neuron SDs. It does not cover differences between network realizations or simulator versions.

## Results

| Averaging rule | Excitatory (nS) | Inhibitory (nS) | Matches 1.15 and 3.56 nS? | From the printed Figure 4 values |
| --- | --- | --- | --- | --- |
| Weighted by population size (the code's logged mean) | 1.185 | 4.093 | Neither | 1.18 / 4.09 |
| Pooled over the recorded neurons | 1.153 | 3.563 | Both | 1.15 / 3.55 |
| Unweighted mean of the four populations | 1.153 | 3.562 | Both | 1.15 / 3.55 |

One sampling standard error is ±0.0035 nS (excitatory) and ±0.011 nS (inhibitory) for the equal-weight averages, and ±0.005 and ±0.015 nS for the size-weighted ones. Full-precision values, the per-population means and SDs, and the extracted text are in `conductance-averages.json`.

The printed values alone already point the same way. With ±0.05 nS of rounding, the size-weighted inhibitory average (4.09 nS) cannot give 3.56 nS, while the unweighted one (3.55 nS) can. Benchmark-004 tested only the weighted version. Our run's full-precision means make the test sharper: the size-weighted excitatory average, 1.185 nS, no longer rounds to 1.15 nS.

## Interpretation

Both values in the text are reproduced by an average in which each population counts about equally: either pooled over the recorded neurons, or the plain mean of the four population means. The two differ by at most 0.0012 nS, so this check cannot tell them apart. Weighting the populations by size reproduces neither value. "Averaged across all model neurons" most likely refers to the recorded model neurons, not to all neurons in the model.

This remains an inference. The code that produced the text values is not available, and other averaging rules could round to the same numbers. The case rests on two observations:

1. The printed values alone exclude size weighting for the inhibitory value (4.09 ± 0.05 nS against 3.56 nS).
2. With our run's full-precision means, the equal-weight rules round to both text values and size weighting rounds to neither. Size weighting misses the excitatory value by 0.035 nS, about seven sampling standard errors.

The three-digit agreement of the equal-weight rules is consistent with them but is not strong evidence by itself, because one sampling standard error of the inhibitory average (±0.011 nS) exceeds the rounding tolerance. The second observation also assumes that our single run is close to the paper's run, which the 24 printed matches support but do not prove.

So the discrepancy most likely reflects how the averages were taken. It probably does not reflect a difference between the paper's model and its figure, or between our run and the paper's run. This is the likely explanation for the open issue recorded in benchmark-004, protocol 006 and protocol 007. The frozen protocols are not edited, and the 24 Figure 4 means remain the declared comparison target.

One consequence for later biological comparisons: the paper compares 3.56 nS with the 4.9 nS it cites from cat recordings (reference 44: Monier, Fournier and Frégnac 2008). The size-weighted model average is 4.09 nS. Which model average, if either, suits that comparison depends on which neurons and layers the cat recordings sampled. We have not checked that source.

## Limits

- Inference only, as above.
- One run. Our means match the printed Figure 4 values at their precision. The three-digit match assumes our run is close to the paper's run, which the 24 printed matches support but do not prove.
- The pooled and unweighted rules differ by at most 0.0012 nS, so this check cannot tell them apart.
- The sampling standard errors are rough. They treat the recorded neurons as random samples and ignore differences between network realizations and simulator versions.

## Files

- `average_conductances.py` → `conductance-averages.json`
- `test_average_conductances.py`: a rerun reproduces the packaged output byte for byte; the result is as stated; a modified input is rejected before any output is written.
- `draw_averages.py` → `figures/conductance-averages.png` / `.svg` (computed values only)
- `sources/antolik-2024.xml`: the article text used (CC BY 4.0)
- `manifest.json` (SHA-256 and size of every file except itself and the zip) and `reference-evidence.zip`

To rerun, place this folder next to `run-011`, `benchmark-004` and `protocol-007`, as on the website under `/research/`. Then run `python3 average_conductances.py --research .. --output .` and `python3 -m unittest -v test_average_conductances`.

## Attribution

Text and Figure 4 values: 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), CC BY 4.0. The XML copy is unchanged. Code: LSV1M v1.0, MIT license. Cat conductance measurements cited by the paper: Monier C, Fournier J, Frégnac Y (2008), *Journal of Neuroscience Methods* 169(2):323–365, [doi:10.1016/j.jneumeth.2007.11.008](https://doi.org/10.1016/j.jneumeth.2007.11.008). Population means: our run 011.
