# conductance-022: protocol 022's declared conductance follow-up

13 November 2026. Journal note 26, "What the suppressed cells received".

Protocol 022 (the plaids pilot, journal note 22) declared this follow-up on 26 October 2026, before its pilot ran
(`protocol-022-public/frozen/spec.md`, "Follow-up, declared now and run after the pilot"). It rebuilds offline, from
recorded spikes and exported connection lists, the synaptic conductances each layer 4 excitatory cell received, and
makes three readings. **It is exploratory by declaration, uses one network instance (the pilot's instance A), and
changes no outcome:** protocol 022's GO and protocol 023 revision 2's CONFIRMED stand whatever is found here.
Simulation only; no claim about cat visual cortex.

## Result

| Step | Outcome |
|---|---|
| Ground truth of the replay method (toy network) | 60 of 60 trains exact; two planted errors fail |
| V1, replay of 017's feedforward-only run | PASSED at recorded attempt 1 |
| V2, replay of 017's intact run, all five lists | PASSED, same attempt, tool unchanged |
| V3, replay of the pilot's lesion run | PASSED, same attempt, tool unchanged |
| Readings | made twice, byte-identical (`results/readings-a.json`, `readings-b.json`) |

Validation numbers (criteria: counts identical in at least 98 % of cell-and-step pairs; worst group-mean deviation at
most 1 %; at least 90 % of recorded spikes with a replayed spike within ±0.2 ms; spike ratio 0.99 to 1.01):

| | recorded spikes | pairs identical | worst group deviation | near share | spike ratio |
|---|---|---|---|---|---|
| V1 | 12,260,678 | 1,427,952 of 1,428,000 (0.99997) | 0.00017 | 0.99972 | 1.00002 |
| V2 | 1,695,598 | 1,427,992 of 1,428,000 (0.999994) | 0.0022 | 0.99994 | 1.00002 |
| V3 | 7,815,513 | 1,163,939 of 1,164,000 (0.99995) | 0.00024 | 0.99964 | 1.00003 |

Readings (intact run; 95 % intervals from 10,000 whole-step resamplings, seed 2202; all 8 of 8 combinations counted):

| Reading | Value | 95 % interval | Remark (verdicts as the code wrote them) |
|---|---|---|---|
| A_aff, afferent additivity | 0.622 | 0.603 to 0.642 | identical in the intact and lesion runs (largest difference 0.0 nS) |
| MM1, m_I, change of inhibition | −0.145 | −0.186 to −0.102 | the opposite direction, beyond the margin (Holm at 0.05) |
| MM2, m_E, change of intracortical excitation | −0.389 | −0.430 to −0.345 | supported (Holm at 0.05, beyond the margin of 0.10) |
| Afferent XOS | 0.860 | 0.810 to 0.909 | **with the declared caution: not a suppression index** (mean cross drive 1.010) |
| Spike XOS, recomputed | 0.584392 | 0.552 to 0.614 | equals the pilot's published value |
| Spike XOS minus afferent XOS | −0.275 | −0.329 to −0.221 | carries the same caution |

## What must be read beside the numbers

- **The amendment.** The declared text counts "12 group-and-plaid combinations" and needs "at least 9 of its 12". The
  declared stimulus set holds 8 (two cell groups × four block A plaids); twelve is a leftover from an earlier draft
  with six plaids (the design draft is not published). Amended before any of the pilot's recorded runs was opened:
  a reading needs 6 of its 8. A literal alternative, three declared groups × four plaids = 12, was not taken: the 90°
  group has no test grating in a 45°/135° plaid, so MM1 and MM2 are undefined for it and could never reach 9. A_aff
  could have been computed for the 90° group too; it was kept to the same two groups, which is a choice. Under the
  literal count the readings would not be made. The full text is in `frozen/WORKLOG.md` and in the `note` of A_aff,
  MM1 and MM2. In the pooled readings 3 of 8 single-combination intervals include zero for m_I and 1 of 8 for m_E.
- **No cause.** The conductances are rebuilt from recorded spikes. Less intracortical excitation is also a consequence
  of similarly tuned cells firing less; the readings describe the suppressed state and cannot separate cause from
  consequence.
- **The afferent mean is not orientation-tuned.** For the 45° group the mean afferent conductance is 1.052 nS for the
  blank, 1.244 for its own grating at 30 %, 1.258 for the orthogonal grating and 1.267 for the plaid. Protocol 022's
  validity rule for a suppression index fails, so the afferent XOS does not say how much of the suppression the input
  stage carries (mean cross drive over the four own gratings 1.010; the rule allows 0.10). A window mean cannot see
  the input's timing within a step, which presumably carries orientation; that was not measured. The declared text
  calls A_aff "the suppression in the drive layer 4 receives"; with an untuned mean we read it only as the mean input
  growing less than additively, a reading made after seeing the numbers. That question stays open.
- **One instance,** exploratory, on a laptop that can make rare one-off execution faults (journal notes 24 and 25):
  every replay and the readings were computed twice and compared.
- **What the validations do not show.** They show that replayed cells fire as the recorded cells did. No conductance
  was recorded in the model runs; the offline conductance code was checked against the simulator's own recorded
  conductances only on a test network (`frozen/test_conductance_nest.py`).
- **Pilot data seen before the freeze.** The connection lists were exported by the pilot's runs (the same network
  instance as 017's runs) and the replay used them from V1 on. And one look at a recorded run: the first 3,000
  characters of the pilot's intact export (its header: run time, command line, resolved seeds and weights) were
  printed to learn the file layout. No spike time, count or measure was read.
- **Spike counts.** The validations count spikes after the excluded first 300 ms (V3: 7,815,513). `frozen/WORKLOG.md`
  quotes 7,901,609 for V3, a count over the whole run.
- **The independent review** (one fresh reader, 12 findings, none high, all applied) is recorded in the working record
  only; the review itself is not published.

## Order of events (UTC, 13 November 2026)

1. Criteria code and tests; ground truth of the replay method. Two earlier tries are kept: the method was already
   exact, but the toy network fired under the 500-spike minimum fixed beforehand, so the run was lengthened.
2. 01:50 gate of V1 written; V1 passed at 02:53.
3. 02:53 gate of V2 written; V2 passed at 04:16.
4. While V1 and V2 ran, and until 04:29: conductance and readings code written and tested without the pilot's
   recorded runs; one independent review (12 findings, none high, all applied); the amendment recorded.
5. 04:30 manifest frozen (`frozen/MANIFEST-frozen.txt`, original SHA-256 `d5b1fb35…708c96`), then the gate of V3.
6. V3 passed (about 04:56). Readings computed twice from the frozen code.

## Contents

- `frozen/`: the code as frozen (`replay_l4.py`, `criteria.py`, `pool.py`, `conductance.py`, `readings.py`,
  `groundtruth.py`, three test files), `MANIFEST-frozen.txt`, the plan (`plan.md`) and the working record
  (`WORKLOG.md`; it gained its last section after the freeze, so its hash differs from the manifest's entry).
- `groundtruth/`: the ground-truth result and the two earlier tries.
- `attempts/a1`, `a1-v2`, `a1-v3`: for each validation the gate written before the run, the run script and its
  output, the criteria per presentation, memory records, the pooled result. `attempts/replayed-sha256.txt` lists the
  ten replayed spike files (about 128 MB, kept with the project, not published); the a and b files of each
  presentation have equal uncompressed content (the `.gz` files differ only in the gzip header's time stamp).
- `results/`: the readings, twice, with logs and memory records.
- `figures/`: the two figures of note 26 with data, scripts and `captions.md`.
- `manifest.json`: SHA-256 of every file here. Twenty-eight files are copies with an absolute folder path replaced
  by `<workspace>` or `<home>` (paths written as `~/…` are left as they are); each is listed under `redactions` with its original's SHA-256, so the frozen manifest's
  entries can be checked against the originals' hashes recorded there.

Inputs not republished here: the pilot's exports and connection lists (`plaids-022`), protocol 017's exports
(`protocol-017`), and protocol 022's frozen `plaid_stats.py` (`protocol-022-public`), whose SHA-256 the readings
record.

## Attribution

Model: Antolík et al. (2024), https://doi.org/10.1371/journal.pcbi.1012342, CC BY 4.0; LSV1M v1.0 (MIT); Mozaik
v0.4.0 (CeCILL); NEST 3.4 (GPL); PyNN 0.11.0 (CeCILL). The follow-up's code, replays, readings and figures are ours.
