# Protocol 022's declared follow-up: the conductance rebuild (work folder)

Plan: `docs/specs/2026-11-13-conductance-rebuild-022.md`. Declared text: `docs/specs/2026-10-26-protocol-022-plaids-r2.md`, lines 235 to 250. Exploratory by declaration; it changes no outcome.

## State, 13 November 2026
- `criteria.py`, `test_criteria.py`: the four declared validation criteria; 8 synthetic tests pass.
- `groundtruth.py`: the replay method on a closed-loop toy network of the model's cell and synapse models (60 excitatory, 20 inhibitory aeif_cond_exp cells, 40 inputs, tsodyks_synapse, 6 s). Open-loop replay reproduced 60 of 60 excitatory trains exactly (647 spikes); a delay wrong by 0.1 ms left 1 of 60 identical and a wrong tau_rec 12 of 60, and both fail the criteria. Result: `groundtruth-result.json`. Two earlier tries are kept (`*.try2-465-spikes.json`, `*.try3-472-spikes.json`): the method was already exact, but the toy fired under the 500-spike minimum fixed beforehand, so the run was lengthened, not the minimum lowered.
- Finding: NEST 3.4 refuses `spike_generator -> tsodyks_synapse`. Each recorded cell is replayed by a generator that fires 0.1 ms early into a `parrot_neuron` over a static 0.1 ms connection; the parrot is the tsodyks source.
- No 017 run and no pilot export has been read. No declared attempt has been spent (0 of 3).

## Next
The real replay tool on 017's feedforward-only run (development on 017's runs is what the declared blinded debugging allows), then recorded attempt 1 of V1.

## 13 November, 02:53 UTC: V1 PASSED at recorded attempt 1
attempts/a1/v1-result.json: pooled pair share 0.99997 (needs 0.98), worst group deviation 0.00017 (needs at most 0.01), near share 0.99972 (needs 0.90), spike ratio 1.00002; 12,260,678 recorded spikes over both presentations; each presentation's two replays equal. V2 started with the tool unchanged (attempts/a1-v2/GATE.md).

## 13 November, 04:16 UTC: V2 PASSED (attempt 1, tool unchanged)
attempts/a1-v2/v2-result.json: pooled pair share 0.999994, worst group deviation 0.0022, near share 0.99994, spike ratio 1.00002; 1,695,598 recorded spikes over both presentations of 017's intact-A; replays equal. V2 licenses MM1 and MM2.
No pilot export has been opened. Next, in this order: (1) the conductance code (offline, per source class, NEST's tsodyks update written out) and the readings code, tested on synthetic data and cross-checked against NEST's own recorded conductances, all without pilot data; (2) one independent review; (3) the manifest frozen and its hash recorded in HANDOFF; (4) V3 on the pilot's lesion run; (5) the readings each validation licenses, computed twice.

## Amendment of 13 November 2026, fixed before any pilot export was opened
The declared follow-up (protocol 022 revision 2, lines 248 to 250) speaks of "the 12 group-and-plaid combinations" and "at least 9 of its 12". The declared stimulus set does not hold twelve.
- Origin: the design panel's draft (`../plaids/design-panel-r2.json`: "g in {G45, G135} and the 6 orthogonal plaids ... the 12 combinations ... at least 9") had six orthogonal plaids. Revision 2 cut the set to four block A plaids and kept the old counts in the follow-up's text.
- What exists: {45, 135} x four block A plaids = 8 combinations.
- The reading taken: a reading needs at least 6 of its 8 combinations, the same three quarters.
- A literal alternative, not taken: the three +-15 deg groups x four plaids = 12. The 90 deg group has no test grating in a 45/135 plaid, so MM1 and MM2 are undefined for it; they would reach 8 of 12 at most and could never be made. The draft that wrote "12" named only G45 and G135.
- Consequence, stated beside every reading in the result: under the literal count MM1 and MM2 would not be made. These readings are exploratory by declaration and change no outcome.
Other choices fixed at the same time, from the independent review of 13 November (one fresh reader; no high finding; 12 findings, all applied):
- Holm stays a family of two even if one of MM1 and MM2 is not made (its p counts as 1).
- The afferent XOS is made only if each of its four own-single denominators is at least 3 standard errors; its mean cross drive is reported, with protocol 022's caution where it lies outside +-0.10.
- Readings are licensed from the validations' result files, not by hand; V3's replay must be of the lesion export read.
- "In both networks": the two runs share every LGN spike and the lists, so the afferent table should be one; the result records whether it is.
- The lesion run gets the afferent readings only. The spike XOS and its difference from the afferent XOS are made for the intact run, as declared.
- V3's criteria count all 97 steps, the warm-up included: a superset of the analysed steps.
- One earlier look at pilot data, before this fence existed: on 13 November the first 3,000 characters of the pilot's intact export were printed to learn the export's layout (its header: run time, command line, resolved seeds and weights). No spike time, count or measure was read.

## 13 November, 04:56 UTC: V3 PASSED; the readings were made (twice, byte-identical)
- V3 (attempts/a1-v3/v3-result.json): pair share 0.99995, worst group deviation 0.00024, near share 0.99964, spike ratio 1.00003 over 7,901,609 recorded spikes of the pilot's lesion run; replays equal; tool unchanged. One recorded attempt covered V1, V2 and V3; two of the three declared attempts were never used.
- Frozen files checked against MANIFEST-frozen.txt (d5b1fb35...) before the readings: all match.
- results/readings-a.json and readings-b.json: SHA-256 6bf7e5e0c773816db7c5a722feb1888312f31f03aa9f363d9425988712389b4e, both.
- Integrity: the spike XOS recomputed here is 0.584392, the pilot's published value, with its interval.
- A_aff = 0.622 (95 % 0.603 to 0.642), 8 of 8 combinations: the mean afferent conductance to a plaid is 62 % of the sum of the two singles' evoked conductances. The afferent table is identical in the intact and lesion runs (largest difference 0.0 nS).
- Afferent XOS = 0.860 (0.810 to 0.909), WITH protocol 022's caution: the mean cross drive is 1.010 (0.994 to 1.027), so the mean afferent conductance in window G is not orientation-tuned (45 deg group: blank 1.052 nS, own grating at 30 % 1.244, orthogonal grating 1.258, plaid 1.267), and this value is not a suppression index. Spike XOS minus it: -0.275 (-0.329 to -0.221), carrying the same caution.
- MM1 m_I = -0.145 (-0.186 to -0.102), 8 of 8, p 0.0002: the opposite of the asked direction, beyond the margin. The second grating does not raise the inhibition onto cells tuned to the first; it lowers it (45 deg group: blank 3.28 nS, test 5.73, plaid 5.13).
- MM2 m_E = -0.389 (-0.430 to -0.345), 8 of 8, p 0.0002: supported. The second grating lowers the intracortical excitation onto cells tuned to the first by 39 % of what the test grating evoked (45 deg group: blank 0.28 nS, test 0.97, plaid 0.63).
- Limits to state beside these: exploratory by declaration; one network instance (A); conductances are rebuilt from the recorded spikes, so a lower recurrent excitation is also a consequence of fewer spikes in similarly tuned cells, and these readings do not show cause; window means cannot see the temporal modulation that carries orientation in the afferent drive; the "9 of 12" amendment above; simulation only.
