# Felisyn plaids 022: the model circuit did not combine two gratings like a relay (pilot, outcome GO)

Results of protocol 022 revision 2, the plaids pilot, dated 28 October 2026. A plaid is two drifting gratings shown on top of each other. The pilot asked whether the cat circuit answers a plaid with the sum of its answers to each grating alone, as a relay would, or with something else.

- **Declared** at 2026-10-26T23:15:38Z, before any plaid run. Spec SHA-256 `8f8878fc60e5378224def8866068d3e3a1ca03ca4a48b69fa69446f5bee662d3`.
- **Frozen** at 2026-10-27T18:39:13Z by a manifest with SHA-256 `a1e2e2d92a7bfe4855da5b2f03e97674bfaa8ef1a0849c8d60faabc116c8b3d0`. The protocol package `protocol-022` holds the frozen files and that manifest. Its public copy, `protocol-022-public`, has three lines with private strings redacted.
- **Run** on the research server from 27 October 18:39:39 to 28 October 00:40:21 UTC.
- **Analysed once** on the frozen contract, on 28 October from 04:26:23 to 04:26:39 UTC.
- **Outcome: GO**, by the declared rule.

Simulation only. This is one network instance of a 2 mm reduction of a published model of cat primary visual cortex, not a biological measurement and not biological validation. No token, wallet, exchange, key or real order was involved.

## The question

In the spec's words: "In one instance of the 2 mm reduction of LSV1M (`pynn_seed` 5), at protocol 017's timing, does the intact circuit combine two superimposed drifting gratings differently from a relay? Here a relay means the sum of the circuit's own above-blank responses to each grating alone."

Three measures on the intact network answer it. Each is exactly 0 for a relay as defined here, so the network's own tuning and overall gain cancel out of it. Anything non-linear that acts when both gratings are present, such as saturation in the input stage or in the cortex, can move it away from 0. The spec says each is 0 "whatever the network's tuning, gain or contrast saturation". That holds only for a relay as defined: the design review had flagged the sentence before the declaration, and it was not corrected.

- **XOS, cross-orientation suppression.** Take layer 4 cells that prefer one grating. Show their preferred grating, then a grating at right angles to it, then both together. XOS is how far the response to both falls below the sum of the two single responses, in units of the response to the preferred grating. It is valid only where the orthogonal grating alone barely moves these cells from the blank level.
- **Dc, dominance.** In a plaid of unequal contrasts, do the cells that prefer the stronger grating keep a larger share of their single-grating response than the cells that prefer the weaker one? A gain shared by both groups cancels out, including uniform suppression.
- **ds, readout shift.** The controller reads one direction from layer 4 with protocol 017's frozen readout. For a plaid of two gratings 45 degrees apart, does that readout point further toward the stronger grating than the sum of its readouts of the two gratings alone? It is in the controller's signal units, where 45 degrees is 1 and the decision threshold is 0.25.

Why the controller project cares: a controller that receives two signals at once needs to know whether the circuit passes their sum on, or lets one suppress or outweigh the other.

## Design

The full design is the frozen package `protocol-022`, with manifest SHA-256 `a1e2e2d92a7bfe4855da5b2f03e97674bfaa8ef1a0849c8d60faabc116c8b3d0`. It is published as [protocol-022-public](https://www.felisyn.com/research/protocol-022-public/README.md), with the [spec as declared](https://www.felisyn.com/research/protocol-022-public/frozen/spec.md). It holds the spec (`spec.md`), the contract (`contract.json`, SHA-256 `128b5bba…`), the sequence (`sequence-r2.json`, SHA-256 `41ecee87…`, seed 231), the analysis code, the server scripts and their tests. The public copy keeps these in its folder `frozen/`, with three lines that hold private strings redacted, as its README lists, and every other file as the frozen manifest pins it. The launcher refused to start the queue unless all 34 server files matched their frozen hashes: 15 against this manifest, and the 19 model files against protocol 017's frozen list.

- **Network and conditions.** Network instance A of the 2 mm reduction, in two runs on the identical sequence and identical input spikes:
  - **intact**, the published connectivity;
  - **plain lesion**, all twelve intracortical projections at zero weight, the input unchanged.
- **Stimulus.** Gratings of 0.8 cycles per degree drifting at 2 Hz. Each step is 210 ms of grating or plaid, then 147 ms of mean-luminance blank, after a 504 ms lead-in. One presentation of 97 steps: a warm-up and 96 analysed steps, 35.1 s of model time.
- **The 15 analysed conditions.**
  - Block A, gratings at right angles (45 and 135 degrees): the blank (8 repeats); each grating alone at 15 and 30 % contrast (8 repeats) and at 60 % (4 repeats); four plaids (8 repeats each) at 15 + 15, 30 + 30, 30 + 15 and 15 + 30 %.
  - Block B, gratings 45 degrees apart (90 and 135 degrees): the 90 degree grating alone at 15 and 30 % (4 repeats each), and two plaids (4 repeats each), 135 at 30 % with 90 at 15 %, and 135 at 15 % with 90 at 30 %.
- **Cells and windows.**
  - Groups: layer 4 excitatory cells whose preferred orientation lies within 15 degrees of 45, 90 or 135 degrees (915, 721, 1,071 cells).
  - Group spike counts are taken 20 to 230 ms after each onset (window G).
  - The readout uses 40 to 300 ms (window D), with protocol 017's frozen readout weights.
- **Intervals.** A paired bootstrap over whole steps, never over cells: 10,000 replicates, seed 2202, the same draws for both runs. The 95 % intervals are for reporting and the connection readings; the 90 % intervals for the relay-like test. Holm's method over the three primaries at family-wise 0.05.
- **The outcome rule,** checked in this order:
  - NOT INTERPRETABLE;
  - GO, if at least one primary passes Holm and its estimate is at least 0.10 from 0;
  - RELAY-LIKE, if all three 90 % intervals lie inside ±0.10;
  - INCONCLUSIVE otherwise.
  - Each outcome has a fixed text, written before the runs.
- **Calibration before the runs.** The spec reports synthetic pilots built from protocol 017's noise:
  - false GO under a relay at most 3.55 %;
  - coverage of the primaries' 95 % intervals 0.934 to 0.966.
- **Details the spec leaves to the code** are listed in `design/analysis-choices-r2.md` (SHA-256 `dc1c14290849d2d8…`). That file's SHA-256 was recorded in the project handoff at the declaration. It is not one of the frozen manifest's files.

## The runs

The queue started at 18:39:39 UTC on 27 October. Its guard found no other project service busy, and it ran three jobs one after another on the research server, each at its first attempt.

| Run | What it ran | Start to end (UTC) | Duration | Exit | Memory peak (unit) | Largest resident size | CPU steal |
| --- | --- | --- | --- | --- | --- | --- | --- |
| `m22p-intact-a1` | intact network, the declared sequence | 27 Oct 18:39:39 to 21:33:45 | 2 h 54 min | 0 | 5.26 GiB | 5.27 GiB | 3.6 % |
| `m22p-lesion-a1` | plain lesion, the same sequence and input | 27 Oct 21:33:45 to 28 Oct 00:24:00 | 2 h 50 min | 0 | 5.65 GiB | 5.62 GiB | 1.9 % |
| `m22p-connections-a1` | builds the intact network, writes the connection lists onto layer 4, stops before any stimulus | 28 Oct 00:24:00 to 00:40:21 | 16 min | 0 | 4.29 GiB | 4.26 GiB | 1.6 % |

- **Memory.** The intact and lesion runs share one service unit, whose memory cap was 6.35 GiB. The lesion's peak is that unit's peak, and it rose above the intact run's, so it is the lesion's own. No run reached the cap: memory.events shows 0 max, 0 oom and 0 oom_kill events in every run.
- **CPU steal.** It is measured over each run from the host's CPU counters at start and end. The shared host lost 1.6 to 3.6 % of its CPU time to steal.
- **Server files.** All 34 server files had the same SHA-256 at the start and the end of every run.
  - 15 of them are the files of the frozen manifest.
  - The other 19 are the model files pinned through protocol 017's frozen server list.
- **The exports** were fetched and checked against the server's own SHA-256 at 04:24 UTC on 28 October.
  - Intact: 25,552,690 bytes, SHA-256 `952b469b9a8e74a4…`.
  - Lesion: 46,569,465 bytes, SHA-256 `42484edd1e15aec6…`.
  - They are too large for this site and stay private, listed in `data-sha256.txt`.
- **The analysis.** `analyze_plaids.py` ran once, unchanged from the frozen package, from 04:26:23 to 04:26:39 UTC on 28 October, with exit 0 (`results/analysis-run-2026-10-28.log`).
  - It used Python 3.14.3, numpy 2.4.2 and scipy 1.17.1 on macOS 26.2 (arm64).
  - Its maximum resident set size was 1.26 GB.

## The result

### Declared outcome: GO

The fixed GO text, as the analysis wrote it (`results/results.json`, field `text`):

> In this one network instance the intact circuit combined two superimposed gratings differently from a relay of its single-grating responses, by at least the declared margin: an orthogonal grating lowered layer 4 cells' response to their preferred grating (cross-orientation suppression), XOS = 0.584 (95 % interval 0.552 to 0.614); at unequal contrasts the cells preferring the stronger grating kept more of their single-grating response than those preferring the weaker one (winner-take-more, as normalisation predicts), Dc = 0.713 (95 % interval 0.651 to 0.782); the controller's frozen readout of a plaid of gratings 45 degrees apart pointed 0.342 signal units further toward the stronger grating than the sum of its readouts of the two gratings (ds = 0.342, 95 % interval 0.295 to 0.391). This is a property of the whole model circuit (input stage and cortex) at these contrasts and this timing, and the input stage alone was expected to produce part of it. It does not show that the intracortical connections compute it: the shared input stage is non-linear, and suppression or dominance of the afferent drive shows in spikes only in tuned cells, which need the connections. A confirmatory protocol on fresh network instances follows, declared before its runs.

### The three primary measures (intact network)

| Measure | Estimate | 95 % interval | 90 % interval | p | Holm threshold | Passes Holm | At least 0.10 from 0 |
| --- | --- | --- | --- | --- | --- | --- | --- |
| XOS, cross-orientation suppression | **+0.584** | +0.552 to +0.614 | +0.557 to +0.609 | 0.0002 | 0.0167 | yes | yes |
| Dc, dominance at unequal contrasts | **+0.713** | +0.651 to +0.782 | +0.661 to +0.770 | 0.0002 | 0.0250 | yes | yes |
| ds, readout shift toward the stronger grating | **+0.342** | +0.295 to +0.391 | +0.303 to +0.383 | 0.0002 | 0.0500 | yes | yes |

- **p values.** Each p is the smallest that 10,000 bootstrap replicates can give, 2 / 10,001: no replicate reached 0.
- **Holm and the margin.** All three pass Holm, and all three estimates are at least 0.10 from 0. So each of the three triggers GO.
- **The relay-like test** is false: no 90 % interval lies inside ±0.10.

### What the numbers say, in plain words

- **Cross-orientation suppression (XOS).** Take the 45 degree cells at 30 % contrast. Their preferred grating alone raised them 2.85 spikes per cell above the blank, and the orthogonal grating alone 0.05. A relay would answer the plaid of both with 2.90. The plaid gave 1.04, a drop of 1.86, which is 0.65 of the preferred response. Over the four groups and contrasts XOS ran from 0.50 to 0.65, with a mean of 0.584.
- **Dominance (Dc).** In the plaid of 45 degrees at 30 % with 135 degrees at 15 %, the cells preferring the stronger grating kept 0.76 of the sum of their single-grating responses, and the cells preferring the weaker one 0.10. In the mirror plaid it was 0.85 against 0.17. The cells of the stronger grating kept most of their response, and the cells of the weaker one lost most of theirs.
- **Readout shift (ds).** Alone, the 135 degree grating at 30 % reads as 1.07 signal units and the 90 degree grating at 15 % as 0.01. Their two readouts add up to a direction of 0.76. The plaid itself read 1.02, 0.258 further toward the stronger grating. In the mirror plaid, with 90 degrees stronger, the sum points to 0.39 and the plaid read −0.03, close to the 90 degree grating alone (−0.04). For scale, the controller's decision threshold is 0.25 signal units.
- **Decisions (secondary S5).** Under protocol 017's rule, 7 of the 40 plaid steps were decided on a different side (buy, sell or hold) from the one the linear prediction gives. All 4 steps of the plaid 135 degrees at 15 % with 90 degrees at 30 % read as hold, where the linear prediction gives buy.

### Components of the primaries (secondary S1, both runs)

| Component (S1) | Intact: estimate (95 % interval) | Plain lesion: estimate (95 % interval) |
| --- | --- | --- |
| XOS, 45° cells at 15 % | +0.501 (+0.416 to +0.576) | +0.628 (+0.561 to +0.694) |
| XOS, 135° cells at 15 % | +0.573 (+0.508 to +0.635) | +0.633 (+0.566 to +0.699) |
| XOS, 45° cells at 30 % | +0.653 (+0.600 to +0.702) | +0.840 (+0.802 to +0.877) |
| XOS, 135° cells at 30 % | +0.610 (+0.565 to +0.653) | +0.816 (+0.780 to +0.853) |
| XOS, classic form 1 − D(plaid) / D(preferred), mean | +0.575 (+0.545 to +0.602) | +0.104 (+0.074 to +0.133) |
| Cross drive, D(orthogonal) / D(preferred), mean of four | +0.010 (−0.014 to +0.034) | +0.625 (+0.608 to +0.642) |
| XOS, 45° cells minus 135° cells at 15 % | −0.072 (−0.176 to +0.028) | −0.005 (−0.056 to +0.045) |
| XOS, 45° cells minus 135° cells at 30 % | +0.043 (−0.026 to +0.111) | +0.023 (−0.011 to +0.057) |
| Dc, 45° at 30 % + 135° at 15 % | +0.761 (+0.666 to +0.874) | +0.130 (+0.117 to +0.143) |
| A, cells preferring the stronger grating | +0.756 (+0.671 to +0.848) | +0.720 (+0.692 to +0.750) |
| A, cells preferring the weaker grating | +0.103 (+0.052 to +0.148) | +0.555 (+0.529 to +0.582) |
| Dc, 45° at 15 % + 135° at 30 % | +0.665 (+0.591 to +0.741) | +0.121 (+0.109 to +0.133) |
| A, cells preferring the stronger grating | +0.846 (+0.788 to +0.912) | +0.735 (+0.706 to +0.766) |
| A, cells preferring the weaker grating | +0.170 (+0.127 to +0.215) | +0.577 (+0.552 to +0.603) |
| Dc, first plaid minus second | +0.096 (−0.025 to +0.227) | +0.009 (−0.008 to +0.026) |
| ds, 135° at 30 % + 90° at 15 % | +0.258 (+0.195 to +0.323) | +0.433 (+0.369 to +0.496) |
| ds, 135° at 15 % + 90° at 30 % | +0.427 (+0.347 to +0.509) | +0.447 (+0.384 to +0.517) |
| ds, first plaid minus second | −0.169 (−0.278 to −0.060) | −0.013 (−0.103 to +0.072) |

- **XOS in the lesion is not a suppression index.** The lesion's cross drive fails the declared condition (within ±0.10), so its XOS values above are reported only, as declared (S4).
- **A is the additivity of a plaid:** the group's plaid response over the sum of its two single-grating responses (S2). 1 is a relay.
- **Order differences** are the first listed plaid minus the second.

### Other declared secondary measures (no decision role)

| Measure | Intact | Plain lesion |
| --- | --- | --- |
| S3 parallel control, 45° cells, 15 + 15 % plaid against 30 % single: S_par | +0.683 (+0.636 to +0.726); informative | +0.431 (+0.400 to +0.463); informative |
| S3 parallel control, 135° cells, 15 + 15 % plaid against 30 % single: S_par | +0.715 (+0.679 to +0.751); informative | +0.496 (+0.465 to +0.527); informative |
| S3 parallel control, 45° cells, 30 + 30 % plaid against 60 % single: S_par | +0.675 (+0.621 to +0.723); saturated, uninformative | +0.323 (+0.294 to +0.351); informative |
| S3 parallel control, 135° cells, 30 + 30 % plaid against 60 % single: S_par | +0.667 (+0.631 to +0.702); informative | +0.366 (+0.340 to +0.392); informative |
| S4 lesion XOS |  | +0.729 (+0.693 to +0.764); "not a suppression index here" (cross drive +0.625) |
| S4 lesion XOS, classic form |  | +0.104 (+0.074 to +0.133) |
| S4 lesion Dc |  | +0.125 (+0.116 to +0.134) |
| S4 lesion ds |  | +0.440 (+0.393 to +0.487) |
| S4 dds = ds intact − ds lesion (paired) | −0.097 (−0.161 to −0.034) |  |
| S5 share of the 40 plaid steps whose decided side (buy, sell or hold, 017's rule) differs from the linear prediction's | 0.175 (0.100 to 0.175) | 0.700 (0.625 to 0.800) |
| S6 LGN X_ON spike additivity (identical input in both runs) | 45° at 15 % + 135° at 15 %: 0.715; 45° at 30 % + 135° at 30 %: 0.602; 45° at 30 % + 135° at 15 %: 0.731; 45° at 15 % + 135° at 30 %: 0.747 | same input |
| S6 LGN X_OFF spike additivity (identical input in both runs) | 45° at 15 % + 135° at 15 %: 0.666; 45° at 30 % + 135° at 30 %: 0.574; 45° at 30 % + 135° at 15 %: 0.717; 45° at 15 % + 135° at 30 %: 0.708 | same input |
| S6 linear-rectifier reference | 45° at 15 % + 135° at 15 %: 0.637; 45° at 30 % + 135° at 30 %: 0.637; 45° at 30 % + 135° at 15 %: 0.709; 45° at 15 % + 135° at 30 %: 0.709 |  |
| S7 layer 4 inhibitory groups, XOS | +0.566 (+0.542 to +0.589) (cross drive +0.060) | +0.689 (+0.659 to +0.717) (cross drive +0.543; not estimable) |
| S8 contrast at half response (c50), 45° cells, % (descriptive fit) | 14.0 (12.7 to 20.6) | 22.2 (19.2 to 28.8) |
| S8 contrast at half response (c50), 135° cells, % (descriptive fit) | 14.7 (13.0 to 21.6) | 21.9 (19.3 to 26.9) |
| S9 layer 4 excitatory rate in the blank steps, Hz | 2.08 | 21.32 |
| S9 layer 4 excitatory rate in the lead-in, Hz | 2.20 | 23.44 |
| S11 groups in D, readout in G: XOS / Dc / ds | +0.559 (+0.529 to +0.586) / +0.644 (+0.595 to +0.695) / +0.373 (+0.317 to +0.430) | +0.792 (+0.771 to +0.813); not estimable / +0.094 (+0.088 to +0.100) / +0.494 (+0.440 to +0.550) |
| S11 onset [20, 70): XOS / Dc / ds | +0.538 (+0.472 to +0.595) / +0.627 (+0.515 to +0.748) / +0.275 (+0.156 to +0.396); not estimable | +0.801 (+0.718 to +0.877); not estimable / +0.044 (+0.029 to +0.058) / +0.272 (−0.396 to +1.506); not estimable |
| S11 sustained [70, 230): XOS / Dc / ds | +0.597 (+0.563 to +0.628) / +0.743 (+0.669 to +0.826) / +0.405 (+0.337 to +0.474) | +0.670 (+0.615 to +0.720); not estimable / +0.239 (+0.216 to +0.264) / +0.380 (+0.334 to +0.426) |
| S11 offset [230, 300): XOS / Dc / ds | +0.494 (+0.426 to +0.553) / +0.459 (+0.374 to +0.547) / +0.275 (+0.181 to +0.369); not estimable | +0.824 (+0.788 to +0.859); not estimable / +0.035 (+0.025 to +0.045) / +0.507 (+0.350 to +0.679); not estimable |
| S12 refit with previous summed contrast (multiplicative, per 100 %, centred): XOS / Dc / ds | +0.584 / +0.712 / +0.344 | +0.729 (not estimable) / +0.125 / +0.440 |
| S12 refit with linear time trend (per run, centred): XOS / Dc / ds | +0.585 / +0.712 / +0.342 | +0.728 (not estimable) / +0.125 / +0.439 |

- **S3** compares a plaid of two gratings at c % with a single grating at the same total contrast, 2c %: S_par = 1 − D(plaid) / D(single at 2c). It is "saturated, uninformative" where the single grating's response hardly rises from c to 2c, as declared.
- **S11** recomputes the primaries with the two windows swapped (group counts in 40 to 300 ms, the readout in 20 to 230 ms), and in the onset (20 to 70 ms), sustained (70 to 230 ms) and offset (230 to 300 ms) windows. "Not estimable" marks a value that fails its declared validity condition there.
- **S12** refits P1 to P3 with the previous step's summed contrast as a covariate, or with a linear time trend. The intact estimates change by at most 0.002.
- **Where the rest is.** S2 (the additivity of every plaid and group) and S10 (the phase geometry of each plaid) are in `results/results.json`, with all intervals, as are the per-condition values behind S5, S8 and S9.
- **S6** is labelled, as declared, "upstream of thalamocortical depression; not the drive layer 4 receives". Its reference is a linear rectifier with no spontaneous firing.

### The connection readings (declared texts)

The connection readings were made, because the outcome is GO and none of the conditions that would remove them holds:
- the lesion run completed and passed its 30 % check;
- no refusal fired on the lesion export;
- the LGN spike trains and the neuron tables are identical between the runs.

These readings help choose the confirmatory design. They never change the outcome.

**At the readout: SAME DIRECTION WITHOUT THE CONNECTIONS.**
The gate needs an intact effect first: ds_int's 95 % interval must exclude 0 and |ds_int| must be at least 0.10. ds_int = +0.342 (+0.295 to +0.391), so the gate passed. Without the connections, ds_les = +0.440 (+0.393 to +0.487). The paired difference dds = ds_int − ds_les = −0.097 (−0.161 to −0.034). The analysis wrote the text below (`results.json`, field `connection_readings`).

> In this one network instance, without the intracortical connections (at a 7 to 8 times higher layer 4 rate and without its tuning), the readout departed from a relay in the same direction, ds_les = 0.440. The lesion's own saturation can produce this, so it does not show that the intact readout's departure comes from the feedforward path.

**The group measures (fixed text, because the lesion's cross drive is at least 0.10).** Lesion cross drive +0.625 (+0.608 to +0.642).

> Without the connections layer 4 cells respond to both gratings (the orthogonal grating alone drives them at 0.625 of the preferred response), so the plain lesion cannot show suppression or dominance in group spikes; this pilot does not answer whether the connections compute the intact network's XOS or Dc.

## Validity checks

- **No refusals.** The `refusals` field of `results.json` is empty, and both runs are "analysed".
- **Frozen check passed.** The analysis checked the frozen manifest (`a1e2e2d92a7bfe48…`) and its 5 analysis files before reading any export. It also checked protocol 017's package against that package's frozen manifest (`383e67ef1bc8f614…`).
- **Identical input.** The LGN spike trains were identical between the runs for 3,600 of 3,600 LGN neurons: 1,800 ON and 1,800 OFF cells.
- **Identical neuron tables,** as declared for instance A in both runs, and an independent recount of the spikes matched. Layer 4 excitatory groups: 915, 721, 1,071 cells. Inhibitory groups: 208, 203, 272 cells.
- **The 30 % validation check passed** in both runs, against protocol 017's values:
  - intact, L4E group 45 deg, 45 deg single at 30 %: 3.30 against 2.97, ratio 1.112 (band 0.80 to 1.25);
  - intact, L4E group 135 deg, 135 deg single at 30 %: 2.28 against 2.18, ratio 1.046 (band 0.80 to 1.25);
  - intact, X_ON, 45 and 135 deg singles at 30 %: 8.03 against 8.06, ratio 0.996 (band 0.90 to 1.10);
  - intact, X_OFF, 45 and 135 deg singles at 30 %: 5.72 against 5.73, ratio 0.998 (band 0.90 to 1.10);
  - lesion, L4E group 45 deg, 45 deg single at 30 %: 9.80 against 9.80, ratio 1.000 (band 0.80 to 1.25);
  - lesion, L4E group 135 deg, 135 deg single at 30 %: 10.31 against 10.24, ratio 1.007 (band 0.80 to 1.25).
- **The connection lists are complete.** The export job finished with 0 errors and wrote 10 projections onto layer 4, 4,306,718 connections in all. Its chunked reader matched PyNN's own reader on the cross-check projection. V0's 7 declared structural checks all pass (files, counts, post_coverage, delays, synapses, weights, orientation). Each count equals the export job's log and protocol 017's intact run log at the printed precision.
- **The bootstrap** used 10,000 replicates, seed 2202, the same draw for every summary and both runs. The draw's SHA-256 is `ab6b4f38cb6e98fe…`.
- **The order.** The 96 analysed steps followed `sequence-r2.json`. No condition ran more than 2 times in a row, as declared.
- **The server files** were unchanged through every run (see The runs).

## What this shows, and what it does not

**What it shows.** In this one network instance, at these contrasts and this timing, the intact model circuit did not add two gratings like a relay. By all three declared measures it combined them differently from the sum of its single-grating responses, by at least the declared margin:
- an orthogonal grating lowered the response to the preferred one;
- at unequal contrasts the stronger grating's cells kept more of their response;
- the controller's readout of a plaid of gratings 45 degrees apart pointed further toward the stronger grating than the sum of its readouts of the two gratings.

**What it does not show, in the declared text's own terms.** "This is a property of the whole model circuit (input stage and cortex) at these contrasts and this timing, and the input stage alone was expected to produce part of it. It does not show that the intracortical connections compute it."

In plain words:
- **The input stage.** The input stage before the cortex is non-linear: the LGN stage saturates, and the LGN to layer 4 synapses depress strongly. Both are the same in the intact and the lesioned network. The pilot cannot say how much of the suppression or dominance they produce.
- **Why the lesion cannot settle it.** Suppression or dominance of the input shows in a group's spikes only when the cells are tuned to orientation. In this model the tuning itself needs the connections. Without them, the orthogonal grating alone drives the "45 degree" and "135 degree" cells at 0.63 of their preferred response, so the lesion cannot show suppression or dominance in group spikes. Its own values of these measures are reported (S1, S4), and as declared they are not read as suppression or dominance.
- **The readout without the connections** also pointed further toward the stronger grating than the sum of its single-grating readouts, by 0.440 signal units. The declared reading is "same direction without the connections". It notes that the lesion's own saturation can produce this, so the readout result does not show that the intact lean comes from the feedforward path either.

**The difference at the readout has no declared reading.** The intact lean is 0.097 signal units smaller than the lesion's (95 % interval 0.034 to 0.161). A reading of "connection-dependent at the readout" needed the opposite sign, so the difference is reported without a reading.

**Not a claim about cats.** The spec: "Neither outcome is a claim about cat V1. Simulation only."

## Limitations

- **One instance, one sequence.** One network instance (A, `pynn_seed` 5), one stimulus order (seed 231), one timing and one readout. The intervals cover the step-to-step noise of this one instance only. They say nothing about how other instances of the network would behave, which is the confirmatory protocol's job.
- **A reduced model.** The model is a 2 mm reduction of LSV1M, not the full published model and not a cat. Its input stage saturates less than cat LGN and has no pooled suppressive field, so its split between input stage and cortex is a statement about the model only.
- **Onset responses.** Counts over 210 ms after onset are dominated by the onset response. They cannot be compared with sustained (F1) measurements in cats.
- **Only layer 4.** As declared, only layer 4 is compared, because layer 2/3 is silent in the lesion: S9 gives 0 Hz for its excitatory and inhibitory cells in every condition of the lesion run. All three primary measures rest on layer 4, its groups and a readout of its excitatory cells.
- **Contrasts and timing.** Contrasts were 15 to 60 %, with each plaid component at most 30 %. Nothing in this model had been measured below 30 % before this pilot.
- **One spatial pattern per plaid.** Both components start at phase 0, so each plaid is one fixed spatial pattern, and it falls differently on the two groups (S10). In block A, 73.4 % of the 915 cells at 45° sit where the two gratings add in phase (mean cosine of the relative phase +0.390), against 33.4 % of the 1,071 cells at 135° (−0.284). In the 135° cells the additivity follows that phase: for the 15 + 15 % plaid, A is 0.36 in the quarter of the cells least in phase and 0.52 in the quarter most in phase, while in the 45° cells it stays between 0.47 and 0.52. So the group values, and the differences between the groups in S1, hold for this layout only. As the spec also notes, the plaids move at the intersection-of-constraints velocity, and the readout is π-periodic: it reads orientation, which repeats every 180°.
- **The readout shift pools two plaids.** ds is the mean of the two block B plaids, 4 steps each, and their values differ: +0.258 for 135° at 30 % with 90° at 15 %, and +0.427 for the mirror plaid (difference −0.169, 95 % interval −0.278 to −0.060). These two plaids also have the least balanced stimulus history in the design, as the spec notes: the steps before them had a mean summed contrast of 22.5 % and 48.75 %. Refitting with the previous step's contrast as a covariate changed ds by 0.002 (S12).
- **Different operating points.** The plain lesion removes all twelve intracortical projections together, and it changes the operating point: without the connections, layer 4 excitatory cells fired 6.4 to 8.7 times as often over whole steps of the stimulus conditions (S9, 0 to 357 ms after onset: the grating or plaid and the grey that follows it), and 10.2 times as often in the blank steps. The declared texts speak of "7 to 8 times". Inhibition and recurrent excitation are not separated here.
- **Precision of p.** p values at the bootstrap's floor (2 / 10,001) mean only that no replicate crossed 0. They are not more precise than that.
- **The noise model.** The declared misfit check compares each condition's raw residual variance with the fitted noise model. It is reported and never acted on. It flagged 5 of the intact run's 855 condition and summary pairs at p < 0.001, and 0 of the lesion run's 855. None of these pairs enters P1 to P3.
- **A reporting gap.** The spec lists a third XOS form beside the primary (S1): the relay-denominator form, 1 − D(plaid) / [D(preferred) + D(orthogonal)]. `results.json` has no field of that name. For each group and contrast it equals 1 − A, where A is the additivity of the equal plaid in S2, so its per-cell values and intervals can be read from S2. Its mean over the four cells, with an interval, is not in the output.
- **No decisions.** No decision, executor or token was involved. The decision sides in S5 come from protocol 017's rule applied offline to the readout.

## What follows

- **A confirmatory protocol** on fresh network instances, declared before its runs. This is the declared GO text's own next step.
  - The spec fixes what a GO hands on: "the triggering measure exactly as defined here, its conditions, and δ = 0.10".
  - The confirmatory sample size "comes from the pilot's noise parameters only, never from its estimates".
- **The declared follow-up on the exported connections.**
  - It rebuilds each layer 4 cell's afferent, recurrent and inhibitory conductances offline.
  - Its replay code is developed and validated only on protocol 017's runs, before it reads any pilot export. It must then reproduce the pilot's own lesion run before any reading on pilot data.
  - It is exploratory and never changes this pilot's outcome. It is the part designed to estimate how much of the suppression the input stage carries.
- **Status on 28 October 2026.** The confirmatory protocol has not been written yet. The follow-up was declared in this spec, and it has not been run.

## Before the freeze: the tier 2 incident

Before the freeze, a check on the server ("tier 2") had to confirm that the new stimulus code builds the declared frames with the pinned simulation software.
- **The first attempt failed** at 14:37 UTC on 27 October. Two checks passed, but all four stimulus checks failed with the same error, raised inside Mozaik: `AttributeError: 'NoneType' object has no attribute '__name__'`.
- **The cause** was a bug in the check script, not in the stimulus code or the design. The script loaded the two stimulus modules from their file paths but never registered them in `sys.modules`. The framework looks a class's module up there, so every stimulus the script built failed. The real runs import the stimulus code normally and would not have hit it.
- **The fix** registers each module before running it, the standard importlib recipe. Nothing else in the script changed.
  - The script's SHA-256 went from `5791f564…` to `1f5528f7…`.
  - The failed version is kept as `receipts/tier2/tier2_imagen_check-v1-failed-20261027.py`.
  - A stale freeze test was fixed at the same time, without changing any assertion.
  - The spec, the contract, the sequence, the stimulus code, the analysis and the queue did not change.
- **The rerun.** A first rerun, from 17:55 to 18:16 UTC, never reached the check, because the research server's SSH connection went silent. The launcher ran again at 18:22 UTC, and the check passed at 18:24 UTC: all six checks in 68.6 s. Its receipt is the tier 2 receipt the freeze requires, and the frozen manifest pins the fixed script.
- **The full record** is in `receipts/tier2/`: both receipts and logs, and `tier2-harness-fix-2026-10-27.md`.

## Redactions in this package

Six files contain a local workspace path, the research server's host name or a personal detail. They are published with those strings replaced and nothing else changed. The originals stay in the private workspace.

| File here | What was replaced | SHA-256 of the private original |
| --- | --- | --- |
| `results/results.json` | the local workspace path before `outputs/felisyn-research/protocol-017`, in field `protocol_017.package`, replaced by `<workspace>` | `f02a858e9038c71be75755155db06fb15d7d90003b83022a8a4d144ff7c0db2e` |
| `receipts/connections-v0.json` | the local workspace path in the checker's arguments (6 places) and input paths (4 places), replaced by `<workspace>` | `5143cc559ee7a1e575b98d4e655b6dfe416e0e17b28b7f4e4c94d7f58e785753` |
| `receipts/pilot/intact.log.gz` | the research server's host name in the last line (an MPI connection message), replaced by `redacted-host`; then gzipped | `640bd5363147fd2c7628043935638a9c4d6e0b0204e5fc916d647377dd4eb342` |
| `receipts/tier2/receipt-20261027T182426Z.json` | the research server's host name in field `host`, replaced by `redacted-host` | `6ed21ebd154e9315a92af7ae2e63673ee285db4bbd656d11187aa56fa2ab433c` |
| `receipts/tier2/receipt-20261027T143746Z.json` | the research server's host name in field `host`, replaced by `redacted-host` | `01d1c822f5e9bb79e63be025c3a90e902a0ceeffd19402093ddf57205ff4389a` |
| `receipts/tier2/tier2-harness-fix-2026-10-27.md` | a personal detail about the cause of the dropped connection, in the "Tier 2 rerun" bullet, replaced by `a local network outage` | `02b86fc646d2862e53f18c7c26619a183a92ec0471b85c5a69fd0ced517a66c7` |

The three run logs are published gzipped. Uncompressed, `lesion.log.gz` and `connections.log.gz` give the original logs byte for byte, and `intact.log.gz` differs from its original only by the host name above. The originals, uncompressed:

| Log here | Bytes | SHA-256 of the original |
| --- | --- | --- |
| `receipts/pilot/connections.log.gz` | 723,861 | `e2ae43188ad083b988dfb6b66ebba4e577ead8372c1b9bfaa1433f3b6827e48e` |
| `receipts/pilot/intact.log.gz` | 724,115 | `640bd5363147fd2c7628043935638a9c4d6e0b0204e5fc916d647377dd4eb342` |
| `receipts/pilot/lesion.log.gz` | 723,954 | `db591e085ef94806a0629b86ac03115c8e2691e42d94632599c0a6d91ae46b41` |

- **The receipt the fix record cites.** `tier2-harness-fix-2026-10-27.md` cites the passing tier 2 receipt by the SHA-256 of its private original, `6ed21ebd154e9315…`, which the frozen manifest also pins. The copy published here, with the host name replaced, has SHA-256 `a287ff76040ef75e…` (Files, below).
- **Private names left in place.** `design/analysis-choices-r2.md`, `receipts/connections-v0.json` and the failed tier 2 script name private project records: workflow runs by ID (`wf_…`), and files by their paths in the private workspace (under `work/` and `docs/`). Those records are not published. The files that name them are kept as recorded, so these names are left unchanged.

## Files

- `README.md`: this file.
- `results/results.json`: everything the declared analysis computed (redacted as listed above).
- `results/analysis-run-2026-10-28.log`: its run log, with outcome, time and memory.
- `design/analysis-choices-r2.md`: the analysis details the spec does not state, recorded before the runs.
- `receipts/pilot/`: the queue's receipts for the three server jobs:
  - start, end and exit receipts, the attempt and the command;
  - the FELISYN environment;
  - SHA-256 of the 34 server files at start and end;
  - `time -v`, memory and CPU-steal receipts;
  - the gzipped run logs;
  - the queue log.
- `receipts/connections-v0.json`: the declared structural checks of the connection lists (redacted).
- `receipts/tier2/`: the tier 2 receipts and logs (passing and failed), the fix record and the failed script version.
- `data-sha256.txt`: SHA-256 and size of the private exports and connection lists.
- `figures/`:
  - three figures as PNG and SVG;
  - `draw_plaids.py`, which draws them from `results/results.json` alone;
  - `captions.md`, with captions and alt text.
- `manifest.json` and `plaids-evidence.zip`:
  - `manifest.json` gives the size and SHA-256 of every file;
  - the ZIP holds every file and the manifest.

## Provenance

**Private data, by hash and size only.** The two run exports hold the recorded spike trains, and the eleven files of the connection lists are a data product for the declared follow-up. They are too large for this site and stay on the research server and in the project's private workspace. They were checked against the server's own SHA-256 when fetched. `data-sha256.txt` lists them as SHA-256, size in bytes and name.

| File (private) | Bytes | SHA-256 |
| --- | --- | --- |
| `exports/intact.json.gz` | 25,552,690 | `952b469b9a8e74a482f704854a0b5462f6c68798ba3466e1a15aa0e893eae281` |
| `exports/lesion.json.gz` | 46,569,465 | `42484edd1e15aec60ac58b8fffc5f160e5e325fe5ad598743860ed6cd4d92c05` |
| `connections/V1AffConnectionOff.npz` | 6,742,506 | `e357017bb0e05a502bd9412acf976736add77ee675c771fbfaeabd9481a5fc6c` |
| `connections/V1AffConnectionOn.npz` | 6,720,126 | `6855f3dab53f0484390c814f3cf51fd56d5fa928e9169c65ae40e98253b3af3b` |
| `connections/V1AffInhConnectionOff.npz` | 1,669,286 | `ceb6e0b22ebafaeafbaa304586137d6b5f83529a0e7d3c128b756b236a7f2a9c` |
| `connections/V1AffInhConnectionOn.npz` | 1,675,726 | `be4a8466ecc3a147d1b6f0f562d597b2344150a4770c8e42029801a711ecdd70` |
| `connections/V1L23ExcL4ExcConnection.npz` | 12,609,446 | `2480c7cbe6be128b16ee8294033b358fd7aaa5572510eab47f60048eeab00fd6` |
| `connections/V1L23ExcL4InhConnection.npz` | 2,220,466 | `75c30843aa949a9c98d01dc0738d86f19ea589ee10dbfa2edf5bc9fe0dc41cfa` |
| `connections/V1L4ExcL4ExcConnection.npz` | 34,545,986 | `500785864f712f233706ca9d838fb68266a04a243148a1ff2cf74ff71a14192f` |
| `connections/V1L4ExcL4InhConnection.npz` | 5,194,966 | `28b291351b0e6613b6e08084a8c2b7788abf31dbd94e4de5839b2d818470df08` |
| `connections/V1L4InhL4ExcConnection.npz` | 12,395,506 | `4d6b221818b0f425c1de934dea44f95cd668bf270022c11fa5c81aecc18d2374` |
| `connections/V1L4InhL4InhConnection.npz` | 2,370,206 | `47ab3bc30d21189a3c4fb72d230ad0c6998988915dbf1811bb8cb9a79ca26670` |
| `connections/index.json` | 1,307,646 | `1e6628b644d72c02338317bc22e5902448d0af6e08b1617dcc0f73a4f96762b7` |

**Every file in this package**, except this README, `manifest.json` and the ZIP. `manifest.json` also lists this README.

| File | Bytes | SHA-256 |
| --- | --- | --- |
| `data-sha256.txt` | 1,429 | `4c8d747ce832406ab5fe178a2175b71bee34578fef66fac71b1db6ac89856072` |
| `design/analysis-choices-r2.md` | 6,453 | `dc1c14290849d2d893a9070ed8418298ea48d7c6660bbbf17ee2fb3fad3a2a30` |
| `figures/captions.md` | 6,240 | `4b25b47568bc45410639893dff29f935d8a5136ff79d2f827325f2df78c4ee40` |
| `figures/draw_plaids.py` | 18,483 | `01f0d876d0840be7591cf110f55060fa5bbcbde7aa2259035726ddc2fc16faab` |
| `figures/plaids-primaries.png` | 248,810 | `051c103168b3b11e3c6a21426059936fafac7bc14768eaae374d212bdd8b7a84` |
| `figures/plaids-primaries.svg` | 15,419 | `3e9b733988f781c5317764ec50bc74cafa4f8ce549a3268385f97d5056b8dce1` |
| `figures/plaids-readout.png` | 227,719 | `38344ea8d32ace2dcd513ce67d4bf70b6ad2ffc10b3fa3d8192cdd1eddd2ac89` |
| `figures/plaids-readout.svg` | 28,237 | `3b4dcb36bddbb87abb3eda3c7c878e2d5e56996bcc235e344905c1101411b215` |
| `figures/plaids-suppression.png` | 236,641 | `c4a8cd685ef5a665476e1ad1a11800e63c21052b8090a668a579eefe2ea4846e` |
| `figures/plaids-suppression.svg` | 53,416 | `669fa6bf1b138fdeebbeaae2aaa64557d51e258ab3e37b1362681413cbca8c19` |
| `receipts/connections-v0.json` | 28,897 | `f5b87bb0f5cd4700d83a087f99cf2e2bfbd60c600b0c1c2544e5b61427f4b93c` |
| `receipts/pilot/connections-attempt.txt` | 2 | `4355a46b19d348dc2f57c046f8ef63d4538ebb936000f3c9ee954a27460dd865` |
| `receipts/pilot/connections-command.txt` | 237 | `02ed1f574b2d9f6d6c2b0c10c1599ba374f8176cf3de4cf68bdc6c7b6d6e6db7` |
| `receipts/pilot/connections-end-utc.txt` | 21 | `2b32ec3da4d1c6177897a4d6a9fac8b74b48cd8ac24e0f42db2a0bf4730723c5` |
| `receipts/pilot/connections-env.txt` | 343 | `8815d4a2f0f3ef97c6c9a36454bf716c56042721dad25f6f379794421581fa9e` |
| `receipts/pilot/connections-exit.txt` | 2 | `9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa` |
| `receipts/pilot/connections-files-sha256.txt` | 997 | `50e7c8c627a049c43cbbdbc68e2a0d08fce1be0ef299b2cbb78c99050c3e8027` |
| `receipts/pilot/connections-memory-end.txt` | 447 | `90455e71e948b609da52152da292ff66be19988b3d2015037df10de95a41c456` |
| `receipts/pilot/connections-memory-start.txt` | 440 | `9d22d1c28acbe31ed9ca15fde234a6ba95790f2894210c26b631788f1e690c7f` |
| `receipts/pilot/connections-sha256-end.txt` | 4,104 | `311180506010317097e83b25efba2c40a03db1426525adc8f34868d40eac3ac0` |
| `receipts/pilot/connections-sha256-start.txt` | 4,104 | `311180506010317097e83b25efba2c40a03db1426525adc8f34868d40eac3ac0` |
| `receipts/pilot/connections-start-utc.txt` | 21 | `d8d2ec609d7872baaa15555258c17c389c232f91674d785fcdef270c84beb197` |
| `receipts/pilot/connections-steal.txt` | 670 | `824781c6dd032fefe688c7ce4e221d63b68a24115c2c3f06473d07ede1094ab9` |
| `receipts/pilot/connections-time-v.txt` | 977 | `66da37c80dd48831a4010b7a4333140a2ce26c5ab8e7348ea640f440a58f2f05` |
| `receipts/pilot/connections.log.gz` | 23,571 | `16cf359aac99f883e1b666d063fa54ba6c8489db9dc646c356787152656fe8a0` |
| `receipts/pilot/intact-attempt.txt` | 2 | `4355a46b19d348dc2f57c046f8ef63d4538ebb936000f3c9ee954a27460dd865` |
| `receipts/pilot/intact-command.txt` | 230 | `83e6da6972e62ad407c35b5e38f8cf0cda95cfa3d1431a9d9173bbdfd0eac13a` |
| `receipts/pilot/intact-end-utc.txt` | 21 | `01b4a6b601f599f8cca57d6335c1754c26b101e863f2abe157b88b8c4435f1c9` |
| `receipts/pilot/intact-env.txt` | 312 | `799be91a4d2ba6796fbde8c452a85cb62034c493512576fbccc79a093c35b0b5` |
| `receipts/pilot/intact-exit.txt` | 2 | `9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa` |
| `receipts/pilot/intact-memory-end.txt` | 447 | `5cd0362263ae3c74461dcbcbc6f50848624c39387d221227be1a4c178092d57c` |
| `receipts/pilot/intact-memory-start.txt` | 439 | `531ee8fc3aa1c35daa0cfde505eae3445e268525f6fec576f5ed894a5f9c46d8` |
| `receipts/pilot/intact-sha256-end.txt` | 4,104 | `311180506010317097e83b25efba2c40a03db1426525adc8f34868d40eac3ac0` |
| `receipts/pilot/intact-sha256-start.txt` | 4,104 | `311180506010317097e83b25efba2c40a03db1426525adc8f34868d40eac3ac0` |
| `receipts/pilot/intact-start-utc.txt` | 21 | `2b7e37053931fc7979f8683e13eb9fc5429b525f2666a82933168390dcf8e9dc` |
| `receipts/pilot/intact-steal.txt` | 670 | `e3a3754aef112ccec9af67cb100988aa1e3b8027850ff033d3996a0205def79b` |
| `receipts/pilot/intact-time-v.txt` | 972 | `3a4726c9b5f6256417bf893fba3c31719169983dc2e72c45f8ee25c977ac8504` |
| `receipts/pilot/intact.log.gz` | 23,817 | `67647c51e9b406c85ffc61290bd4d65ed38ad9c7d6a312ed3fb9531e59b0f063` |
| `receipts/pilot/launches.txt` | 92 | `db5b8befd348a16d7bd7edc6b038737362a6bf9734447d76b946a4b5e6b699cd` |
| `receipts/pilot/lesion-attempt.txt` | 2 | `4355a46b19d348dc2f57c046f8ef63d4538ebb936000f3c9ee954a27460dd865` |
| `receipts/pilot/lesion-command.txt` | 524 | `3fa356621a05e148eba02a76d9ea3c2ba7e56c15bb48f22bf2ad61c8c0c37f01` |
| `receipts/pilot/lesion-end-utc.txt` | 21 | `d8d2ec609d7872baaa15555258c17c389c232f91674d785fcdef270c84beb197` |
| `receipts/pilot/lesion-env.txt` | 312 | `b65d592e385288ea6f3fdc3d35f6eb3fd693064a848f7aa73d7488f22e560a2b` |
| `receipts/pilot/lesion-exit.txt` | 2 | `9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa` |
| `receipts/pilot/lesion-memory-end.txt` | 447 | `9865ce2695f8f9bf55df57a3abc048f5d4513c769a9a83f0437322a34b893d7a` |
| `receipts/pilot/lesion-memory-start.txt` | 447 | `427b7b4ec0238ed3df6ba1b4586a14014706ddf1b01285ac149695aea704065d` |
| `receipts/pilot/lesion-sha256-end.txt` | 4,104 | `311180506010317097e83b25efba2c40a03db1426525adc8f34868d40eac3ac0` |
| `receipts/pilot/lesion-sha256-start.txt` | 4,104 | `311180506010317097e83b25efba2c40a03db1426525adc8f34868d40eac3ac0` |
| `receipts/pilot/lesion-start-utc.txt` | 21 | `01b4a6b601f599f8cca57d6335c1754c26b101e863f2abe157b88b8c4435f1c9` |
| `receipts/pilot/lesion-steal.txt` | 670 | `cd440732989f3824a526ff7f2bc3c372c697ed7046f96c328955858092718ac3` |
| `receipts/pilot/lesion-time-v.txt` | 1,266 | `40104c584d4414ca504ab4eb9998873362b78eb8e9457954111dad6b9748f786` |
| `receipts/pilot/lesion.log.gz` | 23,703 | `8133e103137079bd24fe4a3266cd9fe5ef9566da7df68305bac19a02e709be55` |
| `receipts/pilot/manifest-sha256.txt` | 83 | `aeef35f3e36a47baa175841bd8a1f70757fd153ee3293a8e590f7fb9f8853479` |
| `receipts/pilot/queue.log` | 908 | `308c353176bf5d4eebd1873b5d878f26fcbe9df058b4460a9bc57f5bd7d050a2` |
| `receipts/pilot/runs-done-utc.txt` | 21 | `d8d2ec609d7872baaa15555258c17c389c232f91674d785fcdef270c84beb197` |
| `receipts/tier2/receipt-20261027T143746Z.json` | 12,433 | `64be6af25a701d1a68239be7bb69a35dd1ac622c156a844f4b274fa491000030` |
| `receipts/tier2/receipt-20261027T182426Z.json` | 11,469 | `a287ff76040ef75e98aba5493179b892aa7746f4a53110bb0627da59b4170cca` |
| `receipts/tier2/tier2-20261027T143746Z.log` | 227 | `012681d0b4608e9658471d8c8ddbe0de8c0b8e1ec5ed7636d8f017a440f97625` |
| `receipts/tier2/tier2-20261027T182426Z.log` | 226 | `51cadccd7bc04dff2b2494e9232141fc8e9d7eec1498fb0d03dddf7e468d905f` |
| `receipts/tier2/tier2-harness-fix-2026-10-27.md` | 3,381 | `fc54916e2558bec3d11fba81134b39bd606f81342a2f6692b3aebbaa2f10eeb9` |
| `receipts/tier2/tier2_imagen_check-v1-failed-20261027.py` | 21,939 | `5791f564572781a3306e362f912d9217f71ef3fd3c25c06d497bd8ce738f3e38` |
| `results/analysis-run-2026-10-28.log` | 2,416 | `afab0e8cd6d5bf3c3eedf72075357878289e835a19753b0439e35bebdee7893e` |
| `results/results.json` | 493,869 | `d4e525d237c768eb917ec0fcf3835f5eebdfdb4ecc2ea8dbc2ed7a61d55e0707` |

## Credits and licences

- **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 20(8): e1012342](https://doi.org/10.1371/journal.pcbi.1012342), published under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
- **Software.**
  - The LSV1M model code v1.0 is under the MIT license.
  - The Mozaik simulation framework and PyNN are under CeCILL.
  - The NEST simulator (3.4) is under the GNU GPL, version 2 or later.
  - The imagen pattern library is under its BSD license.
- **Figures.** Set in Latin Modern Roman (GUST Font License), with STIX for the minus sign. Drawn with matplotlib.
- **This work.** The runs, analysis, figures and text are our own, made with protocol 022's frozen code.
