# Felisyn confirm 023r2: protocol 023 revision 2 ended CONFIRMED over six fresh instances

The result package of protocol 023 revision 2, the confirmatory plaids protocol with paired runs, dated 13 November 2026. The protocol ran six fresh network instances of the 2 mm cat circuit model, each twice on identical inputs, and asked whether the plaids pilot's departure from a relay (protocol 022, [journal note 22](https://www.felisyn.com/journal/two-gratings-at-once)) holds on average over them. All six instances settled with an agreeing pair, the declared analysis ran once on the frozen package, and the declared outcome is **CONFIRMED** on all three measures.

- **Declared** at 2026-11-11T00:55:21Z, before any revision 2 run. Spec SHA-256 `b7072b0bc7c56d2fba3bea44a411e2847324e06258f774bd31b40ec3183af980`, recorded in the project's running record at the declaration. The design is published as [protocol-023r2-public](https://www.felisyn.com/research/protocol-023r2-public/README.md), with the [spec](https://www.felisyn.com/research/protocol-023r2-public/frozen/spec.md) (27 home folder paths redacted; that README says how to check it against the recorded hash).
- **Frozen** at 2026-11-11T09:57:37Z by a manifest with SHA-256 `de6693f140984f8dfc6f8dec638a5339129a79f21621537b040655fd47fe39be` (80 files).
- **Run** on the 8 GB laptop that replaced the lost research server: the queue started at 2026-11-11T09:58:51Z (`receipts/queue/launches.txt` names the frozen manifest's SHA-256) and ended at 2026-11-12T23:54:28Z, rc=0, after 37:55:37: 29 attempts, 14 completed runs, 15 attempts stopped at the disk floor and not counted, six instances settled by an agreeing pair.
- **Analysed once,** on the frozen contract, right after the queue: fetch, blind checks of the six exports to analyse (each computed twice), the summary, then `analyze_confirm_023.py` under memcap (126.5 s, peak footprint 0.958 GiB, exit 0, the analysis computed twice and written only when both computations agree). `results/results.json` was written at 23:59:27 UTC on 12 November (its file time). Its SHA-256, `0dedf8ad6ea1745e6937b26c1577366faf4a0962c2e07f57cfb1abccd19582e0`, was recorded in the project's running record before any other reading of this revision's exports.

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

## The declared outcome

The analysis wrote this text (`results/results.json`, field `text`), fixed in the spec before any run:

> On six network instances of the same 2 mm model that had never been built (pynn_seed 8, 10, 11, 12, 13 and 14: new cell positions and connections, each with its own stimulus order), the intact circuit combined two superimposed gratings differently from a relay of its own single-grating responses, on average over the instances, in the direction instance A's pilot found: an orthogonal grating lowered layer 4 cells' response to their preferred grating below the sum of their single-grating responses (cross-orientation suppression), mean XOS 0.586 (95 % t interval 0.548 to 0.624; one-sided p below 0.0001; instance values 0.641, 0.554, 0.581, 0.587, 0.542 and 0.609); 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), mean Dc 0.745 (95 % t interval 0.710 to 0.781; one-sided p below 0.0001; instance values 0.806, 0.761, 0.737, 0.716, 0.720 and 0.731); the controller's frozen 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, by a mean of 0.343 signal units (95 % t interval 0.313 to 0.373; one-sided p below 0.0001; instance values 0.342, 0.322, 0.344, 0.323, 0.398 and 0.327). For each confirmed measure the one-sided t-test of the six instance estimates passed Holm over the three measures at a family-wise level of 0.025, and the mean of the six estimates is at least the declared margin of 0.10; that is a condition on the estimates, not an interval. The instances share the model's one orientation map, its LGN noise stream and its plaid layout, and these contrasts, this timing and this readout, so the result covers new positions and wiring on that map only. It describes the whole circuit, input stage and cortex together, and does not show that the intracortical connections compute it. Simulation only; no claim about cat visual cortex.

Declared next step (field `what_follows`): "A confirmed XOS or Dc is a property of this model circuit as a whole, input stage and cortex together, as a mean over fresh instances of its map. Whether the cortex computes it stays open: that needs a comparison that keeps tuning (the inhibition-only variant) or the conductance follow-up declared in protocol 022, each in its own declared protocol. A confirmed ds changes no controller: a two-signal controller would still need its own declared encoder, with c1 + c2 at most 100 % and contrast independent of orientation."

The spec asks that two limits be printed beside any outcome text of this protocol. First, because its outcome texts call the instances ones that "had never been built": the server half of the seed search ran with round 3's script on 3 November 2026 at 00:29:38Z, the final script ran only on the Mac, and the server's files can no longer be read. Second, the paired-rule sentence: each analysed instance's export was one of two runs with identical inputs that agreed on every spike train, and S15 gives the share of completed runs that disagreed with their instance's agreeing pair, an estimate of the one-off fault rate that runs low. S15 here: 2 of 14 completed runs (0.14, 95 % Clopper-Pearson interval 0.018 to 0.43; the note in `results.json` says why this runs low and why the interval is approximate).

## What the protocol asked

In the spec's words: does the plaids pilot's departure from a relay hold, on average, over six fresh network instances of the same 2 mm reduction that have never been built? It is asked for each of the three measures that triggered the pilot's GO (XOS, Dc and ds), with the same definitions and the same margin of 0.10. Six slots (pynn_seed 8 to 13) and one reserve (14), each with its own declared stimulus order; seed 7 was built and seen by protocol 023, so it is left out, with 5 and 6. For each measure, a one-sided t-test of the six instance estimates in the pilot's direction, with Holm's correction over the three at a family-wise 0.025, and the mean at least the declared margin of 0.10 from 0. The pilot itself is [plaids-022](https://www.felisyn.com/research/plaids-022/README.md) and journal note 22.

**The paired rule**, revision 2's answer to the one-off execution fault that ended protocol 023 ([journal note 24](https://www.felisyn.com/journal/the-confirmatory-test-stopped-at-its-first-run)): each instance is run twice with identical inputs, and it counts only when the two runs agree on every one of the 18,600 neurons' spike trains and on every field of the export but the run name and the run time, with both blind checks passed and the same. A disagreement starts a third and, if needed, a fourth run; the earlier run of the agreeing pair is the one analysed. A refused export is a one-off fault unless an agreeing pair repeats it; protocol 023's early stop is replaced by this rule.

## The result

| Measure | Pilot (instance A) | The six instance estimates | Mean | 95 % t interval | One-sided p |
| --- | --- | --- | --- | --- | --- |
| Suppression, XOS | +0.584 | 0.641, 0.554, 0.581, 0.587, 0.542, 0.609 | +0.586 | +0.548 to +0.624 | 9.4e-08 |
| Dominance, Dc | +0.713 | 0.806, 0.761, 0.737, 0.716, 0.720, 0.731 | +0.745 | +0.710 to +0.781 | 2.0e-08 |
| Readout shift, ds | +0.342 | 0.342, 0.322, 0.344, 0.323, 0.398, 0.327 | +0.343 | +0.313 to +0.373 | 4.3e-07 |

Holm's thresholds at family-wise 0.025 are 0.0083 (Dc), 0.0125 (XOS) and 0.025 (ds); all three p values pass, every mean is at least 0.10 from 0, and on every measure all six instances lie on the pilot's side (the instance estimates are in pynn_seed order 8, 10, 11, 12, 13, 14). The figures in `figures/` plot these values from this package's own records.

## What ran

| Run | Start (UTC) | Wall clock | Peak footprint (GiB) | Outcome |
| --- | --- | --- | --- | --- |
| `s8-a1` | 11-11 09:58 | 1:16:58 | 3.17 | in the agreeing pair |
| `s8-a2` | 11-11 11:16 | 2:23:32 | 3.18 | completed; its cortical trains differ from a1's (the pair disagreed) |
| `s8-a3` | 11-11 13:41 | 3:03:24 | 3.19 | in the agreeing pair |
| `s10-a3` | 11-12 01:18 | 1:54:28 | 3.17 | in the agreeing pair |
| `s10-a4` | 11-12 03:13 | 1:27:47 | 3.18 | in the agreeing pair |
| `s11-a1` | 11-12 04:42 | 1:13:47 | 3.18 | in the agreeing pair |
| `s11-a2` | 11-12 05:56 | 1:20:32 | 3.20 | in the agreeing pair |
| `s12-a1` | 11-12 07:17 | 1:37:44 | 3.20 | in the agreeing pair |
| `s12-a2` | 11-12 08:56 | 1:39:35 | 3.17 | in the agreeing pair |
| `s13-a1` | 11-12 10:36 | 1:12:38 | 3.17 | in the agreeing pair |
| `s13-a2` | 11-12 11:50 | 1:10:38 | 3.19 | completed; refused by the blind check (lead-in identity, X_ON 1,799 of 1,800) |
| `s13-a3` | 11-12 13:01 | 1:26:58 | 3.18 | in the agreeing pair |
| `s14-a5` | 11-12 18:13 | 2:38:40 | 3.19 | in the agreeing pair |
| `s14-a6` | 11-12 20:53 | 2:59:38 | 3.18 | in the agreeing pair |

Attempts not in this table were stopped at memcap's disk floor of 3 GiB before completing (15 in all: nine on seed 9, two on seed 10, four on seed 14), are not counted under the declared rule, and have their receipts here. `receipts/queue/disk-waits.txt` lists the queue's 13 waits for free disk.

- **Slot 9 fell to the reserve.** All nine attempt names of seed 9 (`a1` to `a9`) were consumed by disk-floor stops in 7 h 46 min, with zero completed runs; the disk was being filled by another program on the laptop. At 00:31:52 UTC on 12 November the queue settled s9 "attempts used up", and the reserve, seed 14, took the slot under the declared replacement rule (`receipts/queue/replacements.txt`). No slack remained: every remaining seed had to settle valid.
- **Two one-off faults among the 14 completed runs, both absorbed by the paired rule.** s8-a2 completed and passed the blind check, but its cortical spike trains differ from a1's from 541.4 ms on (`receipts/queue/s8-a1-a2-compare.json`); a third run agreed with a1 on every train. s13-a2 was refused by the blind check itself: 1,799 of 1,800 X_ON lead-in trains identical (`receipts/queue/s13-a2-server-check.json`), the same fault class that ended protocol 023, this time costing one extra run instead of the protocol; a third run agreed with a1. S15 (in `results/results.json`) records every seed's runs.
- **No seed needed its fourth completed run,** no run was killed for memory, and no agreeing pair was refused.

## The checks and the comparisons

The blind checker (unchanged from protocol 023 but for the replacement of the early stop) reads only the run's records, the 30 % single gratings and the LGN lead-in, computes no plaid quantity, and computes its verdict twice. It ran right after each completed run (the "server check"; since the move to the laptop that machine is the laptop) and again, after the fetch, on each analysed export (the local check, under a second Python; `checks/check-*.json`, all VALID). The pair comparison (`controller/compare_runs_023.py` in the frozen package) reads spike times and records only: verdicts and differences are in `receipts/queue/s*-compare.json`. The summary of every seed by the paired rule is `checks/instances.json`.

## Limits

The spec states nine limits of the Mac run and asks that they stand beside every result of this protocol. In short:

1. **A new platform.** The pilot ran on x86-64 Linux; these runs ran on an arm64 Mac (Apple M2, 8 GB, macOS 26.2, Apple clang, libc++, one rank), where cortical voltages can differ from the server's in the low bits.
2. **Two engine patches,** the only changes to pinned sources: the LGN noise pair order and NEST's block size ([laptop-move](https://www.felisyn.com/research/laptop-move/README.md)).
3. **A memory wrapper** around the unchanged runner, shown byte-identical on a short run.
4. **The second-machine check is lost.** Both checks run on the laptop, under two Pythons, so a fault of this machine could show in both.
5. **The server's half of the seed search:** the limit quoted beside the outcome above.
6. **What the laptop's identity with the pilot rests on:** the tiny regression and G4, one instance on one sequence.
7. **One shared laptop,** on a nearly full disk; another program's disk use cost 15 attempts and one slot here.
8. **What G4 shows:** protocol 022's pilot reproduced exactly on this laptop, spike for spike; it does not show that a fresh seed behaves.
9. **One-off faults and the paired rule.** Runs on this Mac fail now and then in a way only an exact repeat shows. The paired rule catches such a fault anywhere in the export, at the cost of more runs. It does not catch: the same fault in both runs of a pair (an argument, not a measurement, puts an identical pair below about 1 in 700,000 under the leading untested candidate); a deterministic fault that every run of an instance repeats (the blind checks, the engine pinning and G4 still catch classes of these, but a deterministic platform difference that changes only cortical spikes of a fresh seed would pass, as in protocol 023); a fault striking both computations of a check or the analysis the same way; and a systematic fault that varies run to run, which would end seeds "no agreeing pair" and show in S15.

Beyond the spec's nine: the result is a mean over six instances that share the model's one orientation map, its LGN noise stream, its plaid layout, these contrasts, this timing and this readout (the outcome text states this); and the between-instance intervals rest on six values, so they are wide relative to the within-instance ones.

## Redactions

Every file here is a byte-for-byte copy of its source in the project's private workspace except those listed in `manifest.json` under "redactions": 250 files differ only by exact replacements (the laptop's home folder path, its host name, its battery identifier), each token absent from the original, so putting the original strings back gives the original files; the builder checked that for every one. The manifest gives each original's SHA-256, the replaced lines and the counts. In summary:

| Files | How many | Copies replaced in each | What was replaced |
| --- | --- | --- | --- |
| `checks/check-s*-a*.json` | 6 | 1 | the home folder's path, replaced by `<home>` |
| `checks/instances.json` | 1 | 1 | the home folder's path, replaced by `<home>` |
| `receipts/queue/launch-environment-*.txt` | 1 | 1 to 3 | the home folder's path, replaced by `<home>`; the laptop's battery identifier, replaced by `id=redacted`; the laptop's host name, replaced by `redacted-host` |
| `receipts/queue/queue.log` | 1 | 1 | the home folder's path, replaced by `<home>` |
| `receipts/queue/s*-a*-archive.mem.json` | 29 | 3 | the home folder's path, replaced by `<home>` |
| `receipts/queue/s*-a*-compare.mem.json` | 8 | 7 | the home folder's path, replaced by `<home>` |
| `receipts/queue/s*-a*-env.txt` | 29 | 4 | the home folder's path, replaced by `<home>` |
| `receipts/queue/s*-a*-memory-end.txt` | 29 | 1 | the home folder's path, replaced by `<home>` |
| `receipts/queue/s*-a*-server-check.json` | 14 | 1 | the home folder's path, replaced by `<home>` |
| `receipts/queue/s*-a*-server-check.mem.json` | 14 | 4 | the home folder's path, replaced by `<home>` |
| `receipts/queue/s*-a*-sha256-end.txt` | 29 | 554 | the home folder's path, replaced by `<home>` |
| `receipts/queue/s*-a*-sha256-start.txt` | 29 | 554 | the home folder's path, replaced by `<home>` |
| `receipts/queue/s*-a*.log.gz` | 29 | 1 to 3 | the home folder's path, replaced by `<home>` |
| `receipts/queue/s*-a*.mem.json` | 29 | 1 | the home folder's path, replaced by `<home>` |
| `results/analysis.mem.json` | 1 | 9 | the home folder's path, replaced by `<home>` |
| `results/results.json` | 1 | 1 | the home folder's path, replaced by `<home>` |

The 29 model run logs (`s*-a*.log`) are published gzipped (level 9, mtime 0) after their replacements; decompressed, each reverses to its original. The six analysed exports are byte-for-byte copies, unredacted (they hold no private string).

## What is here

| Path | What it is |
| --- | --- |
| `results/results.json` | The declared analysis's output: outcome, fixed text, the rule's statistics, every instance's measures, S15 and the frozen check (redacted: one path). Original SHA-256 `0dedf8ad6ea1745e6937b26c1577366faf4a0962c2e07f57cfb1abccd19582e0`. |
| `results/analysis.mem.csv`, `results/analysis.mem.json` | The analysis's memcap log and record (126.5 s, peak footprint 0.958 GiB, exit 0). |
| `checks/check-*.json` | The local blind check of each analysed export (all VALID, computed twice). |
| `checks/instances.json` | The summary of every seed by the paired rule, from the queue's receipts, the comparisons and the checks: six valid instances, analysed 8, 10, 11, 12, 13, 14. |
| `receipts/queue/` | The queue's 597 receipts as fetched, byte-identical to the queue's own: the launch record and environment, the queue log and exit, the per-seed statuses, the replacement record, the disk waits, and every attempt's command, environment, times, exit, memory, heat, run-folder hashes (before and after), archive record and log; every completed run's server check with its log and memory; the eight pair comparisons with their logs. |
| `receipts/data-sha256.txt` | The fetch's record of all 14 completed exports' sizes and SHA-256 on the laptop. |
| `exports/` | The six analysed exports, byte for byte (the earlier run of each agreeing pair): spike trains of all 18,600 neurons, neuron tables and run records, about 145 MB in all. The other eight completed exports stay in the private workspace; their SHA-256 are in `receipts/data-sha256.txt`. |
| `figures/` | The two figures of journal note 25 (PNG and SVG), their CSV data, the two scripts that made them from this package's published records and the pilot's published results, and their captions with alt text. |

## Credits

- **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), [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). Its code, LSV1M v1.0 (MIT), ran with the 2 mm reduction's overrides; none of its files is copied here.
- **Software.** Mozaik (CeCILL), PyNN 0.11.0 (CeCILL), NEST 3.4 (GPL) with the NEST step current module (GPL) and two patches offered under the terms of the files they change, GSL 2.7.1 (GPL).
- **This work.** The protocol, the runs, the checks, the comparisons, the figures and their scripts are Felisyn's own.
