# decisions-024: protocol 024, the two-signal decision rule

13 November 2026. Journal note 27, "A rule for two signals at once".

**Outcome: PASS.** On the six network instances of protocol 023 revision 2, the declared rule took the intended action
in 56 of 56 test steps on every instance, 336 of 336 in all (the rule needed at least 51 per instance and 320 in all).
Computed twice; the two result files are byte-identical (`results/decisions-a.json`, `decisions-b.json`).

An engineering rule on recorded model runs. Held data analysed again, not fresh runs. No market signal, no action
taken. Simulation only; no claim about cat visual cortex.

## The rule

Two signals arrive together as two superimposed gratings for one step (210 ms on, 147 ms grey): signal A at 45°,
signal B at 135°, each signal's strength as its grating's contrast.

- z(k): protocol 017's frozen readout of step k (all layer 4 excitatory cells, balanced weights, spike counts 40 to
  300 ms after onset), unchanged.
- Calibration per instance: z0 = mean z over the 8 blank steps; S = half the difference of the imaginary parts of the
  mean z for A alone and for B alone at 30 % (8 steps each). Refused if S < 1,500.
- v(k) = −Im(z(k) − z0) / S. Action B if v > 0.32, action A if v < −0.32, otherwise hold.
- The threshold 0.32 was fixed on the pilot's instance A (the midpoint between the largest |v| of its equal plaids,
  0.22, and the smallest |v| of its single gratings at 15 %, 0.42) before the six instances were read for the rule.

## Result

| Stimulus (contrast in %) | Intended | Steps as intended, six instances | Mean v per instance, range |
|---|---|---|---|
| A 15 + B 15 | hold | 48 of 48 | −0.18 to +0.03 |
| A 30 + B 30 | hold | 48 of 48 | −0.15 to +0.09 |
| A 30 + B 15 | A | 48 of 48 | −0.87 to −0.70 |
| A 15 + B 30 | B | 48 of 48 | +0.60 to +0.82 |
| A alone, 15 | A | 48 of 48 | −0.76 to −0.64 |
| B alone, 15 | B | 48 of 48 | +0.50 to +0.64 |
| A alone, 60 | A | 24 of 24 | −1.43 to −1.23 |
| B alone, 60 | B | 24 of 24 | +0.99 to +1.28 |

- Scale S per instance: 4,837 to 5,533 (pilot: 5,586).
- Closest calls: the largest |v| among steps meant to hold is 0.294, the smallest among steps meant to act is 0.404.
  One test step lies 0.026 from a threshold.
- **Relay rule, reported beside it (no decision role):** the same thresholds on the sum of the two single gratings'
  mean v gives the intended action in 19 of 24 plaid-and-instance cases. It holds in 4 of 6 instances on
  "A 15 + B 30" (relay v +0.06 to +0.22 there) and acts in 1 of 6 on "A 30 + B 30" (relay v −0.40). The circuit's own
  mean response is on the intended side in 24 of 24.
- **Block B, described only:** a grating at 90° alone gives hold (48 of 48 steps); B at 30 % with 90° at 15 % gives
  B (24 of 24); B at 15 % with 90° at 30 % gives hold (24 of 24), although B alone at 15 % gives B. A stronger third
  orientation masks a weaker signal; the rule was not designed for that case.
- The tool's readout equals the readout protocol 023's analysis published for every one of the 576 steps
  (`published_z_checked`).

## What must be read beside the outcome

- **Held data, known in part.** The six exports already existed, and protocol 023's analysis had published each
  instance's per-step readout in `confirm-023r2/results/results.json`; the outcome could have been computed from
  that file. The declaration states that no per-step or condition-mean readout of the six was read, printed or used
  for the rule before it (only key names were listed). The held-out status rests on that statement. The three
  published measures (XOS, Dc, ds) made a PASS the expected direction.
- **The lesion run passes too.** On the pilot's instance A the rule gives 56 of 56 with the cortex's own connections
  at zero weight (`frozen/design-pilot.json`). A PASS says the rule is reliable on this model, input stage and cortex
  together, not that the intracortical connections compute the decision. No lesion runs exist for the six instances.
- **Narrow.** A 3 × 3 grid of strengths (0, 15, 30 %) plus singles at 60 %, one orientation pair, one timing, one
  orientation map shared by all instances, single-step decisions, calibration from steps of the same run, a threshold
  chosen on one instance.
- **Review before declaration:** one independent reader; 1 high finding (the spec first said the six instances'
  per-step readouts had not been computed, which was false; corrected as above), 2 medium, 4 low, all applied before
  the freeze. The review itself is not published.

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

1. Design on the pilot only (`frozen/explore_pilot.py`, `frozen/design-pilot.json`).
2. Tool, tests, spec; review; fixes.
3. 14:05:41 frozen manifest written and its SHA-256 recorded in the project's running record: the declaration.
4. Right after (both runs done by 14:06:11): the tool run twice on the six exports: PASS, identical files.

## Contents

- `frozen/`: `spec.md` (the declared text; one sentence of private planning replaced, recorded in the manifest),
  `decide_024.py`, `test_decide_024.py` (15 tests), `design-pilot.json`, `explore_pilot.py`, `MANIFEST-frozen.txt`.
  The frozen manifest lists the originals' hashes; where a published copy has a replacement, `manifest.json` gives the
  original's SHA-256 under `redactions`.
- `results/`: the two result files and the two run logs.
- `figures/`: the two figures of note 27 with data, scripts and `captions.md`.
- `manifest.json`: SHA-256 of every file here.

Inputs published elsewhere: the six exports (`confirm-023r2/exports`), the pilot (`plaids-022`), protocol 017's
readout (`protocol-017`), protocol 022's frozen `plaid_stats.py` (`protocol-022-public`).

## 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 rule, its test and the figures are ours.
