# decision-map-025: protocol 025, the decision map on fresh instances

21 November 2026. Journal note 28, "A decision map on fresh instances".

**Outcome: PASS.** On three network instances built for this test (pynn_seed 15, 16, 17; no earlier run of them was
found, see below), the rule of protocol 024, unchanged, took the intended action in 56 of 56 scored steps on every instance, 168
of 168 in all (the rule needed at least 51 per instance and 160 in all). Computed twice; the two result files are
byte-identical (`results/map-a.json`, `map-b.json`).

An engineering rule on model runs. No market signal, no action taken. Simulation only; no claim about cat visual
cortex.

## What was run

- The rule of protocol 024 with nothing changed: z is protocol 017's frozen readout (all layer 4 excitatory cells,
  balanced weights, spike counts 40 to 300 ms after onset); z0 the mean z of the instance's 8 blank steps; S half the
  difference of the imaginary parts of the mean z for A alone and B alone at 30 % (8 steps each; refused below
  1,500); v = −Im(z − z0)/S; action B if v > 0.32, action A if v < −0.32, otherwise hold.
- Signal A is a grating at 45°, signal B one at 135°, superimposed for one step (210 ms on, 147 ms grey). Strengths
  are contrasts on a 5 × 5 grid: 0, 7.5, 15, 22.5, 30 % for each signal. Four steps per cell, eight in the three
  calibration cells: 112 analysed steps per instance after one warm-up step, in an order drawn per instance.
- **Scored (56 steps per instance):** equal strengths above zero: hold (4 cells); strengths that differ by 15
  points or more: the action of the stronger (10 cells, including one signal alone at 15 or 22.5 %).
- **Described only:** the eight cells whose strengths differ by 7.5 points (the transition zone), and the
  calibration cells.
- Each instance was built and run twice. It counts only if the two exports are equal in every field except the run
  name and the run time. All three pairs agreed at the first two runs (8,384,489 spikes in instance 15); no run
  failed, none disagreed, the reserve instance 18 was not used and never built.

## Result

| Cells | Intended | Steps as intended, three instances | Mean v per cell and instance, range |
|---|---|---|---|
| Equal: 7.5, 15, 22.5, 30 | hold | 48 of 48 | −0.07 to +0.04 |
| A stronger by 15 or more (5 cells) | A | 60 of 60 | −0.95 to −0.66 |
| B stronger by 15 or more (5 cells) | B | 60 of 60 | +0.55 to +1.02 |

- Scale S per instance: 5,530, 4,967, 5,704.
- Margins: the largest |v| among steps meant to hold is 0.180, the smallest among steps meant to act is 0.470. The
  closest scored step lies 0.140 from a threshold.
- **Transition zone (described only), steps of the three instances:**

| Cell | A | hold | B | Mean v per instance |
|---|---|---|---|---|
| A 7.5 alone | 6 | 6 | 0 | −0.31 to −0.29 |
| B 7.5 alone | 0 | 11 | 1 | +0.25 to +0.27 |
| A 15 + B 7.5 | 12 | 0 | 0 | −0.50 to −0.43 |
| A 22.5 + B 15 | 12 | 0 | 0 | −0.55 to −0.43 |
| A 30 + B 22.5 | 10 | 2 | 0 | −0.50 to −0.34 |
| A 7.5 + B 15 | 0 | 0 | 12 | +0.40 to +0.43 |
| A 15 + B 22.5 | 0 | 3 | 9 | +0.34 to +0.44 |
| A 22.5 + B 30 | 0 | 1 | 11 | +0.42 to +0.43 |

  In the six two-signal cells, 66 of 72 steps took the stronger signal's action and 6 held; none took the weaker
  signal's action. One signal alone at 7.5 % sits at the threshold: 7 of 24 steps act.
- **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 27 of the 30 scored two-signal cells (10 per instance). It holds in all three
  instances on "A 15 + B 30" (relay v +0.24 to +0.31; the circuit's own mean is +0.62 to +0.78).
- Run times 1 h 1 min to 2 h 40 min per run, peak memory 3.19 to 3.21 GiB, all under the cap of 3.7 GiB.

## What must be read beside the outcome

- **Fresh runs, declared first.** The frozen manifest was written at 2026-11-13T14:58:29Z; the first run started at
  14:58:30Z. A search of the run folders and the workspace on 13 November found no run of pynn_seed 15 to 18, and no
  plan of the project named one; an earlier research server is lost and could not be searched.
- **Partly known beforehand.** Of the 14 scored cells, 4 were also cells of protocol 024 (15 + 15, 30 + 30, 30 + 15
  and its mirror) and 2 more are its single gratings at 15 %; its result made a PASS on those likely. The other 8
  contain 7.5 or 22.5 % and had not been shown to this model. A rehearsal of the tools before the declaration showed
  four cells of this grid for one step each on an older instance (`frozen/smoke-result.json`); the readouts were
  not used for the rule.
- **Margins are not comparable with protocol 024's.** That test had twice the instances and twice the steps per
  cell; wider extremes are expected there.
- **Three instances, four steps per cell.** A 168-of-168 score bounds the error rate of scored steps only loosely
  (with no error in 168, a rate up to about 1.8 % is still compatible at the 95 % level).
- **The scored cells leave the hard cases out by design.** Differences of 7.5 points were declared as a transition
  zone and not scored. The map there is the descriptive result: the rule usually acts for the stronger signal, and
  sometimes holds.
- **No lesion runs.** Protocol 024 found that the rule also works on the pilot with the cortex's own connections at
  zero weight. Nothing here changes that: a PASS says the rule is reliable on this model, input stage and cortex
  together.
- **Narrow.** One orientation pair, one timing, one orientation map shared by all instances, single-step decisions,
  calibration from steps of the same run, strengths up to 30 %.
- **Scoring date.** The runs ended on 14 November at 03:39 UTC. The results were computed on 21 November with the
  frozen tool; nothing was changed in between (the tool's hash is in the result file and in the frozen manifest).
- **Review before declaration:** one independent reader; 1 high finding (disk headroom of the run queue), 3 medium,
  all applied before the freeze. The review itself is not published.

## Contents

- `frozen/`: `spec.md` (the declared text), `queue_025.sh`, `decide_025.py`, `compare_025.py`,
  `archive_mozaik_025.py`, `make_sequence_025.py`, `test_025.py`, `sequences/`, `ref/`, `smoke-result.json`,
  `MANIFEST-frozen.txt`. The frozen manifest lists the originals' hashes; where a published copy has a private path
  replaced, `manifest.json` gives the original's SHA-256 under `redactions`.
- `results/`: the two result files.
- `exports/`: the analysed export of each instance (the earlier run of its pair). The later run's hash is in
  `manifest.json` and in `receipts/runs/sK-a1-a2-compare.json`.
- `receipts/`: the queue's log and output, and for each of the six runs its command, environment, start and end
  times, exit code, memory record, machine state, archive record and log (logs gzip-compressed after the path
  replacement); the comparison record and status of each instance.
- `figures/`: the two figures of note 28 with data, scripts and `captions.md`.
- `manifest.json`: SHA-256 of every file here.

Inputs published elsewhere: protocol 017's readout (`protocol-017`), the rule and its threshold (`decisions-024`),
the lead-in reference (`confirm-023r2`).

## 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.
