# three-step-decisions-026: protocol 026, decisions over three consecutive steps

21 November 2026. Journal note 29, "Three steps did not beat one".

**Outcome: PASS, at the minimum.** On three network instances built for this test (pynn_seed 18, 19, 20), the
declared rule (the action of the mean of three consecutive readings) took the intended action in 26, 25 and 24 of 26
scored blocks, 75 of 78 in all. The pass marks were at least 24 per instance and at least 75 in all. Computed twice;
the two result files are byte-identical (`results/steps-a.json`, `steps-b.json`).

**The remedy under test did not help.** In the 36 close blocks, where one signal is 7.5 points ahead, the mean of
three steps decided for the stronger signal in 33. The first step alone did so in 33 as well. The declaration said
beforehand that a PASS alone would not show three steps to be better than one; the within-run comparison shows no
gain.

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

## What was run

- Per step, v is protocol 024's, unchanged: z is protocol 017's frozen readout; z0 the mean z of the instance's 8
  single blank steps; S half the difference of the imaginary parts of the mean z for A alone and B alone at 30 %
  (8 single steps each; refused below 1,500); v = −Im(z − z0)/S. **A block's decision is the action of the mean of
  its three v:** B above +0.32, A below −0.32, hold between.
- Signal A is a grating at 45°, signal B one at 135°; each step 210 ms on and 147 ms grey. A block shows one pair of
  strengths for three steps in a row. Per instance: one warm-up step, 28 blocks and 24 single calibration steps,
  109 steps, in an order drawn per instance.
- **Scored, 26 blocks per instance:** 12 close blocks (the six pairs 7.5 points apart, twice each): the stronger
  signal's action; 8 equal blocks (7.5, 15, 22.5, 30 % on both, twice each): hold; 6 clear blocks (A 15, B 15,
  A 30 + B 15, A 15 + B 30, A 22.5 + B 7.5, A 7.5 + B 22.5): the stronger's action.
- **Described only:** 2 weak blocks per instance (one signal alone at 7.5 %).
- Each instance was built and run twice and counts only if the two exports are equal in every field except the run
  name and the run time. **Seven runs were made.** Instance 18's first two runs disagreed
  (`receipts/runs/s18-a1-a2-compare.json`); the declared third run equalled the first on all
  7,873,595 spikes, so the first was analysed. Instances 19 and 20 agreed at their first two runs. No run failed;
  the reserve instance 21 was not used and never built.
- Declared at 10:07:26 UTC on 21 November 2026 (frozen manifest SHA-256 `163f6288…2268b5dc`); the first run started
  at 10:07:35 UTC; the queue ended at 21:54:44 UTC.

## Result

| Kind | Intended | Blocks as intended | Steps as intended | Mean v of a block, range |
|---|---|---|---|---|
| Close, A ahead by 7.5 (3 pairs) | A | 18 of 18 | 54 of 54 | −0.59 to −0.45 |
| Close, B ahead by 7.5 (3 pairs) | B | 15 of 18 | 43 of 54 | +0.27 to +0.45 |
| Equal | hold | 24 of 24 | 72 of 72 | −0.15 to +0.04 |
| Clear | the stronger's | 18 of 18 | 54 of 54 | −0.93 to −0.67 and +0.50 to +0.74 |

- Per instance: 26, 25, 24 of 26. Scale S: 5,498, 5,051, 5,781.
- **The three misses** are all holds where B led: A 22.5 + B 30 in instance 19 (mean +0.315) and in instance 20
  (+0.309), and A 15 + B 22.5 in instance 20 (+0.270). No block and no step took the weaker signal's action.
- **Close blocks under other readings (declared reports, no decision role):** first step 33 of 36, second 30,
  third 34, mean of the first two 32, mean of all three 33.
- **No change over repeats:** the mean v toward the stronger signal in the close blocks is 0.441 at the first step,
  0.441 at the second and 0.440 at the third.
- **The two sides are not symmetric** (read after the result, not declared; three instances). By construction the calibration means of A alone and B alone at 30 % differ
  by exactly 2. They were −1.04 and +0.96 in instance 18, −1.10 and +0.90 in 19, −1.14 and +0.87 in 20: the grey
  screen's reading lies nearer the B side, so the same threshold leaves B less room. Equal blocks lean the same way
  (−0.15 to +0.04).
- **Weak blocks (described):** A alone at 7.5 % held in instances 18 and 19 (−0.28, −0.30) and acted in 20 (−0.35);
  B alone at 7.5 % held in all three (+0.18 to +0.23).
- Every single calibration step decided as its own stimulus says (blank hold, A 30 A, B 30 B; 72 of 72).
- Run times 1 h 17 min to 2 h 18 min per run; peak memory 3.19 to 3.22 GiB, under the cap of 3.7 GiB.

## Limits

- The pass is at the minimum on two of three marks (24 in one instance, 75 in all). One more held block would have
  given FAIL.
- The first step of a block is a within-run baseline, not a repeat of protocol 025: it often follows a three-step
  block of another pair.
- One block length, one way to combine steps, one orientation pair, one timing, one orientation map shared by all
  instances, calibration from single steps of the same run, three instances, 12 close blocks per instance.
- No runs without the cortex's own connections: the rule reads the whole model, input stage and cortex together.
- Instance 19's second run (s19-a2) shared the machine with an unrelated short model run from 18:22 to 19:07 UTC.
  Its export equals the first run's in every field compared.

## Files

- `frozen/`: the declared spec, the queue, the scoring tool, its tests, the sequences and `MANIFEST-frozen.txt`.
- `results/`: the declared results, computed twice.
- `exports/`: the analysed export of each instance (the earlier run of its agreeing pair).
- `receipts/`: the queue's log, every run's log, command, environment, memory record and export hash, the comparison
  records, and the lists of model, queue and runtime files.
- `figures/`: the two figures of note 29 with their data, scripts, captions and alt text.
- `manifest.json`: every file with its SHA-256; copies in which a private path was replaced are listed with the
  original's hash; exports not published here are recorded by hash.

Model: 2 mm reduction of LSV1M (Antolík et al. 2024, CC BY 4.0), simulated with Mozaik and NEST 3.4.
