# EXP-46: does 0.83's episodic memory make the default brain learn the Stone Age chain, read state by state?

- Date: 2026-10-13 | Cadence version: 0.83.0 (`23dba71`) | Program iter: none (teaching track; Will 13:46: "highest value target is to test it for Bernhard")
- Script: `exp/EXP-46_episodic_memory_stone_age.py` | Evidence: `exp/EXP-46-pilot.*`, `exp/EXP-46-full.json/.jsonl/.out`

## Summary
- Question: Cadence 0.83 adds an opt-in episodic memory (`Brain.compose(memory_choice=True)`). Upstream reports, on its own four-step crafting chain, frozen full-chain success 38.1% -> 98.7% with a caregiver, 9.0% -> 97.8% after 80 episodes alone through `live`, 9.1% -> 73.2% from reward alone. EXP-45 showed that on Stone Age V5's chamber full-chain success is mostly in-attempt search (a no-knowledge never-repeat searcher gets .304) and that the default brain learns the chain from no signal (per-state greedy accuracy .33-.42). Does the memory change that on this outside chamber, state by state?
- Setup: EXP-45's replica of V5's chamber (reproduces V5's published per-seed numbers). Brain = the current default composition, `Brain.compose(7, 5, modules=(48, 24), arousal=True)`, with and without `memory_choice=True` (its own defaults, nothing tuned). Signals: natural (world reward only), grade-v5 (V5's guided +-.5), corrective (V5's caregiver labels on executed mistakes). 120 teaching attempts, then 80 attempts alone through `live` with world reward only. Readings on frozen copies after teaching and after the 80 alone: per-state greedy accuracy (primary), greedy full-chain, sampled full-chain over 100 attempts. Seeds 42-53.
- RUN GATE: (1) prediction below, written before the pilot. (2) Window: a 3-seed pilot at the exact run conditions; a 1-seed 10-attempt smoke run was seen (code check only: memory/corrective all four states right, without memory .25). If the pilot puts every memory arm at the ceiling, the full run keeps the design (the claim under test is a ceiling claim) and the write-up says the budget cannot rank memory arms. (3) Variance: seeds vary; arms differ only in memory and signal. (4) Metric: per-state greedy accuracy is primary; full-chain is read against the never-repeat searcher (.304). (5) Chain: train success -> taught acc -> taught sampled -> alone train success -> after acc -> after sampled.
- Prediction (written before the pilot):
  - Check, not a prediction: without memory, corrective on V5's seeds 42-46 lands on upstream's published 0.82 figure for V5 (26.4% sampled), since 0.83 without memory builds the 0.82 composition and the replica reproduced V5 on 0.81.
  - (P1) corrective + memory: per-state greedy accuracy >= .9 (mean) and sampled full-chain >= .9 after teaching (upstream 98.7%).
  - (P2) natural + memory: per-state greedy accuracy >= .75 and sampled full-chain above the never-repeat searcher (.304) after teaching (upstream 73.2%).
  - (P3) after 80 alone through `live`, every memory arm keeps per-state accuracy within .1 of its taught value (upstream 97.8%).
  - (P4) grade-v5 + memory: per-state greedy accuracy >= .75 (the memory makes V5's diluted grade usable on the default brain).
  - (P5) without memory, every signal stays at per-state accuracy <= .5 after teaching (EXP-45's default-brain result carries to the 0.82 composition).
- Decision rule (written before the pilot): P1-P3 hold -> upstream's 0.83 claim replicates on an outside chamber with the per-state instrument, and memory turns search into a learned chain; sampled high but per-state accuracy low -> the memory helps search, not the mapping; P5 fails -> the 0.82 composition alone already fixed part of it, and the memory gain is read against that.
- Pilot (3 seeds, 42-44, exact run conditions, `exp/EXP-46-pilot.out`), read before the full run: memory arms near the ceiling (taught per-state accuracy natural .92, grade-v5 .83, corrective 1.00), so as pre-registered the design is kept and the full run cannot rank memory arms against each other at this budget. Without memory, corrective reads .75 (EXP-45's 0.81 composition gave .42), so P5 is at risk: the 0.82 composition alone may have moved the default brain. No design change; the full run reuses the pilot's three seeds as rows (seeded, unchanged code) and adds seeds 45-53.
- Result (12 seeds; acc = per-state greedy accuracy with SEM; sampled = full-chain over 100 attempts; "after" = after 80 attempts alone through `live`):

  | arm | taught acc | all 4 right | greedy solves | taught sampled | after acc | after all 4 | after sampled |
  |---|---|---|---|---|---|---|---|
  | without / natural | .29 (.04) | 0/12 | 0/12 | .092 | .35 | 0/12 | .129 |
  | without / grade-v5 | .44 (.06) | 0/12 | 0/12 | .006 | .44 | 0/12 | .002 |
  | without / corrective | .58 (.04) | 0/12 | 0/12 | .193 | .58 | 0/12 | .276 |
  | memory / natural | .75 (.08) | 4/12 | 9/12 | .777 | .88 | 8/12 | .857 |
  | memory / grade-v5 | .90 (.05) | 8/12 | 11/12 | .917 | .96 | 11/12 | .997 |
  | memory / corrective | .94 (.03) | 9/12 | 11/12 | .963 | .85 | 5/12 | .955 |

  Check held exactly: without memory, corrective on V5's seeds 42-46 gives .264, upstream's published 0.82 figure for V5 (26.4%). P1 HELD (corrective + memory .94 accuracy, .963 sampled). P2 HELD at the threshold (natural + memory accuracy .75, SEM .08; sampled .777, far above the .304 searcher); bimodal: 2 of 12 seeds stayed at .25. P3 HELD as a retention claim, with a note: no memory arm lost more than .1 on average after 80 attempts alone (corrective -.083, grade-v5 +.062, natural +.125, a gain that a literal "within .1" reading would also flag), but corrective's all-four-right count fell from 9 to 5 of 12 while its greedy full-chain solves rose from 11 to 12: a stage answer drifts (most often has-wood, to craft or warm) and the carried context still gets the attempt through. P4 HELD (grade-v5 + memory .90). P5 FAILED: without memory, corrective reaches .58 (natural .29, grade-v5 .44); on the 0.81 composition (EXP-45) the same arm read .42, so 0.82's quiet context alone moved it.
- Verdict: SUPPORTIVE. 4 of 5 predictions held, P5 failed. Decision rule: P1-P3 hold, so upstream's 0.83 claim replicates on an outside chamber with the per-state instrument: on V5's chamber the memory turns the default brain from in-attempt search (sampled under the .304 searcher) into a learned chain (sampled .78-.96, per-state .75-.94), from every signal including reward alone. P5's failure means the memory's gain over the 0.82 composition is read from .58, not from EXP-45's .42.

## Reflection
- What it means: the strongest outside confirmation of a Cadence release claim we have produced. Upstream's own table (38.1 -> 98.7 corrected, 9.1 -> 73.2 from reward alone, 9.0 -> 97.8 after the caregiver leaves) lands on a chamber upstream did not build, read by an instrument (per-state greedy accuracy) that search cannot pass: .193 -> .963, .092 -> .777, .276 -> .955 sampled, and per-state .58 -> .94, .29 -> .75. Before the memory, the default brain's best full-chain numbers were search; with it, they are a chain. One thing upstream's table does not show and this does: after the caregiver leaves, the corrected chain's stage answers drift through `live` too (all four right fell 9 -> 5 of 12) while full-chain success holds, because the carried context covers the drifted stage. Upstream's caveat names the stall through `step` and says `live` avoids it; here the per-state reading drifts through `live`, and the full-chain reading hides it.
- Process: pre-registered and pushed before the pilot; pilot read before the full run; no design change. The replica check hitting upstream's 0.82 figure exactly (.264) is what lets this read as a test of the release and not of our chamber. Limits: the memory arms sit near the ceiling at 120 attempts, so this cannot rank them or show learning speed; one chamber, episodic with done flags (upstream's no-done continuing chain and the history-free bandit loss are not tested); 12 seeds; reward-alone is bimodal (2 seeds at .25), so its mean .75 sits exactly on P2's threshold; P3's wording ("within .1") was loose for an arm that improved.
- Limit found after the write-up (2026-10-13, from the Noesis cross-read): the chain has four states and teaching and testing use the same four, so this shows the memory brain ACQUIRES the chain, not that it GENERALIZES; a brain that records the caregiver's answers would score the same. A held-out-state test is the next experiment.
- Lesson: a release claim measured with full-chain success can be replicated or refuted only by an instrument search cannot pass; read per-state accuracy beside it, and the drift that full-chain success hides shows up.
- Next: learning speed at smaller budgets (10, 20, 40 attempts) where the memory arms are off the ceiling; the drift after the caregiver leaves (which stage, how fast, `live` vs `step`); the memory on our delayed-credit brain (EXP-45's winner) to see whether the two compose; and a note to upstream with these numbers (Will-gated).