# EXP-58: does the delayed-credit brain take in what a teacher says, where the default memory brain does not? - Date: 2026-10-14 | Cadence version: 0.83.0 (`23dba71`, `~/_cadence_study`) | Program iter: none (teaching track; Noesis-lite gate: no channel has carried content so far) - Script: `exp/EXP-58_content_delayed_credit_brain.py` (EXP-53's chamber, protocol, seeds and readings; EXP-45's delayed-credit brain) | Evidence: `exp/EXP-58-pilot.*`, `exp/EXP-58-full.json/.jsonl/.out` ## Summary - Question: in EXP-52, 53, 54 and 55, a teacher that always gives the right action was worth no more than one giving random labels (after mistakes, at every decision, with and without upstream's surprise rule), on the default composed brain with episodic memory. EXP-45 found a second operating point, the key-door protocol-3 "delayed-credit" brain (actor step .1, bias step .01, lam .95, gamma .95, critic rate 5, modules (32,), working-memory amplitude .3, consolidation .25, no arousal, no episodic memory), that learned the Stone Age chain from every signal. Is the missing content a property of teaching through `step(teacher=...)`, or of the default brain? - Setup: EXP-53 exactly (EXP-48's weather chamber, V5's corrective schedule: the label after an executed mistake in a guided window, 20 attempts alone, per-state greedy accuracy on unseen weather, EXP-53's student seeds and world draws), with EXP-45's delayed-credit brain built for 10 inputs. Amended before any run: this brain has no arousal genes and Cadence's `live` requires them, so the 20 attempts alone run through `step` with world reward only and no labels (the smoke test, not a pilot, surfaced this). Twins per student: know-4 (right action), random labels, untaught. Lesson lengths 40 and 120 attempts, 24 students per length. - RUN GATE: (1) prediction below, written before the pilot. (2) Window: a pilot of 3 students per length at exact run conditions; if the untaught twin already sits at ceiling (per-state > .9 at 40), lengths drop to 20 and 40. (3) Variance: paired twins share seed and world draws. (4) Metric: content = know-4 minus random, paired. (5) Chain: slower actor and strong critic -> a label nudges the policy toward its content instead of just perturbing it -> know-4 above random. - Prediction (written before the pilot): - (P1, headline) The delayed-credit brain takes in content: at 120 attempts, know-4 minus random is >= .10 and more than 2 SEM. - (P2) Know-4 teaching beats no teacher at 120 by more than 2 SEM. - (P3) Random labels give this brain no lift: random minus untaught at 120 is below +.05 (on the default brain, being corrected at all was worth +.25 at 40, EXP-53). - Decision rule (written before the pilot): P1 holds -> content does get through `step(teacher=...)` on the right brain; the default brain's operating point is what blocks it, and Noesis-lite v2 runs on this brain against the random-label null. P1 fails -> neither brain takes in content through this path; the next channel is settled `fit()`-style lessons. - Pilot (3 students per length, `exp/EXP-58-pilot.out`), read before the full run: content +.25 at 40 and +.44 at 120, positive for every pilot student; random labels slightly below untaught (-.06, -.08); untaught .52 and .65, not at ceiling, so the lengths stay; know-4 reaches 1.00 at 120 (a ceiling on that arm, content still readable). No design change; the full run reuses the pilot rows. - Result (24 students per length, paired twins; per-state accuracy on unseen weather after 20 attempts alone; `exp/EXP-58-full.*`): | lesson | know-4 | random labels | untaught | content = know-4 - random (SEM) | know-4 - untaught (SEM) | random - untaught (SEM) | |---|---|---|---|---|---|---| | 40 | .635 | .443 | .458 | +.193 (.033) | +.177 (.030) | -.016 (.035) | | 120 | .932 | .464 | .542 | +.469 (.041) | +.391 (.041) | -.078 (.032) | P1 HELD (content +.469 at 120, 11 SEM). P2 HELD (know-4 beats untaught by +.391, 9.5 SEM). P3 HELD (random labels give no lift: -.078). - Verdict: SUPPORTIVE, 3 of 3. On the delayed-credit brain, what a teacher says reaches the student through `step(teacher=...)`: a right-answer teacher takes it to .93 per-state accuracy on weather it never saw, and a random-label teacher is worth nothing (slightly below no teacher). This is the first channel in the series (EXP-52 to 58) with measurable content. Decision rule: P1 held, so Noesis-lite v2 runs on this brain against the random-label null. - Confound, stated: two things differ from EXP-53, the brain's operating point and the 20 attempts alone (through `step`, because this brain has no arousal genes for `live`). The content difference is measured within this setup and stands; attributing the default brain's failure to its operating point rather than to `live` needs EXP-61 (the default brain with the alone phase through `step`). ## Reflection - What it means: the four nulls before this (EXP-52 to 55) were not "teaching through labels cannot carry content in Cadence"; on a slower-actor, strong-critic brain without episodic memory, it carries a lot. That fits EXP-45 (the same brain learned the Stone Age chain from every signal while the default brain learned from none) and EXP-55's side check (on the default brain, behaviour follows the episodic vote, not the cortex the labels teach). For Noesis-lite: there is finally something to price, and a market can now be tested against the form-matched random-label null on this brain. - Process: pre-registered; a smoke test (not a pilot) surfaced that `live` needs arousal genes, and the protocol was amended before any data and the amendment recorded; pilot read before the full run; ran beside EXP-59 on the remaining cores. Limits: one chamber, the corrective schedule only, 24 students per length, the know-4 arm near ceiling at 120 (.93), and the confound above. - Lesson: when one setup returns null after null, change the thing the setup holds fixed (here the brain) before concluding the channel is empty; one operating point is not the library. - Next: EXP-61 (default brain, alone phase through `step`) to split the confound; then Noesis-lite v2 on this brain.