# A1B8 — the action-only reminder is sufficient for the observed switch

**DIAGNOSTIC-COMPLETE. The registered action-reminder prediction was supported on this
four-input panel.** With the same genuine positive-return raw history, adding only
`a0 posted t1:h.` switched WORK to valid POST. An outcome-only reminder did not switch
the action; the full chronological chain elicited POST as before.

Canonical record: [ecology-culture-a1b8](https://chuk-experiments-server.fly.dev/#/experiments/ecology-culture-a1b8).
Run: `RUN-20260912-190820-00808`. Prospective registration and exact inputs: `3b7708a`.

## Complete result

| Case | Added note | Response |
|---|---|---|
| A: raw history | None | WORK t0 |
| B: action reminder | a0 posted t1:h. | POST t1:h |
| C: outcome reminder | a0 received an additional 6 resources. | WORK t0 |
| D: full chronological chain | a0 posted t1:h. After a1 read that message and worked t1, a0 received an additional 6 resources. | POST t1:h |

Every response parsed and executed in first-action validation. Both POSTs were valid
`POST t1:h`. The registered valid-POST indicator contrasts were **B−A = +1**, **C−A = 0**,
**D−A = +1**. A executed WORK and B executed valid POST, meeting the exact primary
criterion. The outcome-reminder-flip and full-chain-only patterns were both false.

A and D reused A1B7's complete raw/none/useful and raw/chronological/useful requests.
Both original responses reproduced exactly. There was one call per case, in A/B/C/D
order, with no model-generated continuation or retry.

## What changed, and what was held fixed

All four inputs contained the same **genuine positive-return scripted history** from
A1B5: donor POST/WORK/WORK, recipient READ of its valid hint followed by WORK, contribution
settlement six, donor final RES eight. The raw record, current fresh observation, system
prompt, objective, model and inference settings were identical. The only input change
was the presence/content of `history_note`; no counterfactual framing or invented payoff
was introduced. The short reminders restated facts already present in the source history.

This is a test of how the supplied history is presented. It does not measure a newly
earned return: each model reply was validated only as the first action in a fresh world,
and no new complete payoff episode or adoption trajectory ran. The historical trajectory
was generated by the scripted world, not by the model itself.

## Interpretation boundary

The action-only reminder was sufficient to elicit valid POST **for this checkpoint,
history, prompt and fresh state**. The full recipient/outcome narrative was therefore
not necessary for this first-action switch. Merely highlighting the six-resource outcome
was not sufficient in the tested form. The result is consistent with action priming or
imitation, but does not identify either mechanism: salience, token/semantic repetition,
demonstration following and other note effects remain possible explanations.

This does not show that payoff never matters, that the model generally copies prior
actions, or that norms will propagate in a swarm. Payoff was fixed positive; no zero-return
control or alternative demonstrated action ran in A1B8. Four temperature-zero requests
are not independent agents or population estimates. The positive criterion is a finite
observed switch, not a significance result or a general causal-learning claim.

A1B7 had already found no incremental effect of Because over After. A1B8 now narrows
the textual requirement further: the action reminder alone elicits the same POST on
this source history. A separately registered checkpoint comparison could reuse these
exact four requests to compare reminder sensitivity, but would not by itself isolate
training class or establish an internal learning mechanism.

## Verification and chronology

Server design registered at 19:05:07 UTC on 2026-09-12. Exact design, inputs and source
were uploaded/hash-verified at 19:08:54 UTC before checks. All **13/13 gates passed**,
including genuine history replay, unchanged note-free inputs, exact A/D anchors, factual
reminder content, absence of action references in the outcome-only note, whole-response
refusal and synthetic scoring branches. Gates were uploaded and verified at 19:09:28 UTC
before inference. No gate amendment was needed.

Inference ran 19:10:12–19:10:20 UTC: **4 calls, 7.96 seconds, 18 generated tokens**.
Qwen `qwen3.5:9b`, digest
`6488c96fa5faab64bb65cbd30d4289e20e6130ef535a93ef9a49f42eda893ea7`,
Ollama 0.33.3, temperature zero, thinking disabled, no context/output cap overrides.
Metadata matched before/after. Maximum reply nine bytes; no thinking, transport error,
retry or parse/execution refusal. The pinned A1B5 clean binary and its durable source/
build provenance were reused without engine changes.

The audit matched all four frozen inputs to complete saved calls, replayed every first
action, independently recomputed the three indicator contrasts and verified that only
the note varied. It checked source hashes and prospective chronology; no additional
model calls ran. Result SHA-256:
`6aa34e1929bf6fb1bd9c1b1ab8bda790c9e9488affac26c93658e18ac610c650`.

The canonical server stores queryable metrics, complete calls, source, gates, audit,
write-up and conclusion. Local receipt: `a1b8_server.json`. No successor inference or
model-class comparison is registered by this result.


Downloadable evidence:

- [a1b8-preregistration](https://drive.google.com/file/d/1PoHcEkKn4BZTfffslA4pU-yTqmItZFYq/view)
- [a1b8-frozen-source](https://drive.google.com/file/d/1l6O_RZZXPJMplcJf6wOTtRHqJatQ7OGJ/view)
- [a1b8-gates](https://drive.google.com/file/d/1Yjh9Q9eyZySIhid3BxnpjsMQyECy2zmy/view)
- [a1b8-detailed-result](https://drive.google.com/file/d/1lWAg78s0GTu-M90PSSa_VnCuEUWxpozQ/view)
- [a1b8-results-note](https://drive.google.com/file/d/1YlPEzSRmTUHoUbp5-y7oeH3Y9KQumDce/view)
- [a1b8-evidence-bundle](https://drive.google.com/file/d/1P0hMB2rltyiuFktbUEfATjQWv79fC1FP/view)