A repeated rationale does not accumulate evidence

Persistent agents can revisit a thesis many times while receiving very little new evidence. We distinguish two clocks: the number of reasoning turns and the number of observations that can actually change the thesis. That distinction matters when a long history makes a position look well supported simply because the same initial argument has appeared in many summaries.

Our operating-layer controls research reported fabricated rules in historical sell traces and a decline after a combined harness intervention. The intervention included treating previous decisions as context rather than precedent. We use that bounded finding to motivate a broader theory of evidence accumulation, rather than repeat the claim that every retained rationale becomes an invented rule.

Twelve turns can contain one observation

Imagine an illustrative agent that sees a venue announcement at noon and opens a position. It revisits the position once per hour. Each summary says that the announcement supports the trade. By midnight, the history contains twelve confident references, but the underlying announcement still contributes one source event. The agent's later confidence may reflect familiarity with its own text more than additional information about the market.

Now compare a second history with the same twelve turns. It includes the announcement, a verified implementation update and an independently observed change in venue conditions. Those observations may still be correlated, and each has a particular relevance. The histories differ because the second introduces new claims whose truth can be checked independently of the earlier rationale.

Bayesian notation makes the error precise. A likelihood factor belongs to evidence received under a declared model. Copying an earlier observation into a new summary does not create another conditionally independent draw. Multiplying the same likelihood factor twelve times would manufacture confidence. A model need not literally calculate that product to exhibit the analogous behavioral failure: repetition can make a proposition seem progressively more settled.

We therefore distinguish validated memory from accumulated wording. Validated memory records what was observed, when it became available and which source or settlement record supports it. A rationale records why the model chose an action given that evidence. Reusing the rationale can help reconstruct behavior; assigning it the evidential role of a fresh observation changes the inference.

A theory test that holds the event fixed

We propose comparing two histories at the same decision cutoff. Both contain identical market facts, owner instructions and settled account state. One repeats the earlier rationale across retained turns. The other preserves one copy plus a count showing how often the agent revisited it. The model receives the same substantive evidence in both conditions.

The measured difference would be thesis confidence, stated evidential support and the proposed action. A change caused by rationale multiplicity would identify a sensitivity to repetition under that history construction. It would still require a separate experiment to determine whether ordinary runtime summaries create the same effect.

A second contrast adds a genuinely new observation with a known relationship to the thesis. This checks whether the agent responds more strongly to repeated wording than to incremental evidence. Include observations that support and weaken the thesis. Otherwise an apparent improvement could be a generic reduction in trading or confidence rather than better evidence use.

The economic outcome remains downstream. Reduced overconfidence could avoid an oversized action, preserve an appropriate position or produce unnecessary abstention. Account constraints and payoffs determine which occurs. A useful memory design must support these distinctions without using eventual profit as proof that the original reasoning was justified.

This theory asks a sharper question than whether an agent remembers its past. What did the latest turn add that could change the justified belief? We want persistent agents to retain the continuity needed for account management while allowing an old explanation to remain old. More turns should mean more opportunities to inspect evidence, rather than more votes cast by the agent's earlier self.