When does history help? The RP002A application
Autorite series · 4/4 · Application. Three synthetic worlds, seven comparisons and the completed study’s limits.
The previous chapters opened a question, discussed its conceptual stance and followed a method for testing it. This chapter reads that sequence through RP002A: the boundaries of history usefulness.
This application is a finite prediction study on specified synthetic generators, not a deployed product. The account uses the public revision inspected on 10 October 2026, including the second confirmation and package review.
Three worlds
- E0: The present observation provides sufficient task information.
- E1: The present observation aliases different states; observed history can add information.
- E2: Histories vary without adding information relevant to the future task.
Together these conditions expose alternatives to the narrative that more history always helps. They are not claimed to represent every real-world setting.
Four comparison points
B0 uses current observation, B1 a finite history window, and B2 a learned recursive state. B3 knows the generator law and conditions on observed history, without privileged access to realized hidden states.
A difference from B2 therefore does not establish that recurrence alone caused it or is necessary for every task. The comparisons concern specified algorithms under a fixed training budget.
What the completed study found
The second confirmation completed 60 independent units and 120 learned-model fits, with 20 independent replicates per world. All seven predeclared outcomes were retained: two benefit, four equivalence and one indeterminate.
In E1, both finite history and the recursive predictor met the declared benefit criterion over current observation. The four E0/E2 comparisons met predictive-loss equivalence within the stated tolerance.
This equivalence does not establish identical representations or the absence of every historical effect. It concerns predictive loss for the fixed algorithms within the declared tolerance.
Why a small difference remained indeterminate
The final E1 comparison asked whether B2 improved on B1 beyond a predeclared benefit threshold. Its adjusted difference interval was approximately −0.000853 to −0.000568 nats; the required threshold was −0.01 nats.
A negative interval did not clear that criterion. The result remained indeterminate, without being switched to a post-outcome equivalence claim. It is a completed decision, not missing execution.
The margin is a fixture-specific engineering tolerance. Statistical interpretation retains the declared independence and normal-theory assumptions for seven comparisons. It is not an assumption-free or distribution-free guarantee.
A closed package, open questions
The first confirmation remained incomplete and separately preserved. The second attempt was independently authorized; samples and results were not pooled across attempts. Replay of artifacts, saved predictions and analysis was reported. This article starts no new experiment and does not claim to have rerun that audit here.
An internal package review closed RP002A for its defined Cycle 01 commitment. One of 13 cycle obligations is resolved; the cycle and broader representation/history questions remain open. This is not external peer review or a general memory theory.
Returning to the opening question, the application supplies no universal answer. It shows a more specific route: boundaries of useful history become researchable when the distinctions and measurements are explicitly stated.
Continue the series
Sources and revision
Public Autorite revision inspected on 10 October 2026. Links are pinned to the same commit. Explanatory examples are editorial interpretations; reported findings retain their source scope.
This connects to
Chapter 4 of the Autorite series.
Explores the Autorite research program through this step of the series.