RP002A — Preregistration v0.1

Question: In finite partially observed stochastic processes, when does recursively updated history recover predictive distinctions unavailable in current observation?

Scope: finite/discrete/stochastic/partially observed/passive; stationary within episodes; known predictive query; no agent actions/self-modification.

Worlds: E0 present sufficient; E1 observation aliases predictive states/history disambiguates; E2 history varies but is future-irrelevant.

Baselines: B0 current observation; B1 finite raw history; B2 fixed-size recursive learned state; B3 generator-defined oracle.

Primary hypothesis: history-sensitive representations outperform B0 in E1, but not systematically in E0/E2.

Controls: generator sanity checks before model comparison. Primary metric candidate: log loss with effect size/uncertainty and calibration diagnostics.

Interpretation lock: does not establish universal/fundamental memory or full-history retention.

Decision matrix: E0 advantage → leakage/capacity investigation; E1 no advantage → generator/model/optimization diagnosis; B1≈B2 → no demonstrated recurrent-state special advantage; unexpected → SURPRISE/diagnostic before revision.