RP002A exploratory convergence-stage review

Date: 2026-10-09. Run: convergence-20261008-001 (the predeclared run ID retains its original proposal date). Completed once; no retries, extensions or confirmatory evidence.

Approval and provenance

The user approved this separately proposed stage by replying “devam” to the explicit 20-unit, 40-fit, 15-minute freeze/execution request. CONVERGENCE_PLAN.json was frozen and passed CI before docs/research/CONVERGENCE_APPROVAL.json anchored the exact plan and source. The execution source was 6fa2c6a009eb704192fa6a49218c4f54f3629aed; plan SHA-256 was 1dfdabfbf39d2890bd7853b2018b1e8e9765460d13052606a8d4ea2cda18bdc4.

Python 3.9.6, NumPy 2.0.2 and PyTorch 2.8.0 on macOS arm64 CPU, two threads. Elapsed stage wall time: 82.630 seconds, within the declared 900-second deadline. All 20 world/replicate units, 40 learned-model fits and 80 model assessment records completed. E1 had 10 independent replicates; E0/E2 had five each. Each used fresh 256/128/128 training/validation/assessment episodes of 129 observations. Assessment data were excluded from checkpoint selection.

Raw immutable artifacts preserve datasets, seeds, checkpoints, traces, probabilities, losses, per-replicate records, summaries and environment. The manifest lists 322 hashed outputs; it is an additional file. No realized latent-state array was supplied to a predictor.

Artifact review

The audit checks all output hashes and file coverage, the approval/source/plan binding, observed-data replay from the seed schedule, B0/B3 reference replay, saved B1/B2 checkpoint predictions, categorical losses and paired replicate summaries. Checks passed. Replaying checkpoints evaluates saved models without further training; tolerance is explicit and does not claim bit-identical optimization across platforms. CI also checks that committed run bytes and file sets remain unchanged.

Descriptive assessment contrasts

Difference is alternative minus reference natural-log loss; negative values favor the alternative on these exploratory assessment samples. Each observation in the summary is one independent training-replicate mean. Timepoints and episodes are not extra training replicates. SD is a descriptive sample SD, not a confidence interval or power guarantee.

E1 B1/B2 assessment losses were lower than B0 in this stage. The B2-minus-B1 mean is much smaller than the previously proposed 0.01-nat effect threshold. E0 learned models remained above the current-observation table; E2 differences were small. These observations are limited to the declared generators, data sizes, implementations and training budgets. They are not confirmatory benefit/equivalence decisions, novelty evidence or proof that recurrence is necessary. No inferential intervals or p-values were generated.

Convergence diagnostics

A best-near-cap flag means all 50 epochs completed and the selected best epoch was in the final ten. A late-improvement flag means validation loss fell by more than 0.001 nats over the final five recorded epochs. These were declared before outcomes; they are descriptive warnings, not convergence tests.

Overall, 33/40 fits reached the cap and selected their best checkpoint near it; 26/40 continued improving late. The stage therefore does not establish settled optimization. The CPU time allowance was not exhausted; the separate fixed epoch cap was binding for many fits. It was not extended after observing these outcomes.

Disposition and next design work

The approved exploratory stage is complete. The unchanged baseline and numerical confirmatory proposal remain distinct; no scientific Claim was promoted and no Cycle obligation was resolved (0/13).

The measured SDs describe this budget and data size only. Do not automatically transfer them to the proposed 4096-training-episode confirmation setting or select a confirmatory replicate count from them. Small samples and changes in optimization/data can change variability.

Before confirmation, propose a separately frozen convergence plan addressing the binding epoch cap, with fresh seeds and validation-based choices, rather than revisiting these assessment samples. Keep the information-matched references and the weak ideal window-eight/full-history gap visible. Then justify the final optimization budget, variance assumptions, effect/equivalence margin, uncertainty calibration, sample counts and complete analysis implementation. This report authorizes no additional run or discretionary tuning.