RP002A remaining budget-warning review

Date: 2026-10-09. Post-outcome planning review of convergence-20261009-002; no new observations or model training. The recorded convergence disposition remains unresolved.

Evidence and reproduction

design/CONVERGENCE_002_DIAGNOSTICS.json records 40 per-fit validation diagnostics and six world/model summaries. Its source manifest hash and 80 input checksums bind validation episodes, saved training-only B0 tables and recorded validation traces. Reproduce with python3 scripts/rp002a-budget-diagnostics.py. This operation does not read assessment observations or losses, train models, draw new data or amend immutable run files.

B0 and the observed-history B3 filter are evaluated on the same checkpoint-selection validation episodes. Learned losses are taken from the recorded minimum-validation checkpoints. Those validation episodes already influenced model selection, so these comparisons are optimistic planning diagnostics, not independent evaluation. Learned trace losses use float32 cross entropy; references use clipped probability loss. Precision and clipping conventions remain visible. A sampled reference loss is not an empirical lower bound that every learned predictor must exceed.

Where the remaining warnings occur

Of 12 near-cap warnings, E0 accounts for eight (B1: three, B2: five) and E1 for four (B1: three, B2: one). E2 has none. Among these flagged fits, the last-20-epoch endpoint decline ranges from -0.000140 to +0.000111 nats; the last-50-epoch decline ranges from -0.000027 to +0.000265 nats. Positive values mean lower loss at the endpoint. Small endpoint changes do not rule out oscillation, slow progress or later improvement; per-fit last-20 loss ranges remain in the JSON.

E0 gaps and endpoint changes are small on these reused validation samples. E1 B1 remains farther from the observed-history filter than B2. This cannot separate finite-data error, approximation limits, optimizer behavior or selection noise, and it does not imply a fundamental need for recurrence. The earlier ideal window-eight/full-history gap is only about 0.000012 nats. No convergence threshold or scientific effect margin is changed using this review.

Next design choice

Do not reopen this run or retroactively remove its 12 warnings. The next useful question is whether the intended confirmatory data sizes are computationally feasible and how validation traces behave there. Stages 001/002 used only 256 training episodes; their runtimes and variances cannot be silently transferred to 4096 episodes.

BUDGET_TRANSFER_PROTOCOL.md and BUDGET_TRANSFER_PLAN.json propose one fresh replicate in each world, six learned-model fits total, at 4096/1024/2048 training/validation/assessment episodes. The pilot is for runtime and trace behavior only. One replicate per world provides no between-training variance estimate, power justification or scientific decision. The proposal remains unfrozen and unauthorized; its runner and artifact audit require implementation and review before requesting a one-run approval.

After that separately reviewed feasibility step, revisit optimization and statistical calibration rather than copying a favorable exploratory SD into confirmation. The baseline and numerical confirmation proposal remain unchanged; no Claim or Cycle obligation is resolved.