RP002A target-data-size feasibility review

Date: 2026-10-09. Run: budget-transfer-20261009-001. Completed once; exploratory feasibility only.

Disposition

All three independent units and six learned-model fits completed in 348.175 seconds (5 minutes 48 seconds), within the separately authorized 3600-second allowance. Each unit used 4096 training, 1024 validation and 2048 assessment episodes of 129 observations. All fits stopped after 20 epochs without a new validation minimum. None triggered the predeclared near-cap, final-20 or secondary final-five warning. This records absence of these warnings at this setting; it does not prove convergence. The earlier stage's 12 warnings remain part of that immutable record.

There is one independently trained replicate per world. This pilot supplies no between-training SD, inferential interval, power justification, equivalence or hypothesis decision. It does not justify five confirmatory replicates or guarantee full-workload resources. RP002A and Cycle 01 remain PLANNED; Cycle resolutions remain 0/13. No scientific Claim is promoted and no further execution is authorized.

Runtime and validation traces

The total includes data generation, training, selected-checkpoint assessment and artifact writing through the recorded timer. Per-fit times are measured locally and are not cross-platform benchmarks. Lifetime peak process resident memory was 263634944 bytes (251.42 MiB). The macOS raw RUSAGE_SELF.ru_maxrss value is 263634944 bytes; this includes imports and all fits, not incremental allocation or per-model memory. The auditor verifies recorded unit conversion, not the historical peak measurement itself.

Single-replicate assessment diagnostics

Natural-log loss, equally weighted across held-out episodes. Lower is better. These are sampled descriptive diagnostics; episodes/timepoints do not increase the number of independently trained replicates. Checkpoint selection used validation only. B3 conditions on observed history and has no privileged realized hidden-state access.

No across-world pooling or comparison with earlier fresh-seed stages is used for a causal budget effect. Assessment outcomes cannot select a follow-up optimization budget. The prior analytic window-eight review and missing statistical calibration remain relevant; a learned-model difference does not establish necessity of recurrent memory.

Provenance and audit

• Execution source: ea5cd5188681cefd94a83b40afa8d32bc71f1f8d.
• Frozen plan SHA256: e0e55a02902b350cbab04f23f558ce69bad3fa82ef5eb98ef51ed2c0ee65174e.
• Base configuration SHA256: fdd50646ee1e5e6ae659091a0fc9e3e1b6ff47bea09b8f3e1471e8bac0cf34ee.
• Manifest SHA256: 78ae8dfe2b50544ccfe1838a4d3d47a75f0393a3bf536800f09f7a4de9cb1d77.
• Separate approval: docs/research/BUDGET_TRANSFER_APPROVAL.json, bound to frozen source cd0bc4c751517bf84998a406d3ed233609577196, unchanged baseline and this run ID.
• Environment: Python 3.9.6, NumPy 2.0.2, PyTorch 2.8.0; macOS arm64 CPU, two threads. Fresh budget-transfer-proposal-0.1 seeds, no warm start, retry or extension.
• Fifty output hashes plus the manifest are preserved. The auditor verifies file sets/hashes, source/approval binding, regenerated data/seeds, B0/B3 reference predictions, selected learned-checkpoint predictions, losses, trace flags and the no-variance summary without optimization.
• Fifty-seven software tests pass. Baseline, research schema and generated-site checks remain required by CI.

Raw run and manifest · Frozen protocol

Next design work

The intended data sizes are feasible for this one local pilot. Before confirmation, review optimization rationale, variance assumptions, substantive margins and uncertainty calibration with a complete analysis implementation. Any additional training needs a separate fixed scope and authorization; this completed pilot grants none.