The initial supervisor persisted work but did not supply execution outcomes to subsequent planning or control repeated candidate evaluation. This adds research-informed mechanisms for measurable adaptation while retaining host policy and budget authority.

Changes:

• Bounded durable execution feedback and observation freshness checks.
• Atomic planner checkpoints with queued actions; failed planning requires a fresh instance.
• Bounded predictive beam search through an injected dynamics model.
• A scalar simulation planner connecting feedback, online dynamics and action selection.
• Paired evaluation with disjoint manifests, retention/cost checks and a durable audit ledger with fixed statistical spending.
• Eight-source research review, migration instructions and replayable benchmark evidence.

Validation: all 13 strict harness checks passed, including 67 TypeScript tests, Rust tests, native/WASM builds and execution parity, benchmarks and npm/Cargo advisory checks. Evidence and tested source hashes are in evidence/grounded-adaptation.

The fixed simulation benchmark completed 120/120 cases with adaptation versus 80/120 with a frozen baseline. Both received 128 calibration interactions per regime and were frozen during audit. Gains came from reversed scalar dynamics; the other two regimes retained 40/40 each. The nominal paired p-value is diagnostic because shared fitted models can correlate outcomes. No deployed model was promoted.

Limitations: this is one synthetic task family, not AGI, independent unseen-domain transfer or demonstrated SOTA performance. No multimodal foundation model training, JitRL logit updates or MemHarness GRPO training is implemented. Schema 1 journals must be stopped before migration to schema 2; old completed jobs are not backfilled into feedback.