The closed repair loop

One common interaction budget

Sixty fresh worlds covered two independently gated change classes: action rate and repair delay. Frozen, incremental fine-tuning, symbolic repair and hybrid composition consumed the same visible transition prefixes at budgets 0, 8, 16, 32 and 64.

The hybrid could additionally spend counted expert answers. Evaluation measured one-step accuracy, H5 and H8 recursive rollouts, bounded planning, false plans, update time, invariant violations, version retention and leak boundaries.

Recovery across the budget

The result was not a last-checkpoint win. Hybrid-minus-fine-tuning confidence intervals stayed strictly positive for H5 recovery AUC in both shift families, while the longer H8 rollout improved and planning remained non-inferior.

Final multi-step competence

At budget 64, the hybrid preserved unaffected one-step competence while materially improving recursive rollout and planning. No strategy violated a planning invariant, but candidate plans were still checked against the reference environment.

What the resource comparison means

Local verified repair used 69–75% less measured update CPU than fine-tuning on the tiny 1,468-parameter model. That ratio was implementation evidence, not a production claim. The more important comparison was the resource vector: equal trajectories, about one answer, far more rollout recoveries.

The scale question—whether the cost advantage grows with learned-model capacity—was explicitly left to later phases.

The milestone

Phase 6A turned the project from a rule-graph prototype into an evidenced systems thesis. The asset is the loop: gap detection, targeted teaching, typed compilation, verification, hybrid execution, provenance and versioning.

It remained a controlled synthetic result. Industrial generalization, million-parameter competence and uncertainty-calibrated planning were still open.