The neural model remains the default dynamics engine. A verified patch handles only the changed causal slice; explicit version context prevents contradictory regimes from being silently merged.
First, earn a learned baseline
Phase 4A trained a relational action-conditioned predictor and measured one-step, recursive rollout and planning behavior. A scalar confidence heuristic failed to control risk on larger topologies, even when its predictions looked confident.
That failure mattered: output confidence was not epistemic uncertainty. The project stopped threshold tuning and introduced ensemble disagreement prospectively rather than laundering the miss.
A local, observable, repairable shift
The deployment benchmark kept public topology and action names fixed while one relation changed rate. Acquisition observations exposed the mismatch; held-out and stress trajectories remained evaluation-only.
The learned model was wrong on 92.28% of causally affected held-out transitions while retaining 96.48% accuracy on unaffected transitions. A two-law reference replacement restored exact symbolic behavior without changing immutable version one.
Repair versus incremental fine-tuning
Every strategy received the same post-shift observation prefix. Fine-tuning updated weights; the hybrid committed a verified local correction and otherwise delegated to the frozen learned model.
| Budget | Frozen affected | Fine-tuned | Hybrid | Mean corrections |
|---|---|---|---|---|
| 0 | 12.03% | 12.03% | 12.03% | 0.00 |
| 8 | 12.03% | 25.53% | 45.01% | 0.37 |
| 16 | 12.03% | 29.52% | 62.59% | 0.57 |
| 32 | 12.03% | 31.91% | 100.00% | 1.00 |
| 64 | 12.03% | 57.13% | 100.00% | 1.00 |
Why the first conjunction still failed
Applying the active post-shift predictor to old-regime inputs failed the registered retention comparison. The same observable state and action can require two contradictory transitions; an unconditioned deterministic model cannot represent both exactly.
The remedy was architectural, not numerical: an opaque regime/version identifier routes to immutable predictors, and missing context produces explicit abstention. A fresh 30-world replication passed every routing check.
A later topology-scaling result was invalidated when an audit found an oracle-conditioned acquisition probe and exact-input overlap. It was removed from the evidence chain and rebuilt on clean boundaries.
The architectural lesson
A local patch can restore a changed learned mapping immediately, but retention requires explicit context when regimes are observationally contradictory. Versioning is therefore part of the world-model representation, not merely deployment plumbing.
This phase still used idealized structured corrections. Language compilation and the complete end-to-end loop remained open.
Combined totals are descriptive unless the article explicitly says they were a preregistered pooled gate. Machine timings describe the measured local implementation. A failed conjunction remains failed even when several sub-results are positive.