A repair mechanism becomes an architectural asset only if it can govern more than one favorable benchmark. On a fresh prospective confirmation, the same immutable Core handled a second learned-dynamics family and a second typed correction without domain logic entering the Core.
The next risk was architectural, not numerical
The embodied Reacher result established a strong bounded repair outcome, but a single integration could still be a benchmark-specific wrapper. The next question was therefore not whether another score could be improved. It was whether the repair contract itself could survive a different model interface, state geometry and correction type.
The only eligible second family was an existing sealed relational dynamics system with 10,068,868 executed parameters. Selecting from already completed evidence avoided benchmark shopping. The Core manifest was frozen before any fresh world opened.
Detection disposition, counted questions, proposal and commit authority, immutable versions, routing, rollback, trace validation and the complete Core source manifest.
A different model and a different local correction
Reacher repaired a hidden actuator-authority change with an action-input transformation. The second family used routed relational dynamics after a hidden four-relation regime change. Its adapter translated only public host evidence into a typed local parameter overlay.
No resource-process rule entered the Core. The host remained responsible for prediction, training and task metrics; IWMR remained responsible for whether a correction could be asked, verified, committed, routed and rolled back.
Fresh confirmation favored repair on prediction targets
Across four fresh changed worlds, every IWMR route used one bounded answer, produced one exact verified child version and retained the nominal route. Full fine-tuning, a rank-four adapter and the frozen model received the same post-change evidence budget.
IWMR reached all four preregistered H5 competence targets. Full fine-tuning reached none and the adapter reached two. The local patch also retained the strongest mean H5 and H8 affected-state accuracy in this confirmation.
| Strategy | H5 targets | Affected one-step | Affected H5 | Affected H8 | Planning success |
|---|---|---|---|---|---|
| Frozen | 0/4 | 0.1716 | 0.1378 | 0.2619 | 0.500 |
| Full fine-tuning | 0/4 | 0.1447 | 0.0882 | 0.2440 | 1.000 |
| Rank-four adapter | 2/4 | 0.5167 | 0.4941 | 0.4167 | 1.000 |
| Verified IWMR repair | 4/4 | 0.9063 | 0.7181 | 0.7679 | 0.875 |
Full fine-tuning and the adapter both achieved higher final planning success than IWMR on these four worlds. The claim is predictive target recovery and Core portability—not universal baseline dominance.
The governance behavior transferred too
All four changed worlds became actionable and consumed exactly one answer. Every proposal verified, every commit produced a scoped immutable version, every trace contained the required 13 events and every route rolled back to the nominal parent.
Two fresh no-shift controls produced no gap and spent zero questions. Two worlds with missing provenance were marked untrusted before the question, also spending zero. Incorrect commits, invariant violations and evidence leaks were zero.
| Condition | Worlds | Disposition | Questions | Commits |
|---|---|---|---|---|
| Consequential hidden change | 4/4 | Actionable | 4 | 4 exact |
| No-shift control | 2/2 | No gap | 0 | 0 |
| Missing provenance | 2/2 | Untrusted | 0 | 0 |
Locality was real; the fastest baseline stayed faster
The typed overlay represented 0.00276% of learned parameter bytes. Median update CPU was 0.451 seconds for IWMR versus 2.095 seconds for full fine-tuning: 4.64 times lower.
The closed-form adapter took 0.0197 seconds and was 22.9 times faster than IWMR. That comparator remains part of the result. Inference overhead measured slightly negative and is treated as timing noise, supporting non-inferiority rather than a speedup claim.
What this establishes—and what it does not
Together with the sealed Reacher integration, this result establishes functional portability across two learned dynamics families and two typed patch families while the Core source manifest remains byte-identical.
It does not establish universal model compatibility, open-ended language induction, task-level dominance, physical transfer or large-sample generalization. The new confirmation contains four synthetic instrumented changed worlds and exact structured corrections.
The next credibility step is independent handover of this second adapter and its deterministic evidence replay. The package is ready, but external reproduction remains an open gate until a reviewer runs it without implementation help.
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.