Study007 — Bounded pathway preservation and local rescue¶
Frozen5195-byte packet SHA-256 fe180f8a931f9d92d8aa39da337bc819d0dbe4e6d971edcba00324fbc18a91c1. Host-only analysis. No deployed model/game changes.
Gates and interpretation¶
Exact replay:627 actual-target fixtures,13167 complete stage vectors,5643 logit scalars. Composed complete-set/legal/tie outcomes exact on4520 raw boards. This reuses saved target evidence, not new target execution/performance. Additional unselected experts evaluated offline only; baseline observation placeholders are not measured zero activity. Selected zero gate remains a real executed expert.
Scores encode >=0 optimal-set membership, not categorical preferences or probabilities. Wrapper canonicalization, legal masking, full-set inverse and raw tie order are fixed non-neural. All finite boards were previously trained/audited;372/121/134 folds are probe/selection confirmation, not novel-domain generalization. Interventions are OOD and preserve native integer rounding, clamps and accumulation. No strategy labels inferred from damage.
Declared circuit boundary¶
Candidate expert3 selected on training-only necessity contrast (must-block minus matched-neutral active error rates); control expert0 is the predeclared next cyclic ID, not post-hoc best control. Scope is a whole expert output vector after routing, with receiver residual/gates/head retained. It is NOT a full network circuit. Natural matches reuse study005's frozen legal one-mark relocations; preserve raw candidate identity/ply/side/geometry/fold, candidate active in both; gate Linf<=4096/32768. Route-fixed and naturally route-changing contrasts are reported separately. Both patch directions executed. Donor latent vectors are computed in their own canonical frames; no square-axis latent alignment is invented. Raw orbit outcomes use exact declared transforms. Unavailable donor experts are never substituted as observed zeros.
Candidate-only patch, shared-noncandidate natural corruption, all-shared patch and candidate restoration are compared with zero-noncandidate/zero-all/restoration controls. Identity and restoration algebra are exact. Rescue reports newly introduced failures too. Natural pathway transplantation is off-distribution; retaining one output path while zeroing another is a local preservation test only, not natural mediation or full-circuit sufficiency.
| Role/contract | Directed trials | Natural rescues | New damage | Zero rescues | Candidate preserved: empty | Candidate preserved: nonoptimal nonempty |
|---|---|---|---|---|---|---|
| candidate/route-changing | 110 | 36 | 0 | 14 | 4 | 26 |
| candidate/route-fixed | 112 | 17 | 0 | 18 | 0 | 10 |
| control/route-changing | 68 | 6 | 0 | 19 | 0 | 7 |
| control/route-fixed | 64 | 6 | 0 | 8 | 0 | 6 |
Matching coverage (no post-hoc relaxation; cap24/stratum)¶
| Role | Contract | Task | Fold | Prior eligible relocation inventory | Accepted pairs |
|---|---|---|---|---|---|
| candidate | route-fixed | win | 0 | 90 | 5 |
| candidate | route-fixed | win | 1 | 18 | 3 |
| candidate | route-fixed | win | 2 | 15 | 1 |
| candidate | route-fixed | threat | 0 | 145 | 14 |
| candidate | route-fixed | threat | 1 | 17 | 3 |
| candidate | route-fixed | threat | 2 | 28 | 1 |
| candidate | route-fixed | fork | 0 | 167 | 22 |
| candidate | route-fixed | fork | 1 | 28 | 4 |
| candidate | route-fixed | fork | 2 | 36 | 3 |
| candidate | route-changing | win | 0 | 90 | 7 |
| candidate | route-changing | win | 1 | 18 | 2 |
| candidate | route-changing | win | 2 | 15 | 1 |
| candidate | route-changing | threat | 0 | 145 | 17 |
| candidate | route-changing | threat | 1 | 17 | 4 |
| candidate | route-changing | threat | 2 | 28 | 4 |
| candidate | route-changing | fork | 0 | 167 | 15 |
| candidate | route-changing | fork | 1 | 28 | 5 |
| candidate | route-changing | fork | 2 | 36 | 0 |
| control | route-fixed | win | 0 | 90 | 3 |
| control | route-fixed | win | 1 | 18 | 0 |
| control | route-fixed | win | 2 | 15 | 1 |
| control | route-fixed | threat | 0 | 145 | 8 |
| control | route-fixed | threat | 1 | 17 | 2 |
| control | route-fixed | threat | 2 | 28 | 2 |
| control | route-fixed | fork | 0 | 167 | 11 |
| control | route-fixed | fork | 1 | 28 | 3 |
| control | route-fixed | fork | 2 | 36 | 2 |
| control | route-changing | win | 0 | 90 | 4 |
| control | route-changing | win | 1 | 18 | 2 |
| control | route-changing | win | 2 | 15 | 0 |
| control | route-changing | threat | 0 | 145 | 15 |
| control | route-changing | threat | 1 | 17 | 0 |
| control | route-changing | threat | 2 | 28 | 3 |
| control | route-changing | fork | 0 | 167 | 6 |
| control | route-changing | fork | 1 | 28 | 2 |
| control | route-changing | fork | 2 | 36 | 2 |
Full per-direction empty/nonoptimal/tie/full-set and raw-orbit outcomes, exact expert output vectors and restoration records are in outcomes.json. Counterexamples are retained in figure cases007 A-C; lack of support is unavailable evidence, not a null circuit effect.
Reproduction and evidence¶
Run OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=ARCHIVE_ROOT SOURCE_ROOT/.venv/bin/python research678_run.py 007 --out <fresh-directory>. This historical interpreter path supplies dependencies only. Local falken9/scripts imports must resolve inside the isolated snapshot. verification-study007.json records fresh-directory byte equality and preservation. Bundles exclude model binaries, ROMs, secrets, environments and caches; a hash-matching original packet and target fixtures are prerequisites. No external writes or notifications.