Study006 — Router usefulness: membership, order and gates¶
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.
Pair controls and findings¶
Same-board expert outputs remain fixed for all six declared pair memberships. Each pair is ordered by receiver router score (low-ID ties), then receives native selected-score renormalization, receiver actual slot gates or uniform gates. Reversing pair and gates jointly is a pure-order control; reversing only pair is a gate-assignment intervention. Natural donors match fold, ply/side and own/opponent corner-edge-center counts; unmatched receivers remain no-op and coverage is separately reported in donors.json. Gate donors with same ordered route separate gating from pair choice.
Actual gates:0/627 errors. Uniform actual pair:194/627; pair-only reversal:291/627. All six renormalized pairs exactly preserve full sets on19/627 boards; at least one pair on627/627. Best/worst are hindsight outcome bounds, not alternative trained policy choices. Training-selected fixed pair[2, 3], confirmation metrics below.
| Intervention | Empty | Nonoptimal nonempty | Full-set changes | Optimal tie changes | Raw chosen errors /4520 |
|---|---|---|---|---|---|
| actual | 0 | 0 | 0 | 0 | 0 |
| reverse-pair-and-gates | 0 | 0 | 0 | 0 | 0 |
| reverse-pair-only | 147 | 144 | 467 | 58 | 2154 |
| actual-uniform | 86 | 108 | 359 | 48 | 1420 |
| pair-01-renormalized | 89 | 126 | 352 | 36 | 1615 |
| pair-01-actual-slot | 113 | 126 | 370 | 31 | 1779 |
| pair-01-uniform | 54 | 180 | 499 | 57 | 1729 |
| pair-02-renormalized | 69 | 55 | 245 | 35 | 934 |
| pair-02-actual-slot | 65 | 65 | 265 | 35 | 984 |
| pair-02-uniform | 61 | 158 | 459 | 73 | 1616 |
| pair-03-renormalized | 48 | 90 | 253 | 30 | 1043 |
| pair-03-actual-slot | 47 | 95 | 270 | 35 | 1076 |
| pair-03-uniform | 65 | 156 | 414 | 51 | 1665 |
| pair-12-renormalized | 104 | 70 | 314 | 70 | 1262 |
| pair-12-actual-slot | 111 | 71 | 325 | 66 | 1335 |
| pair-12-uniform | 130 | 147 | 483 | 50 | 2035 |
| pair-13-renormalized | 90 | 130 | 321 | 22 | 1623 |
| pair-13-actual-slot | 80 | 134 | 322 | 32 | 1602 |
| pair-13-uniform | 119 | 166 | 486 | 57 | 2102 |
| pair-23-renormalized | 22 | 85 | 192 | 7 | 737 |
| pair-23-actual-slot | 20 | 91 | 195 | 10 | 760 |
| pair-23-uniform | 90 | 135 | 401 | 73 | 1563 |
| natural-alternative-pair-donor | 61 | 98 | 273 | 37 | 1223 |
| natural-alternative-pair-renormalized | 47 | 72 | 214 | 28 | 967 |
| natural-alternative-pair-actual-slot | 63 | 104 | 278 | 37 | 1289 |
| natural-same-pair-gate-donor | 3 | 13 | 38 | 6 | 127 |
| natural-same-pair-gate-renormalized | 0 | 0 | 0 | 0 | 0 |
| natural-same-pair-gate-actual-slot | 0 | 0 | 0 | 0 | 0 |
Matched tactical versus neutral confirmation (training-selected fixed pair)¶
Repeated donors/neutral boards are not independent. Zero matched support remains unavailable; matching is not relaxed.
| Task | Fold | Positive candidates | Matched pairs | Unique neutral | Positive errors | Neutral errors |
|---|---|---|---|---|---|---|
| has_win | 0 | 181 | 98 | 30 | 21 | 9 |
| has_win | 1 | 65 | 13 | 9 | 1 | 0 |
| has_win | 2 | 75 | 23 | 12 | 2 | 3 |
| must_block | 0 | 91 | 52 | 32 | 12 | 10 |
| must_block | 1 | 22 | 6 | 6 | 1 | 1 |
| must_block | 2 | 20 | 2 | 2 | 0 | 0 |
| has_fork | 0 | 15 | 4 | 2 | 1 | 0 |
| has_fork | 1 | 5 | 1 | 1 | 0 | 0 |
| has_fork | 2 | 6 | 4 | 3 | 0 | 0 |
The matched contrasts quantify selective task effects relative to broad damage, but do not establish a named detector or causal strategy circuit. Offline all-expert computation2508 vs1254 normal outputs is additional work.
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 006 --out <fresh-directory>. This historical interpreter path supplies dependencies only. Local falken9/scripts imports must resolve inside the isolated snapshot. verification-study006.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.