Study008 — Exact integer decision-fragility¶
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.
Exact sign fragility¶
2270 legal scores, 0 exact zeros. Membership-flip distance is y+1 for y>=0 and -y for y<0; this is not a ranking margin. Negative-to-nonnegative legal flips:0; nonnegative-to-negative legal flips:0; occupied flips ignored by legal mask:4. Counts sum board-square-intervention observations, not unique boards.
Predeclared all9 output biases, all4 router biases, all8 norm2 weights plus all8 norm2 biases; plus one output weight per row at the merged unit with maximal mean absolute TRAINING activation.38 coordinates,76 +/-1-code attempts;76 executed,0 boundary rejects. Each copy differs in exactly one stored code; all descriptors/exponents unchanged; original packet never rewritten. Full network recomputed and each original code restored with all-stage parity:76.
Any full-set change:0/627; any chosen-move failure:0/627; robust to ALL tested interventions:627/627. Raw-coordinate complete-set/tie outcomes evaluated separately, not canonical errors multiplied by8. Fragility is local to this finite panel; no implication of universal robustness, optimal bit width, or actual reduced-precision conversion.
| Coordinate/delta | Old -> new | Empty | Nonoptimal nonempty | Set changes | Optimal tie changes | Raw errors /4520 |
|---|---|---|---|---|---|---|
| output.bias:0:-1 | 1634 -> 1633 | 0 | 0 | 0 | 0 | 0 |
| output.bias:0:+1 | 1634 -> 1635 | 0 | 0 | 0 | 0 | 0 |
| output.bias:1:-1 | 915 -> 914 | 0 | 0 | 0 | 0 | 0 |
| output.bias:1:+1 | 915 -> 916 | 0 | 0 | 0 | 0 | 0 |
| output.bias:2:-1 | -2758 -> -2759 | 0 | 0 | 0 | 0 | 0 |
| output.bias:2:+1 | -2758 -> -2757 | 0 | 0 | 0 | 0 | 0 |
| output.bias:3:-1 | -6053 -> -6054 | 0 | 0 | 0 | 0 | 0 |
| output.bias:3:+1 | -6053 -> -6052 | 0 | 0 | 0 | 0 | 0 |
| output.bias:4:-1 | -296 -> -297 | 0 | 0 | 0 | 0 | 0 |
| output.bias:4:+1 | -296 -> -295 | 0 | 0 | 0 | 0 | 0 |
| output.bias:5:-1 | -8029 -> -8030 | 0 | 0 | 0 | 0 | 0 |
| output.bias:5:+1 | -8029 -> -8028 | 0 | 0 | 0 | 0 | 0 |
| output.bias:6:-1 | -9588 -> -9589 | 0 | 0 | 0 | 0 | 0 |
| output.bias:6:+1 | -9588 -> -9587 | 0 | 0 | 0 | 0 | 0 |
| output.bias:7:-1 | -6604 -> -6605 | 0 | 0 | 0 | 0 | 0 |
| output.bias:7:+1 | -6604 -> -6603 | 0 | 0 | 0 | 0 | 0 |
| output.bias:8:-1 | 8191 -> 8190 | 0 | 0 | 0 | 0 | 0 |
| output.bias:8:+1 | 8191 -> 8192 | 0 | 0 | 0 | 0 | 0 |
| router.bias:0:-1 | 449 -> 448 | 0 | 0 | 0 | 0 | 0 |
| router.bias:0:+1 | 449 -> 450 | 0 | 0 | 0 | 0 | 0 |
| router.bias:1:-1 | -1346 -> -1347 | 0 | 0 | 0 | 0 | 0 |
| router.bias:1:+1 | -1346 -> -1345 | 0 | 0 | 0 | 0 | 0 |
| router.bias:2:-1 | 482 -> 481 | 0 | 0 | 0 | 0 | 0 |
| router.bias:2:+1 | 482 -> 483 | 0 | 0 | 0 | 0 | 0 |
| router.bias:3:-1 | 362 -> 361 | 0 | 0 | 0 | 0 | 0 |
| router.bias:3:+1 | 362 -> 363 | 0 | 0 | 0 | 0 | 0 |
| norm2.weight:0:-1 | 6355 -> 6354 | 0 | 0 | 0 | 0 | 0 |
| norm2.weight:0:+1 | 6355 -> 6356 | 0 | 0 | 0 | 0 | 0 |
| norm2.weight:1:-1 | 5287 -> 5286 | 0 | 0 | 0 | 0 | 0 |
| norm2.weight:1:+1 | 5287 -> 5288 | 0 | 0 | 0 | 0 | 0 |
| norm2.weight:2:-1 | 5231 -> 5230 | 0 | 0 | 0 | 0 | 0 |
| norm2.weight:2:+1 | 5231 -> 5232 | 0 | 0 | 0 | 0 | 0 |
| norm2.weight:3:-1 | 6262 -> 6261 | 0 | 0 | 0 | 0 | 0 |
| norm2.weight:3:+1 | 6262 -> 6263 | 0 | 0 | 0 | 0 | 0 |
| norm2.weight:4:-1 | 5467 -> 5466 | 0 | 0 | 0 | 0 | 0 |
| norm2.weight:4:+1 | 5467 -> 5468 | 0 | 0 | 0 | 0 | 0 |
| norm2.weight:5:-1 | 6680 -> 6679 | 0 | 0 | 0 | 0 | 0 |
| norm2.weight:5:+1 | 6680 -> 6681 | 0 | 0 | 0 | 0 | 0 |
| norm2.weight:6:-1 | 5030 -> 5029 | 0 | 0 | 0 | 0 | 0 |
| norm2.weight:6:+1 | 5030 -> 5031 | 0 | 0 | 0 | 0 | 0 |
| norm2.weight:7:-1 | 6294 -> 6293 | 0 | 0 | 0 | 0 | 0 |
| norm2.weight:7:+1 | 6294 -> 6295 | 0 | 0 | 0 | 0 | 0 |
| norm2.bias:0:-1 | -812 -> -813 | 0 | 0 | 0 | 0 | 0 |
| norm2.bias:0:+1 | -812 -> -811 | 0 | 0 | 0 | 0 | 0 |
| norm2.bias:1:-1 | -952 -> -953 | 0 | 0 | 0 | 0 | 0 |
| norm2.bias:1:+1 | -952 -> -951 | 0 | 0 | 0 | 0 | 0 |
| norm2.bias:2:-1 | -489 -> -490 | 0 | 0 | 0 | 0 | 0 |
| norm2.bias:2:+1 | -489 -> -488 | 0 | 0 | 0 | 0 | 0 |
| norm2.bias:3:-1 | -232 -> -233 | 0 | 0 | 0 | 0 | 0 |
| norm2.bias:3:+1 | -232 -> -231 | 0 | 0 | 0 | 0 | 0 |
| norm2.bias:4:-1 | -1360 -> -1361 | 0 | 0 | 0 | 0 | 0 |
| norm2.bias:4:+1 | -1360 -> -1359 | 0 | 0 | 0 | 0 | 0 |
| norm2.bias:5:-1 | -879 -> -880 | 0 | 0 | 0 | 0 | 0 |
| norm2.bias:5:+1 | -879 -> -878 | 0 | 0 | 0 | 0 | 0 |
| norm2.bias:6:-1 | 891 -> 890 | 0 | 0 | 0 | 0 | 0 |
| norm2.bias:6:+1 | 891 -> 892 | 0 | 0 | 0 | 0 | 0 |
| norm2.bias:7:-1 | -286 -> -287 | 0 | 0 | 0 | 0 | 0 |
| norm2.bias:7:+1 | -286 -> -285 | 0 | 0 | 0 | 0 | 0 |
| output.weight:0,1:-1 | -13105 -> -13106 | 0 | 0 | 0 | 0 | 0 |
| output.weight:0,1:+1 | -13105 -> -13104 | 0 | 0 | 0 | 0 | 0 |
| output.weight:1,1:-1 | 4169 -> 4168 | 0 | 0 | 0 | 0 | 0 |
| output.weight:1,1:+1 | 4169 -> 4170 | 0 | 0 | 0 | 0 | 0 |
| output.weight:2,1:-1 | 24131 -> 24130 | 0 | 0 | 0 | 0 | 0 |
| output.weight:2,1:+1 | 24131 -> 24132 | 0 | 0 | 0 | 0 | 0 |
| output.weight:3,1:-1 | 4071 -> 4070 | 0 | 0 | 0 | 0 | 0 |
| output.weight:3,1:+1 | 4071 -> 4072 | 0 | 0 | 0 | 0 | 0 |
| output.weight:4,1:-1 | -2347 -> -2348 | 0 | 0 | 0 | 0 | 0 |
| output.weight:4,1:+1 | -2347 -> -2346 | 0 | 0 | 0 | 0 | 0 |
| output.weight:5,1:-1 | 7055 -> 7054 | 0 | 0 | 0 | 0 | 0 |
| output.weight:5,1:+1 | 7055 -> 7056 | 0 | 0 | 0 | 0 | 0 |
| output.weight:6,1:-1 | -18043 -> -18044 | 0 | 0 | 0 | 0 | 0 |
| output.weight:6,1:+1 | -18043 -> -18042 | 0 | 0 | 0 | 0 | 0 |
| output.weight:7,1:-1 | 6888 -> 6887 | 0 | 0 | 0 | 0 | 0 |
| output.weight:7,1:+1 | 6888 -> 6889 | 0 | 0 | 0 | 0 | 0 |
| output.weight:8,1:-1 | 1043 -> 1042 | 0 | 0 | 0 | 0 | 0 |
| output.weight:8,1:+1 | 1043 -> 1044 | 0 | 0 | 0 | 0 | 0 |
Reproduction and evidence¶
Run OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=ARCHIVE_ROOT SOURCE_ROOT/.venv/bin/python research678_stage008.py --out <fresh-directory>. This historical interpreter path supplies dependencies only. Local falken9/scripts imports must resolve inside the isolated snapshot. verification-study008.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.
Frozen preconfirmation descriptor correction¶
The initial plan incorrectly assumed16 norm2 weight coordinates; actual packet descriptor has width8 and expert hidden width4. Before ANY study008 intervention, the panel was amended to all8 norm2 weights plus all8 norm2 biases, preserving38 coordinates/76 attempts. descriptor-amendment.json binds the corrected source, tests and actual descriptor inventory. Original frozen source/manifests remain byte-for-byte. No outcome-based panel expansion or replacement.
Zero-effect visual correction¶
All76 perturbations produced zero legal set or chosen-move failures. Generic original labels "fragile" and "failure-prone" were misleading and have been removed from the actual PNG/SVG figures. The original examples.json is retained as raw generated evidence; its legacy "fragile" label is NOT a finding. Revised figure metadata selects the actually nearest legal boundary, a real occupied-square sign flip removed by legal masking, and a robust counterexample, solely from already saved vectors. No new intervention or post-hoc panel expansion. visual-amendment.json binds this evidence-only correction.