Frozen-model interpretability / study003¶

Outcome¶

New causal evidence is contextual necessity, donor non-interchangeability, and cancellation—not a strategy neuron or a named expert circuit. All predeclared negative and unfavorable results are retained. No model was trained, game changed, package built, or release modified.

Reference and scope¶

Verified all 69 snapshot evidence hashes and 13167 saved stage vectors across 627 canonical boards (including logits/routes/gates). Composed host full-set inverse/legal-mask/original-tie outcomes pass on 4520 raw boards, with zero set/move/tie mismatches. This is comparison to copied actual627 target evidence, not a new 6502 execution.

Packet SHA256 fe180f8a931f9d92d8aa39da337bc819d0dbe4e6d971edcba00324fbc18a91c1. Snapshot 027e1c11c6114312fee552cdf894b248fb3e5be0 is a materialized archive, not a Git working tree. All imports resolve within the isolated study root. Prior study002 and all 69 copied evidence files remain hash-identical.

The fixed experiment-manifest.json was written before any new intervention/held-out evaluation. Selection executed training rows only: all 16 single-zero lesions, then at most 15 contrast candidates excluding E1.h0; 8 upstream lanes; 3 predeclared pairs/groups and 4 donor components. No tuning or expanded search after confirmation. 128 interventions comprise 80 single-dose, 12 group-dose, 32 donor and 4 upstream/restoration runs; identity control is additional.

Folds group canonical symmetry classes: train372/validation121/test134 (seed1729). These are exploratory selection/confirmation folds only: the frozen model had already fitted/audited this finite domain, and study002 already probed it. Raw copies are not independent observations; no confidence interval or significance claim uses 4520 as independent N. Test-only fork support is six canonical boards; classes overlap. Minimax labels audit outputs offline; it is absent from neural inference.

Winning/must-block/fork labels describe availability with the stated priority exclusions, not an expert objective. Fork prevention is bounded: no immediate win/threat, opponent fork available, and candidate leaves no fork after each viable one-ply opponent reply. It is not a complete long-horizon prevention theorem.

Supported: dose-dependent effects, not exclusivity¶

A=expert_hidden[1,0] after GELU; B=expert_hidden[3,3], the training-selected contrast maximizing must-block minus fork move-error rate. Lesions scale only routed observations with the original integer round_div(h*n,4). Unrouted cells are null/unobserved, never activation zero. Doses and natural patches are post-routing and changed zero routes/gates. Upstream changes are reported separately below.

“Bad” includes empty legal sets; sets use legal logits >=0, then inverse-map the complete set before original-coordinate ties. A canonical chosen-move error need not expand uniformly because tie order is original-coordinate; raw outcomes are re-evaluated, not multiplied by eight.

Set/tie categories intentionally overlap: a changed set can retain an optimal chosen move, and a different chosen optimal move is not a failure. optimal_move_set_changed in machine results separates set changes retaining optimal behavior.

Target effects versus covered matched non-target damage¶

Matching is exact on ply, side, own/opponent corner/edge counts, center ownership and ordered expert routes, with max slot-gate difference <=4096/32768. It is not exact squarewise occupancy; that would identify the board and prevent donor contrasts. Rates below are canonical; non-target rates average matched negatives per covered positive, not pooled independent samples. All rows restrict to routed component observations. All-domain comparisons are descriptive, not held-out confirmation.

B shows a descriptive must-block excess versus covered negatives, but most of its whole-domain failures are empty-set damage and it also breaks immediate wins. Confirmation is weak: B must-block validation5/13 routed targets (matched1/2 versus 0% across2 negatives), test1/10 (no matched test negatives). B fork validation0/3 and test0/1 are too small to establish selectivity. A fork validation0/1 and test1/4 routed targets, with no matched target coverage in either fold, do not establish a fork-specific component. A whole-domain fork matched positives0/7 versus9.5% controls is unfavorable to that interpretation.

Concrete necessity example A: canonical170, board [0, 0, 0, 0, 0, 2, 2, 1, 1], h=25125; zero loses its winning fork. Counterexample canonical259, board [0, 1, 0, 1, 0, 2, 2, 2, 1], has a fork with routed A but remains optimal after zero. B block example canonical268 fails; non-block immediate-win example canonical23 also fails. Full cases and further target-unaffected/non-target-failure examples are in counterexamples.json.

Natural donor patches: limited coverage, no semantic transfer finding¶

Across all four components/tasks/class directions: 1044 pair records, 113 unique training donors, 116 unique held-out receivers. These totals contain reuse across tasks/components and are not independent samples. No relaxation was added when matching failed. Donors minimize canonical occupancy Hamming distance, then gate distance, then canonical index. Gate tolerance does not claim exact gating; receiver gates remain unchanged.

Both boards are fed in their own canonical input frame. Patches align the same expert ID and hidden-unit coordinate, not board squares and not route slot numbers. No hidden-unit rotation-equivariance is assumed; raw-to-canonical permutations are saved in raw-alignment.json. Donor indexing, receiver indexing and inactive rejection are independently tested.

E1.h0 fork same-class has zero positive receiver coverage, despite48 covered negative receivers; cross-class covers only2 positive test boards, with0 failures. Do not interpret this as necessity/rescue evidence. For B must-block, same-class causes2/3 positive failures and cross-class3/12; matched E3.h0 causes1/3 and2/12. The donor pools differ, so this is not a balanced class-direction effect. No win/block/fork/prevention feature transplantation is established.

A concrete positive result is contextual non-interchangeability: donor320 to receiver268, E3.h3 3483->123, same must-block class, same ply/side/geometry counts/routes, gate delta0, Hamming4, removes the receiver legal set. It is not universal: same-class donor491 to receiver145 changes B -277->579 yet retains an optimal chosen move (gate delta35, Hamming4); cross-class B must-block has9/12 positive receivers remaining optimal. Identity patches preserve every saved stage/logit/route on all627. Full donor IDs, activation values, class/fold/coverage and distances are in donor-pairs.json; experiment metrics retain matched non-target comparisons, empty sets, set and tie changes.

Pair interaction and path restoration: retain the negative results¶

These are canonical whole-domain counts. E1+E3 are co-routed on only14/627 canonical boards; most combined damage occurs on disjoint routed populations. Only one new failure beyond the single-failure union appears for A+B: canonical3 (test), a bounded-prevention board. The local A+E1.h1 pair instead rescues8 union failures and adds none, including test canonical553 where h0 alone retains the optimal canonical chosen move (but changes its complete set) while h1 alone fails; this cancellation is a counterexample to a simple additive necessity story. Small logit nonadditivity (max27) is compatible with downstream integer rounding/clamps; this experiment does not identify its arithmetic source. Sign/tie thresholding can create behavioral interactions without a special nonlinear circuit.

Training activation-only selection chose C=norm2_output[7] half. It causes 139/627 canonical and 1041/4520 raw bad moves, 106 ordered-route changes (63 expert-membership changes) and 530 gate-vector changes. This is an upstream router-affecting perturbation, not a pure expert lesion. Restorations retain that altered route/gate state; only baseline/upstream jointly observed experts are patched.

A restores2/50 covered upstream failures but introduces2 new failures: net zero canonical benefit. The rescued cases137/414 are both training examples; introduced296 is training and553 is test. Contrast B restores2 but introduces7; joint restoration on14 co-observed cases rescues0/2 upstream failures. This does not identify C->A->fork mediation. M+ board137 demonstrates a real local restoration; M- board553 is its counterexample, newly damaged by restoring A. “Restoration” here is a path intervention, not an identified natural indirect effect; shared upstream damage/rerouting confounds a circuit story.

Unsupported¶

• Dedicated fork, win, block or prevention neuron; exclusive expert strategy job titles; board-orientation-equivariant hidden semantics.

• Reliable same-class/cross-class strategy transfer; fork donor coverage was inadequate and matched controls can be equally or more damaging.

• A selective two-unit strategy circuit or a useful mediated rescue from C. Small per-fold classes and empty-set damage narrow the conclusions.

• Model-held-out/OOD generalization, independent symmetry samples, native-target performance changes or gameplay improvements.

Artifacts and reproduction¶

Figures: necessity-context.{png,svg} and patch-and-restoration.{png,svg}. figure-metadata.json binds A/B/C/P/M+/M- to exact activation/tensor coordinates and case IDs, source hashes and measured inputs. No metaphorical activations or stored-weight changes are shown. Logit differences are exact whole-model intervention effects, not asserted additive contributions.

Machine evidence: baseline-gate.json, fixed experiment-manifest.json, training-selection.json, cases.json, raw-alignment.json, results.json, experiment-vectors.json (full logits/routes/gates/norm2/observed hidden), effect-table.csv, donor-pairs.json, pair-interactions.json, path-restoration.json, identity-controls.json, counterexamples.json, provenance.json, hashes.json. results.json retains all units/doses/groups, class/fold raw+canonical denominators and covered matched-control rates/IDs; table is a flat view, not a second experiment.

Run from ARCHIVE_ROOT; use a fresh nonexistent output directory:

export OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 PYTHONDONTWRITEBYTECODE=1

export PYTHONPATH=ARCHIVE_ROOT

SOURCE_ROOT/.venv/bin/python scripts/model_strategy_study003.py --output artifacts/interpretability/study-003-reproduction-new

SOURCE_ROOT/.venv/bin/python scripts/render_strategy_study003.py --study artifacts/interpretability/study-003-reproduction-new

Validation and exact fresh-run comparisons are recorded separately in verification-study003.json at the study workspace root. Broader legacy integer tests requiring uncopied FP/training artifacts cannot pass in this bounded evidence snapshot; no packages or external data were fetched to conceal that limitation.