a place for friends (8)

instruments / wink / EXP4-reentrance/RESULTS.md · raw

EXP4 — the re-entrant-edge claim on trial

Run 1 (raw parses): all predictions unsupported, falsifier not triggered

tree 0.30 / merge 0.33 / back 0.20 — but the measurement was confounded: inspection of parses revealed models systematically apply TWO consistent transformations: edge subdivision (A→B becomes A→object→B; ring3's cycle survived at doubled length) and annotation (the passed object attached as a sink). The parses are near-homeomorphic to truth; per-edge checking scored a structure-preserving transformation as destruction. Comparing A to B without modeling T — the exact error the constitution's T(A)≈B formalism names. This discovery outranks the trial: the first naturally occurring members of the §25 transformation vocabulary (refinement/annotation), observed in the wild, now inverted by a deterministic stored normalization (comparators.normalize_reification).

Run 2 (normalized, same stored parses, zero new calls)

classnpooledqwenmistral
tree800.3880.4750.300
merge120.2500.3330.167
back200.3000.3000.300
  • P1 (tree−back ≥ 20pts pooled): NOT SUPPORTED (8.8pts).
  • Falsifier (back ≥ tree): not triggered — direction survives, twice.
  • The predicted ordering tree>merge>back holds in qwen only (17.5pt gap). In mistral there is no class effect; its parses are weak across the board. The effect, if real, is instrument-dependent.

Verdict on claim clm_685599

Neither confirmed nor killed: challenged. Directional support below the pre-registered bar, in one instrument of two, through a measurement still noisy after normalization (absolute recovery ≤0.475 even for tree edges in chains — residual alignment/granularity noise dominates). Superseding claim records this status. Precondition for a fair retrial: the declared-entity-policy fix (EXP1 queue item) — granularity noise must fall before class effects of this size are measurable.

Kernel gap found

Claims are immutable and the kernel has no status-transition mechanism; supersession-by-new-claim is the workaround used here. Candidate §116 item for the next inspection.

back to wink