Generated but Not Committed: Locating a Commit-Stage Failure behind a Salient Geometric Misbinding in an Open-Weights LLM

Benchmark-passing behavior in large language models can be soft: an answer that holds at a problem's canonical presentation may be produced by the pull of a salient pattern rather than by the construction the problem demands. We report a fully pre-registered instrumentation campaign on one sharp case. Asked for the single most important invariant of a classical plane-geometry configuration, an open-weights reasoning model (Qwen3-32B) reliably commits to a salient but incorrect answer — the circle of Apollonius on the two similitude centers — in place of the correct apex invariant: the concyclicity of the four common-tangent intersection points on the circle over the diameter joining the two given centers, a construction demonstrably within the model's reach. Our contribution is not that candidate answers coexist, established in prior work, but the identification of a format-side manipulandum — closure, the requirement to commit to a single answer — with a joint behavioral, representational, and probabilistic measurement of where the failure lives. Ablating closure alone surfaces the apex invariant in six of ten candidate lists where the blocked (closure-forcing) prompt yields the attractor deterministically on the 4-bit greedy path; a Jacobian-lens reading of the model's reasoning rollouts shows a rally without a verdict: candidate constructions surface and alternate in the verbalizable readout, commitment to the attractor is never established there, and the scoring falls outside the lens's field of view, while the rare correct escapes are visible and effortful; and teacher-forced measurement finds the commit balance deeply tilted (median -12.4 nats) and unmoved by closure erasure. The failure is therefore a commit-stage selection failure, not a generation failure. The same differential-probing method transfers to a non-geometric domain, and we release the complete pre-registration ledger — twenty-one frozen predictions, including our own drafting errors — as part of the record.

Revision v2.0 (REVISION-ID: TSD-20260826). The token-set rule pre-registered on 2026-07-16 ("a word that splits into subwords is represented by its first token") is a defect; every token set was rebuilt under a repaired acceptance (joshaku, 7 modules, combined md5 89b194be…, 20 tests; SPEC v1.2) and every affected measurement was re-acquired. The readout-side readings marked in the text (P6, F6, P-G2) are superseded as described in the Revision History section; the behavioral findings are unaffected. Limitations (one line): all rank statements are pool-internal (N = 69,940) under band 26..57; the band statistic reads the representation before commitment. Version 1.0 (doi:10.5281/zenodo.22088920) remains available; its superseded values are not counted as evidence.

 

Provenance note (2026-09-24): the Qwen3-32B lens file this work read (n = 80 checkpoint, md5 ede6ade7…) was replaced upstream on 2026-09-21 by a finished 557-prompt lens (neuronpedia/jacobian-lens@496cf110); the readings here remain readings of the checkpoint as distributed, retained at doi:10.5281/zenodo.21449255.