The Hypothesis Graph: a mechanism case study
收藏资源简介:
Artifact backing the paper 'The Hypothesis Graph: Semantic Memory Written by Methodeutics'. A mechanism case study on whether an LLM generalizes a bug fix to a represented predicate or merely tabulates the reported case. On a Verus verifier soundness bug (#2219, post-cutoff) whose correct fix is an XOR-shaped predicate, it isolates where prompt-encoded methods, an execution-gated evaluation tool, and cross-model capability each do or do not lift the fix, and shows that enumeration is inducible by the model while the calibration oracle (external ground truth) is not. Includes the forced-fresh identity-verified grading harness, the case arms (codex, Fable, Sonnet 4.6, Composer 2.5), the out-of-grammar held-outs, the full reasoning trail, the surrounding pilots, and reproduction instructions. An existence proof of a mechanism, not a resolve-rate claim.



