Skip to content

0065: Preserve exact integers in the LLM witness

2026-10-23. A new candidate under the standing implementation authority and decision 0064. No change to Hallmark's classifier statements, Rust proof relation or verifier acceptance rules.

The first real TinyLlama layer failed before proof generation. The upstream Rust loader recomputed a different attention denominator on one of 1,024 rows. A local reproduction identifies head 1, position 16: the integer Q/K dot product is 17,866,521, but the Python witness records its maximum as 17,866,520. The negative shift is outside the exponent table. This is a witness error; the loader's rejection must remain.

The cause is int_gemm_exact's unconditional final .float(). Its float64 dot product is exact, but float32 cannot retain that odd integer. The bounded candidate removes only this cast, preserving the dtype selected from the partial-sum bound. All four callers were inspected: widened output projection, integer-softmax scores, integer attention/value multiplication and quantized Q/K attention. The regression uses int64 dot products and must fail on the original helper before passing on the candidate. No dependency is added to Hallmark's core environment.

Freeze the source commit, patch hash, model revision, calibration, prompt and expected answer in a new run directory. Regenerate inference and every witness segment. Keep numerator arrays for the upstream comparison, check all row maxima and denominators against exact integer products, and retain the failed original. A passing model-root build is only the next gate; the full response still needs actual proofs and caller-pinned mutation rejection.

The candidate may change intermediate values or output text. New measurements must name the patch and must not inherit the original witness's evidence claims. The reviewed eight-token answer is an engineering fixture, not a held-out quality test. Existing classifier experiments and qualification flags remain unchanged.