SASH witness precision regression¶
The real TinyLlama attempt found that upstream int_gemm_exact computes a dot
product in float64 when required, then unconditionally casts it to float32.
Odd integers above 2²⁴ are lost. One recorded attention score was 17,866,521,
while its witness maximum was 17,866,520. The unchanged Rust loader correctly
rejected the resulting negative shift and denominator mismatch.
exact-gemm.patch removes that final cast. Apply it only to SASH commit
eb95bfa887299d1637550228d42086c27cbb2b3b in a separate checkout. It changes the
witness candidate, so regenerate its inference and witness; do not repair old
arrays or reuse their proof claims. It changes no Rust proof or verifier code.
With SASH's locked Python environment installed, from the Hallmark checkout:
git -C "$SASH_SOURCE" apply --check "$PWD/research/probes/sash/exact-gemm.patch"
git -C "$SASH_SOURCE" apply "$PWD/research/probes/sash/exact-gemm.patch"
PYTHONPATH="$SASH_SOURCE" "$SASH_PYTHON" research/probes/sash/check_exact_gemm.py
The four cases compare the real helper against an int64 oracle: a small dot product, positive and negative results above 2²⁴, and cancellation. The original helper fails; the patched helper passes. This regression alone is not a full model proof. See the candidate decision and current evidence.
The patch derives from Apache-2.0 SASH source; its licence and notice are retained here. The checker is Hallmark code under MIT.
Shape reconstruction regression¶
The larger first-layer run serialized a proof but rejected it during verification:
the parameter-only constructor supplied placeholder scales outside the statement's
fold divisor bounds. fold_shape.rs exercises that constructor at alpha 9 and
three boundary/control values. shape-reconstruction.patch changes its dummy
scales; it leaves the proof constraints and acceptance checks intact. This is a
separate Rust candidate, described in decision 0066.
With the pinned upstream build environment, from the Hallmark checkout:
cp research/probes/sash/fold_shape.rs "$SASH_SOURCE/prover/tests/fold_shape.rs"
cargo test --manifest-path "$SASH_SOURCE/Cargo.toml" --release --locked --test fold_shape
# The original constructor fails the alpha-9 case.
git -C "$SASH_SOURCE" apply "$PWD/research/probes/sash/shape-reconstruction.patch"
cargo test --manifest-path "$SASH_SOURCE/Cargo.toml" --release --locked --test fold_shape
The regression alone does not establish proof validity. The execution record must also show the saved proof verifying against its previously enrolled roots and changed claims rejecting. The patch retains the upstream Apache-2.0 notices; the new regression is Hallmark code under MIT.
The delayed-value profile also leaves an activation-scale equality unchecked; see decision 0067. This constructor patch does not repair that relation. The full-response candidate uses the existing non-delayed profile and the original Rust binary instead.
Complete-response caller checks¶
check_response.py is the caller guard used by the TinyLlama proof-only experiment.
It requires the caller's expected prompt/answer IDs and 23 enrolled model,
preprocessed and shape roots. Its parameter digest reproduces the pinned upstream
shape-key encoding, including calibration float bit patterns. It rejects incomplete
file coverage, changed parameters, invalid token IDs, the open-scale Fold profile
and an output head that omits final normalization. These checks remain active
under python -O.
The guard is an importable research module, not a cryptographic verifier or an
admitted Hallmark backend. After bundle(proof_directory, caller_expectations)
succeeds, the real zk-prover boundary-grid --verify must check all 23 proofs,
22 layer links, exact prompt and enrolled model/preprocessed/calibration roots.
The fresh-worker protocol and response evidence
record the complete sequence. Tokenization remains the caller's responsibility.
See the LLM proof explanation for the statement and
measured limits. Unit tests exercise this guard; they do not generate STARK proofs.