conformance — first bless, 2026-10-10

The first stack snapshot of cyber-conformance 0.2 (contract: conformance/specs). It was generated from the origin default branches. Their revisions are recorded in conformance/snapshots/provenance.toml: hemera d5f0a09, strata 494df7a, nox 0b1407a, lens 392abe7, zheng 5733dfd, tade c7d404a, bbg 441e3c9, foculus 8a935e1, cybergraph 42fc251, inf 5054c93, tru 2f3a255, tok e27da0d.

result

• 24 encodings, 127 mechanisms, manifest root h64353c453d8db5101b7dfe6d084da4ea72bfc2e5869ccca47598ec29f6d05586.
• --check ran twice in a row and both runs reproduced the files byte for byte. The zheng prover is deterministic.
• The release-train gate (train_gates.conformance_gate) ran on a fresh target directory (aarch64-apple-darwin): green in 15.4 s, no warnings. The Linux targets were not run locally. The first Friday candidate gives the cross-platform receipt.
• Harness self-tests: 10 integration tests and 11 unit tests passed. They detect a tampered fingerprint, an edited encoding, a removed line, hand-edited bytes, a hand-edited manifest root, a tampered proof fixture and a stale fixture. They also check that drift at beta tier is only reported, that --bless restores the files and prints the diff, and that a delta-tier entry is refused unless --tier delta is given.

findings

1. The tree vectors published in hemera are the fixed-chunk tree. hemera/vectors/hemera.json tree.4k_zeros (66a2b0f7…) equals tree::fixed_chunk_root. It does not equal tree::root_hash (70dd43dc…). Since the content-defined chunking change, root_hash matches the fixed-chunk tree only below about 2 KiB. hemera's rs/tests/vectors.rs tests just empty and hello, so the divergence never failed there. The harness checks the tree section against fixed_chunk_root and snapshots both trees. Follow-up for hemera: label the section, or add a cdc_tree section.
2. zheng's default branch has no ZHENGPF1 profiles. Profile 0 (public), profile 3 (state-public) and the other profiles are on the 0.4 integration line (zheng#33 and the release/0.4 line). The zheng surface covers what master exposes:
   • commit/verify of trace proofs: three quote-only programs and the hash-binding proof [15 [1 42]];
   • the Brakedown PCS open/verify_eval. Each profile becomes a scenario pair when it lands on master. That change moves zheng's entries and requires a bless in the same PR.
3. All 45 single-bit tamperings of the five committed proof fixtures are rejected. The tamper points are k/8 of the length for k = 0..7, plus the last byte. For the trace proofs, 8 of 9 tamperings fail in the verifier and the last-byte flip fails in postcard decoding. For hash-42 and pcs-eval, all 9 fail in the verifier. The snapshot pins these verdicts.
4. The nox pattern-spec test vectors hold on master. These are add (3, cost 3), quote (7, cost 1), cons ([1 2], cost 3), branch (200, cost 5), inv(2), inv(0) = InvZero, and (p−1)² = 1. They are now fixture invariants.