Experimental tagged CCS proof validation — 2026-09-12

Two fresh native Triton7 STARKs passed in the ignored integration test genuine_tagged_ccs_proofs_bind_atom_pair_topology_under_same_matrices. This is test-only integration, not a production Joy/Zheng protocol or a claim that general nox execution is complete. No production files changed in this gate.

Exact gate and results

CARGO_BUILD_JOBS=2 RAYON_NUM_THREADS=4 TVM_LDE_TRACE=no_cache cargo test --release --locked -p trisha-rs --test tagged_ccs genuine_tagged_ccs_proofs_bind_atom_pair_topology_under_same_matrices -- --ignored --nocapture --test-threads=1

The verifier derives one relation for if(axis1, quote7, quote[8,9]) and an atom subject. Its checker assembly is byte-identical for both statements. Input0 selects atom7; input1 selects pair[8,9]; both bind exact native cost3. The test sorts and consistently merges public coordinates, verifies the constant-one pin, and leaves its enforcement to the native CCS checker. It binds all other input/output tag, payload, absent-child padding and cost coordinates.

Both proofs survive encode/decode and genuine native verification. Negative assertions reject the opposite valid statement, reversed pair payload, changed cost and mutated public claim. Crucially the test replaces the complete public-input envelope with the opposite valid branch's coordinates and verifies against that branch's checker. The program is the same and metadata now matches, so rejection is by STARK claim binding, not only envelope mismatch. Both public outputs of the native CCS checker remain its normal empty stream.

Test result: 1 passed, 0 failed, 0 ignored (explicit ignored-test invocation; 1 ordinary test filtered out). Test duration 25.99s. Whole invocation, including dependency compilation and monitor shutdown: 131.089s.

Resource guard and receipts

An owned new process group was sampled every1s. Guard limits: aggregate process group RSS16GiB and host free memory at least15% (memory_pressure -Q). On a violation only that group would be terminated and the failure preserved. The guard did not trigger. Maximum sampled group RSS was 3,479,126,016 bytes; minimum observed host free percentage was 41%. This is sampled RSS, not a kernel-accounted instantaneous peak. No proof parameter/security was reduced; TVM_LDE_TRACE=no_cache changes upstream memory/cache strategy only.

Receipts: /tmp/trisha-tagged-ccs-proof.log, /tmp/trisha-tagged-ccs-proof-memory.jsonl, /tmp/trisha-tagged-ccs-proof-receipt.json; monitor script /tmp/run-tagged-ccs-guard.py. Exit0, no remaining owned process; Trisha build slot released to the full all-target/all-feature checker.

Source identity

The following SHA256 values were captured before execution and checked unchanged afterward. Paths beginning zheng/ are relative to the sibling workspace root; others are relative to Trisha. This lists the test and relevant code/lock files, not an independently inventoried release archive or complete dependency closure.