CAIN-42 evidence library

NULLIUS verifiers, release 2026-10-27

Evidence bundle: 0 claims, 0 of 0 invariants held.

Last reviewed 2026-10-27

SHA-256 manifest

The bundle's own README

NULLIUS verifiers, release 2026-10-27#

Three clean-room verifiers. Each one is a single Python file with zero CAIN imports. It needs Python 3 and the cryptography package. It exits 0 only when the evidence proves the claim against the public keys you pin.

Check the files against SHA256SUMS before you run them.

What changed since the copies inside the 2026-10-12 bundles#

The earlier copies are left unchanged inside their bundles and auditor-kit releases, because those releases record their hashes. Use these files instead. A live audit on 2026-10-27 found the following.

• execution truth: replay could not be checked from the command line. The tool took one receipt, so the

published replayed fixture, which must be FAILED, came back VERIFIED. It now takes several receipts and keeps one record of receipt and authorization ids across them: python3 verify_execution_truth.py R1.json R2.json ANCHORS.json.

• execution truth: the enforcer could hide a sensor. A failed sensor leaves no trace in a receipt, and the

enforcer chooses which observations to include. An enforcer that dropped the one anchored sensor that saw an undeclared write got VERIFIED. Now, if any sensor you anchored did not report, the result is INCOMPLETE (ANCHORED_SENSOR_DID_NOT_REPORT).

• launchpad: fields were shown but never checked. Changing 410 of 1,168 single fields in the sample export

still gave VERIFIED. Examples: a step's visible action (for instance to delete_everything), its state and action_digest, the gate's signed receipt, each record's digest, the dispatch decision id, the order of the history, the export schema and the deployment id. Every one of these is now checked against the signed records. The changes still accepted are only these: one witness statement out of four (three still meet the threshold of 3), the producer's free-text note, an under-claimed quorum flag, a reordering of history sequence numbers that keeps them strictly increasing, and the producer's paging window, which the output now labels as unverified.

• mission journal: a crash on malformed input. A line that was valid JSON but not an object stopped the tool

with a traceback. It now gives a clean FAILED. Nothing was ever accepted.

The mission-journal verifier passed every other attack: 132 single-digit edits, truncation against the witnessed head, reordering, duplication, a swapped trust file, and the same attacks under python -O.

What VERIFIED does not mean#

These verifiers check the evidence against the keys you pin. They do not check the channel you got the keys from, whether a signer told the truth, or anything that happened outside the recorded run. The verifiers were written by the same project as the systems they check, and no third party has reproduced these results.