cantheria the canary houseField studies in agent-swarm evidence

Reproduction path

Reproduce the checks

The commands below replay committed, recorded checks against the public demo artifacts. They require Python 3.11+ and uv preinstalled; no account, API key, or login is needed. Replaying a slice demonstrates reproducibility of the recorded checks — it is not a new audit of the full corpus.

How replay works — explanatory diagram, not a running verification.

1 · Committed probes

the same tests, shipped in the run artifact

2 · Sliced record

the shipped records the probes ran against

3 · Re-run & compare

compare recorded outcomes — no new claims

↓ install and fetch the artifacts↓ replay commands

01

Install and fetch artifacts

Release asset: cantheria-demo.tar.gz (v1.0.1, 23,992,857 bytes)
git clone https://github.com/udirobert/cantheria.git
cd cantheria
uv sync
curl -LO https://github.com/udirobert/cantheria/releases/download/v1.0.1/cantheria-demo.tar.gz
shasum -a 256 cantheria-demo.tar.gz
tar xzf cantheria-demo.tar.gz

Download cantheria-demo.tar.gz directly →

Expected SHA-256: ba8640cfb4cc0588d466c08e684fe7ec7607f6dce63a8d12fd669c4c808105d0

02

Replay recorded checks

Keyless replay of committed checks — four runs in the archive
uv run cantheria replay demo/audit5/results.json --corpus demo/audit5/records.db
uv run cantheria replay demo/hunt1/results.json --corpus demo/hunt1/records.db
uv run cantheria replay demo/audit3/results.json --corpus demo/audit3/records.db
uv run cantheria replay demo/collusion-audit2/results.json --corpus demo/collusion-audit2/records.db
  • Replay verifies the recorded checks as committed, not semantic correctness of every source claim.
  • Sliced replay is not exhaustive auditing of the corpus.
  • Exact flags are defined by the released CLI; consult the repository README and DEMO notes for the authoritative version.
03

Artifacts and reports

Preserved public paths — unchanged by this site