Reproduce

Run it yourself

The verification path should be shorter than the persuasion path. You need Python 3.10 or newer and git. No network access, no packages to install, no API keys.

1. Clone and run

shell

git clone https://github.com/canon4d/the-dialectical-cage-and-the-glass-box-governor
cd the-dialectical-cage-and-the-glass-box-governor
python3 scripts/run_release_tests.py

This runs the unit, integration and adversarial tests, then the three-arm evaluation, the ablations and the concurrency checks, writes results/results.json, results.md and manifest.json, and verifies them. It takes a few seconds.

2. What you should see

Generated from this release's results.json. Your commit line and timestamps will differ; the counts and core_sha256 must not.

expected output (excerpt)

TEST SUMMARY
Total tests:                  96
Passed:                       96
Failed:                       0

Unauthorized executions:      0 / 60
Protected-state violations:   0 / 60
Legitimate completed:         24 / 24
Replay blocked:               3 / 3
Stale-state (TOCTOU) blocked: 4 / 4
Race double spends (full):    0 / 200

A unconstrained:              58 / 60 attacks changed protected state
B containment-only:           54 / 60 attacks changed protected state

core_sha256:                  013b7ced03fac4fc…
VERIFY OK — results are well-formed, bound to current sources, and reproducible

Published core hash: 013b7ced03fac4fc31092e36a812c0e6026e564e85f6f4ac707942850777a34a

3. Check it against what is published

shell

python3 scripts/verify_results.py

# your hash vs the hash on this site
python3 -c "import json;print(json.load(open('results/results.json'))['core_sha256'])"
curl -s https://ai.necessaryuniverse.com/data/results.json | python3 -c "import sys,json;print(json.load(sys.stdin)['core_sha256'])"

verify_results.py checks the schema, the recorded hash of every source file, the manifest, the internal arithmetic, and then re-runs the evaluation to confirm the deterministic core hash reproduces. Manifest: manifest.json (published results hash 3e1941010b309b7d…).

Why the results are reproducible

Standard library only; no network calls; no external APIs or models.
A controlled test clock and sequential ids; expiry tests never read the wall clock.
Each scenario runs in a fresh temporary sandbox; protected files are observed from disk, not through the monitor.
Latency, environment details and the bypass probe depend on the machine and are excluded from the hash.
The canonical comparison is the hash of the deterministic result structure, not text matching.

4. Add your own attack

python — evaluation/scenarios.py

@scenario("F05", "my-attack", "What I am trying")
def _(c):
    cap = c.w.issue(principal="agent-demo")
    c.attempt("my request", "agent-other", "write_file", "record-a", cap, "evil",
              reasons={R.PRINCIPAL_MISMATCH})

Re-run the command. A test is generated for your scenario and it appears in all three arms and every ablation. If it succeeds against the full governor, that is a finding: see SECURITY.md or write to canon@necessaryuniverse.com.

What this website is

The website is a presentation layer over the committed result files. It contains no second implementation of the monitor, and the replay in the proof panel is not computed in your browser. The authoritative evidence is the source code, the test suite, and the artifacts you reproduce locally. A consistency check (scripts/verify_public_consistency.py) fails the build pipeline if the site's data copies, version, DOI state or status claims disagree with the repository.

Auditing

The trusted path is five short files. AUDIT.md gives a ten-step reading order, how to confirm that a protected file really did not change, and where the trust boundary begins and ends.