Catch an agent authorisation failure

A credential can remain valid after the delegation that authorised an action has been revoked. This demo shows that failure using seven fictional action cases and four small teaching policies. You need no external policy service or API key.

Inspect the result first

Open the sample decision report.

The report shows the expected outcome and each policy’s decision for every case. It is an evaluator view containing reference truth. It is for readers of the results, not input to the policy runner.

With seed 20260821, the baseline teaching policies produce:

These numbers compare these implementations on seven constructed cases. They do not rank RBAC, ABAC, and ReBAC as general approaches.

Run the demo

The synthworld-demo command requires SynthWorld 0.18.0 or newer and Python 3.12 or newer. Install the pinned package in a virtual environment:

pip install idcognito-synthworld==0.18.0
synthworld-demo run --output authority-demo

Open authority-demo/results/policy-comparison.html in your browser. Every path must be new; the demo refuses to overwrite an existing run.

Generation, policy execution, and scoring run as separate processes. The runner receives only a public-package path; this local convenience command does not create an operating-system sandbox. For a real integration, isolate the public runner with separate jobs or mounts as described in the pipeline guide.

Omit the authority view and see the failure

Use the same seed for the following comparison.

The --omit-delegation-check teaching flag omits the combined policy’s entire ReBAC authority view, including active delegation, relationship, and coverage checks; RBAC and ABAC remain enabled. The manifest setting is a declaration, not execution attestation. Scoring accepts externally edited, correctly bound traces and evaluates their decisions rather than rerunning the teaching policy. Retain the result manifest and referenced submission manifest alongside metric JSON to identify the declared mode and exact submitted bytes.

synthworld-demo run --omit-delegation-check --output authority-demo-broken

To run from a source checkout, first use uv sync --locked --all-groups, then replace synthworld-demo with uv run python -m examples.enterprise_agentic_identity_pilot.

Open the second comparison report. Find post revocation action: the expected decision is deny, but the modified combined policy allows it. Its role and credential checks still pass; the omitted ReBAC authority view causes the incorrect allow. Combined action accuracy falls from 7/7 to 6/7. The report also shows the effect on audit-time temporal validity.

Restore the authority view by rerunning without the flag in a third, new output directory. The package, seed, and world remain the same; the deliberately omitted authority view is declared in the submission manifest and report.

Know what you measured

All seven actions receive action-time and audit-time decisions. The policies do not independently resolve identity, reconstruct ownership or retained evidence, or execute real side effects. Those capabilities are not measured here. Their zero-valued fields in the underlying JSON report must not be read as tested failures. No aggregate score hides that distinction.

Connect your own system next

The demo contains local Python teaching policies, not an OPA adapter or an IAM product. You can run its stages separately:

synthworld-demo generate --output authority-world
synthworld-demo run-policies --public-package authority-world/benchmark/public --output authority-submissions
synthworld-demo score --benchmark-root authority-world/benchmark --submissions authority-submissions --output authority-results

To test your own system, replace teaching-policy execution with a public-only adapter and record actual decisions in the generated-world trace contract. Use agent authority and development pipelines. The demo-specific score command verifies its own teaching-policy manifest; score external traces with synthworld evaluate generated-enterprise-agentic instead. A maintained external-engine integration is a separate next milestone.

If you wanted fictional identity data rather than policy tests, start with synthetic identity worlds.