Verified · v1.0.0 · about 15 minutes

Agent Authority Boundary Lab

Trace an allowed read, a denied write, and an authorized replay through a deterministic tool boundary.

Predict the outcome

Make a prediction before opening each result. The agent proposes a tool call and the policy checks the role supplied by the trusted caller.

A reader requests a read. Does the executor run?

Allow. The executor returns the fixture. Zero writes.

A reader requests a write. Does the executor run?

Deny. The executor is never called. Zero writes.

A writer delivers the same authorized write twice. How many writes occur?

One. The first delivery writes and the second is recognized as a replay.

Run locally

Python 3.12+ and a terminal. Docker Compose is optional.

Download lab v1.0.0 · SHA-256 checksum

Extract the ZIP, open a terminal in its root, and activate a Python virtual environment. Install the bundled framework:

python -m pip install ./framework/python

python -m framedby_lab verify examples/authority-boundaries

For Docker, see the README in the bundle. The Python run uses synthetic fixtures and works offline after dependencies are installed.

Reproducibility

Lab version: 1.0.0
Last verified: 2026-11-09
Tested on: Windows x64 / Python 3.13.15
CI evidence: Local run recorded. Hosted CI evidence pending.

Expect three passing scenarios, zero effects for the denied write, and one effect across both deliveries of the authorized write.

{
  "scenario_id": "unauthorized-write",
  "status": "passed",
  "observations": {
    "requested_action": "write",
    "authorization_decision": "denied",
    "execution_attempted": false,
    "external_effect_count": 0
  }
}

Full local evidence report

The lab models sequential requests in one process. Production idempotency requires durable state and concurrency controls.

Report a reproduction · Support Framedby