Evidence-gated agent operations
Evidence Harness
Check the project, permission and evidence behind an AI-assisted task. A portable skill supplies operating guidance; a small Python helper records whether the supplied facts allow the requested action.
The control loop
A prompt is not an operating boundary
Model choice and prompt quality do not resolve stale authority, wrong-environment action, unbounded mutation, or unsupported completion claims. The harness turns those failure modes into explicit checks.
Route → authority → sense → bound → verify → receipt → improve
Operating guidance
- Decides whether a task needs the harness at all.
- Loads only the rules relevant to that task.
- Establishes which source is authoritative.
- Senses actual state before consequential action.
- Preserves explicit human authority.
- Asks for reproducible evidence before completion.
What it does not do
- Guarantee correctness, security, safety, or compliance.
- Publish private prompts, product contracts, traces, or operational history.
- Grant permission to deploy, publish, message, spend, or delete.
- Replace project-local authority or host-environment controls.
Public proof
The expected outcome is a refusal
The included fixture asks for a deployment while the observed project differs from the authoritative project and mutation authorization is absent. Evidence Harness records both reasons and executes no infrastructure command.
decision: blocked
reasons:
- mismatch:project
- authorization:missingThe helper evaluates the supplied files and writes a receipt. It does not execute a deployment or enforce an agent's network permissions.
Source and installation
Inspect the artifact before you install it
GitHub is the release authority. The repository includes the skill, specification, receipt schema, synthetic fixture, four behavioral tests, public-hygiene scan, and complete release history.
Replay the example
Download and extract v0.1.0. From the extracted repository root, use Python 3—no model API key is needed:
python3 scripts/release_check.py
python3 skills/evidence-harness/scripts/harness_receipt.py \
--task examples/synthetic-infra-mismatch/task.json \
--authority examples/synthetic-infra-mismatch/authority.json \
--observed examples/synthetic-infra-mismatch/observed.json \
--output /tmp/evidence-harness-receipt.json