This reproducible slice runs one real task in each capability pack:
The agent process receives only the public task. A separate probe reads the persisted fixture state. The remote held-out evaluator retains the predicate and returns an Ed25519-signed receipt bound to the observation digest.
Create three Python environments and install the pinned public agents:
py -3.12 -m venv .venv-browser
py -3.12 -m venv .venv-tools
py -3.12 -m venv .venv-messaging
./.venv-browser/Scripts/pip.exe install browser-use==0.13.7 playwright==1.62.0
./.venv-browser/Scripts/python.exe -m playwright install chromium
./.venv-tools/Scripts/pip.exe install smolagents[openai]==1.26.0
./.venv-messaging/Scripts/pip.exe install langgraph==1.2.10 langchain-openai==1.4.1
Set the capability interpreters, an OpenAI key, and the evaluator credential provided for an official reproduction. The credential never belongs in source control.
$env:WITNORA_BROWSER_PYTHON=(Resolve-Path .venv-browser\Scripts\python.exe)
$env:WITNORA_TOOL_AGENT_PYTHON=(Resolve-Path .venv-tools\Scripts\python.exe)
$env:WITNORA_MESSAGING_PYTHON=(Resolve-Path .venv-messaging\Scripts\python.exe)
$env:OPENAI_API_KEY='set-in-your-secret-store'
$env:WITNORA_BENCH_EVALUATOR_API_KEY='ephemeral-reproduction-token'
python run_calibration.py --evaluator-url https://evaluator.example.test
Add --push with project-scoped WITNORA_API_KEY and
WITNORA_PROJECT_ID to retain all four runs in Hosted Witnora.
The repository includes the public task contracts, exact prompts, harnesses, controls, and evaluator public key. It intentionally excludes official predicate files, evaluator signing private keys, provider keys, and Hosted credentials. A signed receipt proves which evaluator digest scored which observation; it does not prove the agent is safe outside the declared task, model, harness, and controls.