PENROSE
Penrose/Documentation
Back to site

Quickstart

Install

git clone https://github.com/PattersonResearch/Penrose
cd Penrose
python -m venv .venv
. .venv/bin/activate
pip install -e .

Docker is required only for claim paths that execute model-written strategy code. Deterministic claim executors and the keyless verification suite do not require an API key.

Verify the referee

Run the verification controls before trusting a result:

python scripts/eval_suite.py
python scripts/calibration_placebo.py
python scripts/calibration_power_mining.py 40

The eval suite must pass. The placebo must certify no noise signal, and the power-mining control must report PASS.

Configure a model

Paper ingestion and reconstruction use one OpenAI-compatible model endpoint:

export PENROSE_LLM_API_KEY=...
export PENROSE_LLM_BASE_URL=https://your-provider.example/v1
export PENROSE_LLM_DEFAULT_MODEL=glm-5.2

An independent verifier is optional and uses the corresponding PENROSE_LLM_VERIFIER_* variables.

Submit a structured claim

Structured claims avoid lossy prose re-extraction:

penrose run --claims claims.json --json

Or ingest a source:

penrose run --paper path/to/paper.pdf --json

Read Submit claims before using either route.

Inspect the result

penrose verdicts -n 20
penrose triage --json
penrose data-requests

Start the dashboard:

make dash

Then open http://127.0.0.1:8077/. For process controls, see Dashboard operations.

Stop at P9

No agent or unattended process may approve a survivor. Human review is explicit:

python -m penrose.pipeline.p9_review list

Do not run approve unless you are the authorized human reviewer and have inspected the complete receipt, report, provenance, and confirmation state.

Penrose is a falsification referee for quantitative trading claims. A verdict is evidence about survival under the referee, never a trading instruction.