trialdesignbench ¶

TrialDesignBench provides tooling for evaluating whether AI agents can reproduce clinical trial designs from Statistical Analysis Plans and protocols.

This baseline implements workflow step 1:

1. Create a local benchmark workspace.
2. Convert a SAP/protocol PDF to Mathpix Markdown, with optional LaTeX ZIP output.
3. Build the standard TrialDesignBench reproduction prompt.
4. Run the prompt against a locally installed Codex SDK/runtime and save the run artifacts.

Installation¶

uv add trialdesignbench

For development:

git clone https://github.com/BBSW-org/TrialDesignBench.git
cd TrialDesignBench
uv sync

The experimental Codex Python SDK is declared as a Git source dependency for uv environments until it is published on PyPI. From a clone of this repository, uv sync installs both openai-codex and its pinned local runtime. For PyPI-only installs before openai-codex is published on PyPI, add the SDK source explicitly in the consuming project:

uv add "openai-codex @ git+https://github.com/openai/codex.git#subdirectory=sdk/python"

Quick Start¶

uv run tdb init tdb-workspace
uv run tdb configure --workspace tdb-workspace
uv run tdb run path/to/sap.pdf --workspace tdb-workspace --case-id tdb-001

Use --no-codex to exercise only the Mathpix ingestion portion:

uv run tdb run path/to/sap.pdf --workspace tdb-workspace --no-codex

The workspace .env file stores MATHPIX_APP_ID, MATHPIX_APP_KEY, CODEX_MODEL, and optionally CODEX_BIN. The default Codex model is gpt-5.5, and the default reasoning effort is high. The generated workspace .gitignore excludes credentials and output artifacts by default.