The paper's prompts were modified to work with structured outputs, and the author states outright that this is not the official implementation. What it gives you that a summariser does not is decontextualisation: each claim stands on its own, which is the property that makes checking it — by the bundled prompt or by a person — possible at all.
A local Python server implementing the Claimify method from the paper "Towards Effective Extraction and Evaluation of Factual Claims" — a four-stage pipeline that turns text into atomic factual claims.
- Extraction in four stages: split into sentences with their surrounding context, keep only the sentences carrying verifiable propositions, resolve ambiguity or discard what cannot be clarified, then decompose into atomic self-contained claims
- Opinions and speculation filtered out at the selection stage instead of being extracted as facts
- OpenAI structured outputs throughout, so results come back typed rather than as prose to parse
- Each extraction stored as MCP resources — an aggregate at `claim://extraction_<n>_<timestamp>` and one resource per claim at `claim://<slug>`
- A `verify_claim` prompt that has the model check a single claim against authoritative sources and answer VERIFIED, UNCERTAIN or DISPUTED, with at least three referenced sources when verified
- The original language preserved, and logging of every model call and pipeline stage
Python 3.10+ and an OpenAI API key in `OPENAI_API_KEY`, with a model that supports structured outputs — `gpt-4o` or the cheaper `gpt-4o-mini`. Create a virtualenv, `pip install -r requirements.txt`; the NLTK punkt data downloads on first run. Point your client at the virtualenv's Python and `claimify_server.py`. `LOG_LLM_CALLS`, `LOG_OUTPUT` and `LOG_FILE` control the logging.