Proof Engine

Proof Engine is an open-source AI skill for making factual claims come with supporting evidence. It is built around the idea that claims should be proved rather than merely asserted, and it is intended to make gaps in verification visible when a citation cannot be matched or a computation does not hold.

Its workflow combines several concrete steps. The system decomposes a claim into sub-claims and extractable facts, fetches sources from live URLs, and checks quoted text against the page. Numbers are computed in Python, and deterministic computation is part of the proof process. If a fabricated citation or partial match is encountered, the verdict is downgraded rather than hidden. The output can include structured verdicts such as PROVED, DISPROVED, PARTIAL, SUPPORTED, or UNDETERMINED, and the page says each verdict ships with a Jupyter notebook, a PROV-JSON provenance chain, and an RO-Crate 1.1 archive bundle.

The product is shown as a re-runnable Python proof script, and the page says the pipeline can be re-run from source offline in Python. It also states that the model is not part of the verification path, while still being used to draft code, find sources, and formalize claims. The page names academic papers, government data, and reference encyclopedias as source types, and says every URL is logged with HTTP status, fetch mode, and credibility tier. An audit trail is part of the system, along with coverage information for quote matching.

Proof Engine is described as an open-source skill and is available under the MIT license. The page mentions install and submit, and points to a catalog of 140 proofs across 20 domains with 393 sources checked. It is presented as a tool for handling factual assertions, including myths, statistics, mathematical identities, and pitch-deck numbers.