In development
Petra
Petra is an AI harness for science, built to make model-generated claims traceable to their evidence. I’m developing a workbench that connects hypotheses to supporting and conflicting source passages, with people reviewing changes before they enter the trusted record.
From a claim to its source
Language models extract claims and synthesize findings. A graph lets researchers inspect the passages supporting or challenging a hypothesis, keeping disagreements and missing evidence visible.
Review before acceptance
A policy-controlled gateway routes bounded model tasks. People review proposed changes before they enter PostgreSQL, the trusted record. Neo4j provides a rebuildable graph projection with explicit relationship and traversal limits.
What the demonstration shows
The miniature investigation on the homepage is an authored walkthrough using sample research data. It distinguishes a source-linked finding from an unsupported overclaim and leaves the test of factual accuracy open. It makes no model or database calls; it is not a measurement of Petra’s reliability.
