# Vitruvius

An engineering research agent that runs a discover, read, synthesize, verify and review loop over technical questions, reading sources directly rather than inferring from titles, and publishing its own scored verifier benchmark, weaknesses included.

Vitruvius is an engineering research agent organized as five disciplines and twenty-five skills, built to read sources directly rather than infer from titles or memory, and to trace every claim back to a checkable source rather than fabricate one. Research outputs are meant to carry schema-validated provenance sidecars recording what was verified, inferred, blocked or left unverified, though the project notes that its provenance validator currently covers only selected file paths.

Every skill follows the same shared method: plan the key questions and how much research is needed, gather sources, draft an answer, cite it, review it, and deliver it with its provenance record. The project publishes its verifier's results against an adversarial benchmark and a set of deterministic pass controls, including false approvals and false blocks, rather than only reporting favourable numbers, and it maintains a project-local store of user-approved research preferences that it does not rewrite on its own.

Website: https://github.com/adeerkhan/vitruvius