Things I am building because the manual version kept failing.

Veritas connects assessment findings to their source evidence and flags unsupported conclusions for human review. It reads the documents a vendor actually sends and returns findings tied to specific source text; where the evidence falls short, it raises an open question for a person instead of stating a claim as fact.

Evidence chain — how a finding is produced

1. 01Source ingestedDocument parsed and retained with version and date.
2. 02Passage citedA candidate finding must point to specific source text.
3. 03Citation checkedA deterministic check confirms the quoted text exists in the source. Failures are discarded, not smoothed over.
4. 04Supported, or an open questionInsufficient evidence produces a question for review rather than a conclusion.
5. 05Human decision recordedA reviewer accepts, edits, or rejects. The assessment is versioned and the trail preserved.

Core principles

• Human-in-the-loop verificationNo finding leaves the system without a reviewer’s decision attached.
• Deterministic citation checksQuotation matching is code, not a model’s self-assessment.
• Versioned assessmentsRe-assessment creates a new version; prior conclusions remain readable.
• Audit trailsWho changed what, when, and on the basis of which source.
• Regression testingA fixed evaluation set catches drift when prompts or models change.
• Evidence provenanceEvery claim traces back to a document, version, and location.
• Governed AI useModel use is scoped, documented, and reviewable — the same standard I would ask of a vendor.