How verification works
VeraFrame is built on one principle: the thing that generates content must not be the thing that verifies it.
Large language models produce fluent text. They do not produce proofs. When the same model is asked to check its own output, it often confirms whatever it said in the first place. Chaining two LLMs to review each other fails the same way.
VeraFrame splits the workflow into two clearly separated stages: AI generates, a deterministic validator verifies.
The two stages
Stage 1 — AI generates structured output
Given a request and source material, VeraFrame asks the language model to produce an output together with a structured trace: for every factual claim (a number, a date, a term, a name, a quoted clause), the model records which part of the source material the claim came from.
This happens inside a constrained context: VeraFrame narrows the source material to only the blocks relevant to each placeholder or claim, so the model is not guessing from the whole corpus.
Stage 2 — Deterministic validator verifies each claim
The validator is plain Python, not AI. For every claim in the output, it re-opens the referenced source material and checks whether the claim matches.
- Numbers are parsed and compared (with unit awareness).
- Dates are parsed and compared.
- Quoted clauses are matched against source text.
- Completeness is checked against the request schema or placeholder list.
- Conflicts between multiple sources are surfaced rather than silently resolved.
The validator has no opinion and no creativity. It reports what it sees. That determinism is what makes the result auditable.
The four validation modes
VeraFrame exposes the same verification pipeline through four modes, each tuned to a different scenario:
| Mode | What you give it | What you get back |
|---|---|---|
| Generate | Request + source material (+ optional Word/Excel template) | New document with trust report, optionally rendered into your template |
| Chatbot Audit | Existing chatbot response + source material | Verification report on the response you already have |
| Ask | Question + source material | Validated answer with citations |
| JSON API | Input payload + schemas + source material | Schema-validated JSON output for programmatic use |
All four modes share the same trust report format.
The correction loop
Validation produces a report, not a verdict. For any claim the validator flags, you can:
- Approve the claim (it is fine as written, even if the validator could not find a direct source).
- Override a mismatch (you know the source was wrong or incomplete).
- Edit the output text and rerun validation.
Corrections are recorded alongside the output. In regulated deployments the correction history becomes part of the audit trail.
What VeraFrame does not do
VeraFrame does not train or fine-tune models, does not replace your AI assistant, and does not decide what you should publish. Its job is to tell you, with evidence, what is true and what is not in a given AI output so that a human can decide what to do.