AI governance
Veridion uses a language model as a reviewer, not an authority. These are the rules it operates under and the ways it can still go wrong.
The model's job
In AI-assisted runs, the model reads a requirement, the result of each deterministic check, and a short list of candidate passages, then judges whether the wording actually establishes each element. It drafts the rationale shown on the finding.
What it cannot do
- It cannot change data, take actions or call tools. Its only output is a structured JSON answer.
- It cannot cite a passage it was not shown. Every citation is checked; unknown identifiers are discarded and recorded.
- It cannot decide numbers, units or periods. If it disagrees with an extracted value, the finding goes to human review instead.
- It cannot override a detected conflict between documents.
Which model, and when
The model is configurable per deployment. The default is openai/gpt-oss-120b served by Groq, called with temperature zero. A deployment can use another compatible provider, a model running on its own infrastructure, or no model at all — rules-only assessments are fully deterministic. Every run records the provider, model and prompt version used.
Known failure modes
| Risk | Safeguard |
|---|---|
| Invented citations | Citations are validated against the passages shown; invalid ones are discarded and counted. |
| Overconfident or wrong judgements | Deterministic checks decide figures; sharp disagreements and low-confidence answers are routed to human review. |
| Instructions hidden in documents | Passage text is presented as data, the prompt says so explicitly, and the model has no tools to act on such instructions. |
| Provider outage or error | The requirement falls back to the rules-only result and is flagged for review. |
| Behaviour changing between model versions | Runs record the model and prompt version; responses are cached per version; the evaluation suite can be re-run. |
Human oversight
Reviewers can accept a finding, override its status with a recorded reason, or comment. The original assessment is never overwritten. Model adjustments to rule checks are listed on each finding, so a reviewer can see exactly where the model changed the outcome.
Measurement
Rules-only and AI-assisted modes are measured against the same labelled dataset, and the results — including failure cases — are published on the methodology page.
Appropriate use
- Veridion supports evidence review. It does not provide legal advice or make compliance determinations.
- Findings describe the documents provided; conclusions about a company should be reached by a qualified person.
- It should not be used to make decisions about individuals.