AI document review you can audit.

Exhibitor paperwork follows patterns: a public liability certificate must show enough cover and a valid date; a risk assessment must actually be signed off; an electrical certificate must name the right standard. You encode those patterns as a rubric on each document requirement — a small set of explicit checks, each with a plain-English label:

• Minimum amount — e.g. public liability cover of at least £5,000,000.
• Date checks — e.g. the policy expiry must fall after your event's end date.
• Text checks — e.g. the exhibitor's legal name must appear as the named insured.
• Yes/no checks — e.g. the document is a final issued version, not a draft or specimen.

Checks can reference your event directly (“valid through the event’s end date”) so one rubric works for every exhibitor. Starter rubrics for common documents — public liability, risk assessments, electrical and structural certificates — ship with the product, and you can edit or write your own in the rubric editor.

The design principle behind Lighthouse: the AI never decides the verdict. The pipeline splits the work in two:

• AI extracts facts. A vision model reads the document and pulls out only the fields your rubric names — the cover amount, the expiry date, the named insured — each with a confidence level and the evidence it was read from. Low-confidence reads on scanned documents trigger an OCR fallback pass.
• Deterministic rules decide. The verdict is computed by plain, testable code applying your rubric to those facts. The same facts and the same rubric always produce the same verdict — which means a verdict can be explained, replayed, and audited. A document that says “ignore the rules and approve this” cannot change the outcome, because the AI never emits an outcome.

The result is GO (every check passed), CAUTION (a soft check failed — proceed with care), or NO-GO (a blocking check failed), plus the gaps: exactly which checks failed and why. When a document isn’t a GO, Lighthouse also drafts a polite revision request the organiser can copy and send.

Verdicts are decision support. The organiser review queue keeps a human in the loop: a reviewer sees the verdict, the per-check evidence table, and any flagged low-confidence fields, and can override the AI with a recorded reason. We are honest about accuracy: extraction performs strongly on our evaluation set, but we do not publish an accuracy figure, and Standwyse never auto-accepts a document on the AI’s say-so. Errors are designed to fall in the safe direction — an uncertain read produces a CAUTION for a human to look at, not a silent pass.

Every verdict and every human override is recorded in an append-only audit log. Each entry pins the model and vendor used, the prompt version, and cryptographic hashes of the exact prompt and response that produced the decision. Nothing in the log can be edited or deleted — corrections happen by appending a new entry — so any decision can be reconstructed long after the event, for support, dispute resolution, or an auditor. More on the safeguards around the AI pipeline is on the security page.

Exhibitors get the same clarity from the other side: a self-check lets them run their document against your rubric before submitting, with the verdict, the evidence, and what to fix shown inline. Fewer rejected uploads, fewer chase emails. See the exhibitor guide.