How we check our work

AI can be confidently wrong. That’s a strange foundation for a product whose entire job is telling you the truth about a legal document, so we engineered around it. This page explains exactly what gets checked before a report reaches you, and what we deliberately refuse to do.

Every quote is located in your document before you see it

When your report cites a clause, it quotes the policy verbatim, and our server checks that the quoted text actually exists in the PDF you uploaded. The check is mechanical: we extract your document’s text and search for each quote. A quote we cannot locate is removed from the report rather than shown to you. We would rather give you a clause reference to look up yourself than a quotation we can’t stand behind.

The score is arithmetic, not a feeling

The policy score starts at 100 and loses a fixed number of points per issue found: 15 per critical issue (capped at 45) and 6 per moderate issue (capped at 24). Informational notes don’t deduct anything. Our server recomputes this from the issue counts on every report and overwrites anything that doesn’t add up, so the number on your gauge is reproducible math. The report itself shows the working.

We refuse documents we can’t verify

Scanned, image-only PDFs are rejected before analysis, not because the AI can’t read them, but because wecouldn’t then verify its quotes against extractable text. The same applies to wrong documents: upload a car policy to the health decoder and you get a refusal, not a made-up report.

Claim-settlement ratios come from the regulator or not at all

Insurer claim statistics in our reports come only from a hand-verified table sourced from the IRDAI Annual Report. If an insurer isn’t in that table yet, your report shows no ratio rather than a guessed one. The AI is forbidden from “remembering” these numbers, because remembered numbers go stale.

Before launch, every policy type passes a validation gauntlet

We run each insurance type against real, publicly published policy wordings from major insurers, with hand-built answer keys listing the traps each policy actually contains. Machines check the quotes, the math, the format, and the refusals; a human adjudicates every flag. A policy type goes live only after clearing all gates, including zero confirmed fabricated clauses and zero missed traps on keyed policies.

What we don’t claim

• Severity ratings are judgement.Whether a clause is “critical” or “moderate” is our considered opinion, clearly labelled, and you may weigh it differently.
• Money-at-risk scenarios are illustrations.They follow a stated formula using your policy’s own numbers, but your actual claim will differ. The report shows the assumptions.
• This is not financial advice.It’s a plain-English decode of what your document says, built so you can verify every factual claim yourself.

If we still get a fact wrong

Then the accuracy guarantee applies: a demonstrable factual error in a paid report means a full refund. The verification above exists to make that guarantee cheap for us to keep.

Your document, briefly held

Your PDF exists on our servers only long enough to produce your report, and is deleted after the full report is generated, with a daily sweep as backstop. Details in the privacy policy.