In public
Hard questions
The objections a professional raises about an AI-checked report, answered in writing — including the ones we cannot fully answer yet.
A public page with quotes and sources from my client's report is a leak. My GC will never allow it.
Agreed, which is why it does not work that way on any tier, including the free one. The receipt publishes hashes, counts, timestamps, a signature and a position in the chain. The claim text, the quotes and the URLs stay on your side, in the receipt file you hand to whomever you choose, whenever you choose. A reader can independently confirm what you checked, when, and that the record predates the report — without seeing a line of client material. Privacy is not a paid option; selling it for an extra fee would mean we had never handled confidential material.
HOLDS runs at 64–77 %. You want to sign and store for seven years a record where independent models argue with my professional conclusion. In discovery that is not my alibi, it is the other side's exhibit.
Correct, which is why HOLDS is off by default and never enters the sealed receipt as a verdict. We sell you two near-deterministic levels: the source exists, the source says this — facts that do not belong to you and on which you are obliged to be diligent. HOLDS concerns judgement, which is what you are paid for; creating a permanent adversarial record of it is a bad trade, and we will not pretend otherwise. It is enabled by hand, for private pre-flight, with a separate confirmation, and it is kept private.
Courts have rejected 'one AI checked another AI' as independent verification. So your receipt is worth nothing in court.
True, and it is written on our first screen. The receipt is not proof that a citation is good. It is proof that a documented check was performed on a date you can substantiate, and that the record could not have been produced afterwards. That is a different question, and it is asked not in court but at deposition, at renewal, and by a client who wants a refund.
I will just say I checked everything myself. Or attach screenshots.
Until this year that was enough. What changed is that your process became discoverable (CLF v. Shell, 18 May 2026) and your insurer now asks for a written verification protocol. Self-attestation is the one form of evidence that weakens exactly when it is challenged. A screenshot is made by you, undated in any provable sense, and editable in thirty seconds — which is what the person cross-examining you will point out. A receipt is hashed at fetch time, linked into a chain, signed with a key that can be yours, and anchored in a daily public root.
$199 a month for something I use four times a year is absurd. And I have no idea how many claims are in my report, so your price per claim tells me nothing.
That is why we sell the report, not the claim: $9 per report, no account — try it on a real document today. The subscription is not about volume but about continuity of the record: a gap in the chain is worse than no chain, because opposing counsel will ask why the checks stopped in March. $39 a month is the cost of your chain not breaking in the months you file nothing. Count it against your rate: $39 is five minutes of a billable hour.
You are a one-person studio. Where are my receipts in three years?
This is the one objection answered by architecture rather than by promise. The verifier is MIT-licensed, a single binary, works offline and is mirrored on a public code host. A receipt is a self-contained file of about 4 KB; you can export your whole chain at any time with one command. Daily Merkle roots are published to a public repository, not only to our database. If we vanish tomorrow, your receipt still verifies — and you can run that check right now, before you sign up. Our studio portfolio is not on this site: it answers the wrong question.
I could build this in an evening: fetch, 200, ask a model whether the quote matches.
In an evening you will build EXISTS — and you should, which is why EXISTS is free forever, no key required. In an evening you will not build canonicalisation that stays byte-stable across versions, key rotation that does not break historical verification, root publication to a place we cannot be quietly removed from, and — above all — a dated record that began before you needed the proof. A chain you start on the day you are asked is not evidence.