The sceptical questions, answered plainly

What does it cost?

A cent a page. You buy credit ($10 minimum, $5,000 maximum per purchase) and each analysed page spends one cent of it. There is no subscription, no platform fee, no free tier, and no expiry — credit you bought is yours until you spend it, and unused balance is refundable on request. When a document would cost more than you have left, we refuse the whole document up front and tell you the arithmetic — you are never billed for a partial answer.

Do you keep my documents?

No. Analysis records keep bounded excerpts (the concealed text we found and its immediate context), never the document. Rendered page images are never stored. GET /v1/account/export shows you everything we hold; DELETE /v1/account removes it — records deleted, usage reduced to units-and-timestamps as billing evidence, keys revoked. Backups age out within three months of a deletion. Certificates carry hashes, not content, which is why yours keep verifying even after you delete the account.

Is my document sent to an AI model?

No. There is no LLM anywhere in the service. Detection is structural — we parse the file format itself — plus OCR (open-source, run on our own machines) for scanned pages. Nothing leaves our infrastructure.

How do I know what you catch — and what you don’t?

GET /v1/output-surface lists every technique the running build detects, generated from the build itself so it cannot drift. Where a technique has a known blind spot, we say so in the finding (residuals) rather than over-claiming — integrators have built reviewer-facing banners on those declared residuals, and we treat narrowing one as a breaking change. Every finding also carries confidenceBasis.ordinaryDocumentRate: how often that technique fires on ordinary documents, measured on named corpora, so you can decide how to route on it.

What about false positives?

Measured, on real documents of the genre we sell into, and re-measured when a detector changes — the rate ships on the finding itself. When we tune a threshold we publish which corpus moved and by how much in the changelog (e.g. pdf.tiny-font moved from 21% to 3% when measured on 510 real commercial contracts instead of government forms).

Can’t I just do this myself?

You can start, and honestly the first weekend goes well: a DOCX is a zip of XML, and flagging vanished runs, white text, and tiny fonts is a few hundred lines. Three problems then eat the next six months, because we built this for our own pipeline first and they ate ours.

First, hidden is a comparison, not a property. White text is an attack on a white page and a design choice on a dark banner, a shaded table cell, or a themed background — and the answer lives in style inheritance, theme fills, and cell shading, resolved the way the document actually renders. Get it wrong one way and you miss the attack; the other way and you flag the header of every ordinary template.

Second, the false-positive rate is the product. Our first tiny-font check fired on 21% of one document collection and 3% of another — same detector, different genre — and a detector nobody has measured gets muted by its reviewers within a week, after which it catches nothing. Every rate we ship names the corpus and engine version it was measured on, and gets re-measured when the detector changes.

Third, the catalogue doesn’t end. Each format hides text its own way — pictures laid over text in decks, field codes in legacy .doc, PDFs whose text layer disagrees with their rendered pixels, scans that need OCR — and every carve-out you add for legitimate content is a surface an attacker can dress a payload in. Ours are tested by attack, and keeping that true is our full-time job; it probably isn’t yours.

And even a filter that works perfectly leaves you nothing to show for it later. The signed certificate — this document, this verdict, this engine version, verifiable offline years on — is the part a homegrown script never produces. That said: if your exposure is one format and the stakes are low, a hand-rolled check is a reasonable place to start, and the techniques page is a fair map of what to look for. At a cent a page, we think the build-vs-buy math speaks for itself.

Can I try it without paying?

Not without $10 — but trying is risk-free rather than free. There is no free tier: the $10 minimum buys 1,000 pages, the credit never expires, and unused balance is refundable on request, so an evaluation that doesn’t convince you costs you nothing but the time. Signup takes one curl and returns your production key immediately. If you want a hand-held evaluation of any size, email us.

Am I locked in?

The exit is built in: GET /v1/account/export hands you everything as JSON, certificates verify offline against our published JWKS (/.well-known/jwks.json) with our servers switched off entirely, and the same container we run is the artifact an on-prem deployment runs — there is one build, not a hosted fork.

What happens when my credit runs out mid-batch?

Each document is checked against your balance before anything is stored: a document you can afford completes, the first one you cannot is refused whole with a 402 naming exactly how many pages it needed and how many you have. Already-analysed documents stay available. If you want never to hit that wall, auto-refill is strictly opt-in — below X pages, charge $Y — with the consent sentence you agreed to stored verbatim, and one call turns it off.

What’s a certificate, and why would I care?

Every certified analysis is signed. Months later — in a dispute, an audit, a court exhibit — anyone can verify offline that this exact document produced this exact verdict on this exact engine version, without trusting us or even reaching us. certified: false never hides: it names the reason.

We’re tax-exempt — how do we buy credit without tax?

Email your exemption certificate (resale, nonprofit, or government) to [email protected] before purchasing, and we mark the account exempt so checkout stops adding tax. The certificate has to be on file before the purchase — an exemption claimed after the fact is our audit liability, not yours, so we can refund and re-run a purchase but we can’t retroactively exempt one.

Which formats?

DOCX, PDF (born-digital and scanned — OCR is built in), PPTX, XLSX, HTML, RTF, and legacy DOC. The techniques page lists what is looked for in each.