WHAT WE REMOVED from our own website, and why.

Every false thing our website ever said had the same shape: plausible copy, written to fill a space, indistinguishable from a true claim to everyone except the one reader who checks. Not lies, exactly. Placeholders that were never replaced, sentences that borrowed the vocabulary of a field to sound at home in it, numbers that a reasonable person would assume had a source.

A company selling systems whose whole promise is every claim carries its source cannot have a website like that. So in August 2026 we read every sentence on ours the way a sceptical buyer would, and removed what failed. This is the list.

An implied security attestation. The site read as though it held a SOC 2 report. It does not, and never did; a SOC 2 report can only be issued by an accounting firm under a specific standard. The policy pages now say plainly what we hold — nothing — rather than letting a phrase imply otherwise.

A benchmark we had not run. The VELA case study said the system was evaluated against a named industry benchmark. Our own repository's limitations section says the benchmark harness is not yet wired. The sentence now says what is true: the benchmark identified the gap; VELA has not been run against it; no score exists.

Engineering that did not exist. The "how we build" page described automated security scanning, model-graded evaluation suites and versioned benchmark datasets. The build pipeline runs six checks — documentation, types, lint, formatting, tests, build — and none of those three is among them. The page listed the six real checks correctly, two sentences away from the invented ones. The true list and the false one sat side by side, and the false one read as more credible, because it used the industry's words.

Diagrams with their prompts still in them. Four generated architecture schematics, each with a spelling error in a label, one with a node for the scanner that did not run, and captions that were the image generator's own prompt text rendered as if it were a title. Generated imagery is fine; generated type is not, and a diagram is mostly type. Every label on the site is now real text.

A progress bar that measured nothing. The opening screen counted from zero to a hundred per cent on a random timer, under captions about loading system modules. It now shows the name and three true lines for the same twelve hundred milliseconds.

Words. "India's first". "Enterprise-grade". "Universal". "Industry standard". Each was a superlative with no assessor behind it, and each is gone.

Our site has an assistant that answers from a corpus derived from the site itself. On its first adversarial test, a reader asked it for a percentage improvement between two approaches described on a research page. The model — correctly grounded, with the source table in front of it, explicitly instructed not to state numbers outside its sources — computed one anyway: a ratio of two simulated values that appeared in no document.

The grounding did not catch it. The instructions did not catch it. A third layer did: a check that inspects the finished answer, finds a claim with no source, and discards the whole answer in favour of an honest refusal. That layer exists because the first two had already failed under real conditions. It is the same principle as the product: a claim without a source does not ship, and the enforcement has to be a gate, not a request.

Before a sentence goes on the site, find the mechanism it describes. If the answer is a document, a repository, a pipeline step or a live URL, the sentence is fine. If the answer is a plot we would have to generate, a rig we do not have, or a paraphrase we would have to defend, the sentence is deleted.

Under-claiming is always correct. A missing fact costs a reader nothing. An invented one costs the whole site its credibility, and we know that because we paid it once.

Where a gap is deliberate — a number we could show but choose not to, a name we withhold pending consent, a benchmark we have not run — the site says the gap is deliberate, so that the next person to edit it does not helpfully fill it. Those notes are load-bearing. They are the immune system.