Built to be checked

Hallucination is a design failure, not a model defect.

It needs three conditions at once: the model is asked for an answer it has no evidence for; it has no permission to say "I don't know"; and nobody can check. Remove any one and hallucination mostly dies. Our systems remove all three.

"Not publicly disclosed" is a first-class answer, and "couldn't check" is never "none found".

Confident fabrication is the number-one failure mode of AI research tools; one invented contact destroys trust in an entire report. So we never fill a cell to look complete. A brief that says "not disclosed" eleven times and is right about everything else beats one that fills every cell and poisons your trust with a single invention.

The same honesty applies when a source fails: every external dependency has a named degrade policy, so when a data source is down the output says "this could not be verified", never an empty result that reads as "nothing exists". Silent degradation is a lie by omission.

The model extracts; the code decides.

The AI is only ever allowed to report what it saw. Every judgement is made by deterministic, versioned rules: same evidence in, same output out, every time. When a customer asks "why does it say this?", there is a rule to point at, not a vibe.

The AI is the sensor; the software is the brain.

We write software whose input happens to be gathered by AI. Sensors can be noisy, so every reading carries provenance, and the brain never acts on an unsourced reading. That is why the products behave like software - reproducible, versioned, debuggable - rather than like a chatbot with opinions.

The engine is blind to who is asking.

The moment a model knows the answer you are hoping for, it starts finding it. Motivated reasoning is not a human monopoly. So the engine is never told what outcome a scan is hoped to support. It is the AI version of a double-blind trial.

System of record first, then canonicalise, then search.

Facts with an authoritative source come from the system of record (the official registry), never from a search engine's impression. And every search runs on the company's resolved legal identity, because "who is this company officially?" changes the answer to every question asked afterwards. This is why our directors data audits at one hundred percent.

We attack our own products before trusting them.

When we added prompt-injection defences, we did not declare victory; we built a live decoy website loaded with hidden adversarial instructions and scanned it cold. The engine reported the attack payloads as facts about the page, and its output held shape. Defences are proven by attack, not assertion.

Price the outcome, not the token.

Everyone assumes the smaller, cheaper model is the economical choice. We measured cost per finished deliverable across four models: the flagship model was both the cheapest and the best, because per-token price is only one variable. Model choice is a measurement, not an assumption.

The same discipline reconciles estimated spend against the real bill; doing so once exposed a threefold overstatement in our own cost figures, which we corrected the same day.

Data protection is designed in at contract time.

Every personal fact in our briefs is already public: sourced from the official register or the person's own employer, with the source attached. Commercial data aggregators are banned by name in the engine's rules. Erasure was built before rollout, not bolted on after an audit. The same rule that stops invented contacts is also the GDPR posture: one design choice, two wins.

Every artifact is version-stamped.

Every deliverable carries the version of the logic that produced it, like a software build number. Any brief on any desk, weeks later, traces to the exact rules that generated it.