Either it shows you the line,
or it says it doesn't know

A tool that answers everything is indistinguishable from a tool that knows nothing, because it sounds the same either way. This one is built so that the sound of an answer is evidence: if it spoke, it found the words in your file, and it will show you where.

A confident wrong answer does not look like a mistake. It looks like an answer. It gets copied onto a bid, read out at a meeting or typed into a form, and nobody goes back to the file because the file is what they thought they were reading. The error travels under the name of your own records.

A refusal does the opposite. It sends you to the drawer. It costs you the twenty minutes the drawer would have cost anyway, and nothing else. On a job priced by the metre, the gap between those two outcomes is the gap between a number you can stand behind and a number that stands behind you. That is why this product treats a refusal as a result rather than a failure, and why the refusal is the state it spends the most design on.

Most tools in this category answer every question, cite nothing you can open, and never say what they could not read. They are easier to demonstrate and harder to trust, and the person who trusted one and got burned is the reader this site is written for.

Before a passage is allowed to become an answer it has to clear two floors.

Floor one · the score

The engine looks the question up three ways — by exact words, by weighted words, and by meaning — and fuses the three scores onto one scale. The best passage has to score above a fixed floor on that scale. A passage that is merely the least bad of a poor set does not pass.

Floor two · coverage

The passage also has to contain enough of the words in your question. A passage can score well on meaning and still be about something adjacent; this floor catches that. If you asked about the rate on a job and the passage never mentions the job, it does not pass.

The floors are set by hand and are not published as tuned values, because they are not tuned values. Where a corpus has a calibration set — a batch of questions with known answers and known non-answers — an optional third gate can be fitted on top of the two floors. Only one of the proving corpora has one so far.

If both floors pass, the passage is quoted: at most three lines, word for word, with at most two citations, each carrying the document id and the line locator. If either floor fails, the engine returns one sentence and nothing else:

Your records don't contain enough to answer that with confidence.

That sentence is fixed in the engine's source; it is never paraphrased, softened or varied. When you see it, you know a passage was found and rejected, and that nothing was composed in its place.

It would be easy to build a tool that refused everything. It would be safe and useless. What makes a refusal worth having is that the same engine, on the same corpus, answers when the words are there — and proves it by naming the document and the lines. The two states vouch for each other. An answer you can open makes the refusal credible; a refusal that really is a refusal makes the answer worth opening.

There is no language model inside the engine. The only model it loads turns your text into vectors for the meaning leg of the search, and it runs on your machine. Nothing in the answer path can compose a sentence, so nothing in the answer path can compose a wrong one. The wrong answers it can give are of one kind only: a real line from a real document, quoted when it should have declined. That is the failure the gate exists to stop, and it is the failure that has been measured.