The AI that cites everything it says

Every sentence brings its fact.

Not a promise about style: a check that runs over each answer before it reaches the screen.

The citation is mandatory

Each fact is numbered and each sentence has to point at one. Measured before it was built: 7 of 7 sentences tagged and 0 invalid citations over five runs. A sentence with no valid citation drops itself; it is not rewritten and not softened.

7 / 7

Numbers are verified against the facts

Every figure has to appear in what it was given, to half a unit in the last decimal it wrote. If it does not, the answer is retracted and the model is told exactly what it invented.

The scope is checked too

A whole-lap number quoted inside a sentence about one corner is misattributed even when the number is right. Checked per sentence, corners only and decimals only, because a bare “2” matches far too many things.

And it checks which way it points.

A correct number pointing the wrong way is worse than no number, because it arrives as coaching.

The direction gate

If the measurement says “you make 3 MORE steering corrections”, a sentence saying “fewer” is the opposite of what was measured. The rule is positive and scoped: among the sentences citing the finding's own figure, at least one must carry the direction the measurement asserts.

0 / 45

Measured, not assumed

Zero false positives over a 45-entry corpus. That is the number that made it shippable, not the share it catches. A check that rejects correct coaching is worse than no check.

It exists because it happened

The local model wrote “hai 3 meno correzioni” in Italian: exactly the opposite of the measurement, handed to a driver as an instruction. All three flags in the corpus are that same case.

It runs on your machine.

A cloud provider is optional and explicit, not the default path.

Local by default

It talks to a model running on your machine. The telemetry does not leave. Ordering the prompt constant-first took evaluation from 5.98 s to a measured 0.64 s, because every runtime reuses the prefix that does not change.

0.64 s

Your own key, if you want one

You can use Claude, OpenAI, Gemini or OpenRouter with your own key. Then the block of facts does travel, with your permission and knowing what is in it. The key is stored only in your browser and is never synced.

The driver's name is never sent

Not ever, not even with your own key. It was removed from three automatic routes and each has a test asserting its ABSENCE from the composed prompt, which is what actually leaves the machine. A school should not be sending its students' names to a third party.

0 bytes

And it goes quiet when it measured nothing.

An assistant that always has something to say is one you cannot believe when it says something that matters.

Silence is a state, not an error

If nothing on that view was measured, it says so. When the sentence was composed by the analysis rather than by a model, it says that too. Knowing who is speaking is part of being able to check it.

It does not do the arithmetic

Even given both numbers, a subtraction performed by the model is refused. A deliberate trade: accepting a merely derivable number is the gap that lets misattribution back in. If a view needs a difference stated, the analysis computes it.

44 real generations kept

An append-only corpus of real outputs judged by the shipped verifier. It caught an invented “100 m” nobody was looking for. The risk no rule-level test covers is an answer that passes every gate and still is not worth showing.

44

What it still cannot do

The gates bound the numbers and their direction, not what they are called. If the model mislabels a correct quantity, that is undetected - and it is written here rather than glossed over.

Watch it talk about a run of your own.

Send us a file and we will send back the analysis with what Ares says about it and the citations visible.

Send us a file