The calculated mind: it reasons or refuses — and 'cannot bluff' is a property under test

The industry's default failure mode is confident invention: an AI states a fake citation with the same fluency as a true one. Our wager is architectural. If meaning is calculated — reasoned over a small set of semantic primitives — rather than sampled from a distribution, then an answer can be reasoned or refused, never bluffed. Only the surface, the voice, is learned.

This is the lane where we are most careful to show the eval beside the demo, because the strongest results and the most honest gaps sit right next to each other.

No-bluff, at the type level

✅Shipped & measured

Cite-or-refuse is enforced at the type level rather than coaxed with a prompt, and a provenance invariant runs underneath: every word traces back to a meaning it can name. On the clean battery and two honesty batteries we wrote, the mind emitted zero confident-wrong answers, and the no-bluff floor held under a 16-turn adversary that actively tried to make it invent. 🟢

Three qualifications travel with that zero, and all three are ours:

1. It was measured on a non-default code path behind an environment variable, on clean phrasings only. On a 44-turn battery of colloquial phrasing, the default path emitted nine confident-wrong answers. 🔬
2. The cutover to the clean path was never completed. 🔬
3. The first instrument we built to question the mind from outside our own batteries found two confident lies on its first run — Mars and Wakanda, both diagnosed, opposite failure modes. 🔬 We retracted "it never lies" in our own design-of-record on 2026-08-01 and published the whole episode.

We previously published "confabulation 87% → 0.0%" here and on seven other surfaces. No benchmark, corpus, N or baseline model was ever recorded behind those two numbers, so they have been withdrawn rather than restated.

One consequence is small and telling: the voice layer can never invert a negation. "Cannot" cannot silently become "can" — proven on the real wave, not asserted. In a safety context, a flipped negation is the difference between "safe to proceed" and "not safe to proceed", so this is a property, not a nicety.

It cites, from a store that fits on a phone

✅Shipped & measured

Knowledge is compressed hard and kept citeable, with provenance intact. 🔬 Ask it for the capital of Australia and it cites Canberra; ask about Austen, da Vinci, Tokyo, and it answers with the source, from a store small enough to live on a laptop.

We used to print a store size and a fact count here — "60M facts distilled to 22.2M citeable, 20 GB down to 0.70 GB, about 28.6 bytes per fact". Those numbers came from an internal design document, not from a build log: there is no manifest, no du output and no counting script behind them, and our own evidence pass found a 22× disagreement on store sizes between two trees. They are withdrawn until a manifest and a counting command exist.

The crisis floor is handled by meaning, not keywords: it catches "I don't want to be here" and suppresses an essay on the word "perish" — the kind of literal misread that keyword systems make.

The honest gaps — kept visible

🔬Research

Here is where a lesser page would go quiet. We won't.

• A cold stranger scores it 3/10. 🔬 Up from 2/10 (knowledge went 0→5, the crisis response went from FAIL to trustworthy) — but capped at 3 by deflection, noun-grab, a garbled DERIVE step and register misreads. It is not yet a pleasant conversationalist to a first-time user.
• 0.84 is reconstruction, not prediction. 🔬 The mind reaches loss 0.84 on a single 3090 at a median 21 ms per training step and a 377 MiB peak — a training figure, not a serving one, on a genuinely efficient pure stack — but that number measures reconstruction, not next-token prediction. It does not, on its own, prove the mind is better than a trained one.
• "Calculated beats trained at language-model scale" is OPEN. 🔬 This is the real research question, and we mark it open. Closing it needs downstream evaluations, not a lower reconstruction loss.

How it works, briefly

Meaning is composed over a set of semantic primitives with no learned weights; the wave layer renders form and feeling, while a grammar layer gold-fills reference (so it says "Tokyo", not a plausible-sounding wrong noun). Six reply moves speak from the composed meaning rather than predicting the next token — which is exactly why a refusal is available as a first-class outcome.

Why it ladders back

A mind that reasons or refuses is a mind you can trust — and trust is the thing that turns a capable tool into an empowering one. We would rather ship a 3/10 stranger score and publish the two lies our own outside-in test caught than a 9/10 charmer that never mentions them.