- Entry date
- 1 May 2026
- Category
- Models
- Lead over the world
- the thesis, not the lead
- Access
- 🔒 Protected
Compute meaning instead of guessing it, and a system can only do two things with a question: answer it, or refuse. It cannot bluff in between.
- reason or refuse
- the core architectural bet — under test, not proven
- 0 trained weights
- in the reasoning core itself, per the design
- 1 May 2026
- the dated start of this research line
Honest evaluation
This entry marks the thesis's origin; whether 'cannot bluff' holds is a design property under active test elsewhere, not reported as measured here.
Runnable proof — see it work
What we learned — a preview
The honest edges are open even when the engine is not. The full report — what it was, what we built, and the measured internals — is protected.
- This is research-stage, not a shipped reasoning engine — say so plainly.
- 'Cannot bluff' is a design property being tested, not a guarantee — and on 2026-08-01 it failed: the first instrument built to question the mind from outside our own test batteries found two confident lies on its first run. The episode is published at /lab/never-lies-retracted.
- Engine internals are deliberately kept off this page as core IP; this report covers the thesis only.
The deep body of “The calculated mind — where the 'reason or refuse' thesis began” is behind access.
We open the demos, the specs, and the method; we protect the engines, the model internals, and anything that touches commercial, safety-critical work. This report describes an engine — so its details are gated, and no client, company, or project is named. Members can read it in full; if you have a genuine reason to see it, tell us who you are and why.
Proofs & sparks
We demonstrate rather than assert. Each ✅ proof is a visible result with a hard figure.
- It can't bluff0 confident-wrong, on our own batteriesthe no-bluff floor holds under a 16-turn adversary on the clean battery and two honesty batteries; every word traces to a meaning, so it cites or refuses. On a 44-turn battery of colloquial phrasing the default path emitted nine confident-wrong answers, and on 2026-08-01 the first outside-in instrument found two lies — see the retraction.
- Citeable knowledge, compressedcites or refuses, from a local storeeach answer traceable to a source (Canberra, Austen, da Vinci) or refused outright. The store size and fact count we used to publish here traced to a design document rather than a build log, and are withdrawn until a manifest and a counting command exist.
- A mind on one 309021 ms/training step × 377 MiB × one 3090the calculated mind trains at a median 21 ms per step in a 377 MiB peak on a single RTX 3090 over 250 steps, pure stack — 3 of 4 meters green, cold-judge honestly at 3/10. That is a training figure; no serving latency is published.