Concepts
This is the unlearning section.
If you’ve spent the last few years building on generative models, you’ve absorbed a set of habits: prompt engineering, output parsing, retry loops, treating a model’s self-reported confidence as decoration. Almost all of it is unnecessary here, and some of it is actively harmful.
These pages replace it.
The core mental model, borrowed from Kahneman: fast reflexes versus slow deliberation, and why AI systems are splitting along that line. Start here.
Exactly what goes in (state and typed questions) and what comes out (typed answers with probabilities), field by field. The same format works for Kenning, Clef and Jev.
What “no hallucination” means (and doesn’t)
Section titled “What “no hallucination” means (and doesn’t)”Removing token generation removes a whole class of failure: malformed and invented output. It doesn’t remove being wrong, and it doesn’t make adversarial input harmless.
How to read the probabilities, how to verify they mean what they say, and how to pick thresholds from your cost of error instead of from vibes.
Cross-encoders, calibration by temperature, and distillation from a larger teacher: why these models behave so differently from chat models, and what that means for you.