Research
Linear Anti-Hallucination Protocol
A governing discipline for keeping long-running work anchored to evidence and to the intention it started from.
The problem
A model can be fluent, competent and wrong, and nothing in the output separates the three. Confidence is not a measure of correctness, and a system that treats it as one accumulates error quietly.
Over a long project the cost compounds. An unsupported claim made early is repeated back later with more certainty than it was first offered, and by then work has been built on top of it. The error is not in any single step. It is in the absence of anything that would have stopped the claim becoming state.
This is a different failure from a wrong answer to a single question, and it is not addressed by making the model better at single questions.
What it asks for
Intention is preserved, not inferred
What the work is for is recorded once and carried forward, rather than reconstructed from whatever is still in view.
Evidence precedes assertion
A claim is not recorded as settled until something has been checked. The check is mechanical, because a reviewer under time pressure accepts a plausible report.
Provenance travels with the claim
Every recorded fact carries where it came from, so a later reader can tell an observation from an assumption.
Uncertainty stays explicit
Unknown is a valid and recorded state. Silence gets filled in by whoever reads next; an explicit unknown does not.
Drift is detectable
Deviation from the stated intention should be visible as a signal, not discovered later as a surprise.
What this is not
It is not a claim that hallucination has been solved. It is a discipline for reducing the rate at which unsupported claims become the basis for later work.
It is not a published result. There is no benchmark here, no evaluation, and no third-party validation. Where that changes, this page will say so.