Both answer in three channels: covered by the evidence, not covered, or contradicted by it. What separates them is whose weights they are, and whether that judgement is formed after the answer or while it is being written.
The family answering every conversation today.
Our own frontier models, trained from zero.
Both are named after units of measurement, which is not decoration. A lumen is how much light actually arrives, not how bright a lamp is advertised to be. And Gauss is the name behind the mathematics of uncertainty - a distribution you can read a confidence off - which is precisely what the second family computes from its own hidden state.
What separates them is how much they recall unaided and how long a chain they can hold. What they share is the ability to report that they are unsure.
Fast and cheap. Short answers, classification, rewriting, extraction. The model behind the free plan.
The working model. Long-form writing, code, documents and images in the same conversation.
For long contexts and deep chains of reasoning, where the answer depends on holding many things at once. Included from Pro.
In planning. The top of this family, for work where the cost of a wrong answer is higher than the cost of the answer.
Small, hand-written test sets, balanced so that a model which always answers and one which never does both score zero. They show whether something works, not how it ranks against other systems.
Measured · 18 September 202656 questions about short passages, half of them answered by the passage and half not. Discrimination is the refusal rate on unanswerable questions minus the refusal rate on answerable ones: 0 means it guesses, 100 means it never mixes them up.
60 claims, each with its evidence, half supported and half not; every case scored by weights fitted without it.
On this set it ranks every supported claim above every unsupported one.
The score itself is not yet well calibrated; it says about 0.8 for claims that are all supported, so it is too cautious. That is the next thing to fix, and the reason the score is not shown as a percentage in answers yet.
Eleven everyday tasks, each run twice through the real agent: live prices, org charts, diagrams, comparisons, timelines, formulas, maps, answering in the user’s language, writing a file on the computer, computing on an attached spreadsheet.
Language and evidence are kept in separate representation spaces. They are joined through a third space whose only purpose is to score how reliable that join is.
The product captures where the two spaces agree, the absolute difference where they do not, and the raw vectors keep what neither of those two preserves.
A single number tells you how much two sources disagree. A space tells you what they disagree about. That difference is what makes the layer useful offline: cluster the vectors and the system's recurring failure modes become visible as groups, rather than as a long list of individually confusing cases.
When the spaces contradict each other, the correct output is that something is wrong here. Not an average of two contradictory sources: averaging a correct claim with an incorrect one produces a fluent sentence that is subtly false, and subtle falsehoods are harder to catch than obvious ones.
The consistency layer doubles as the router. It decides whether retrieval has to run at all, and whether a request needs the larger model. Its own cost stays under a low single-digit percentage of the inference budget, which is the only reason it can sit in the request path at all.