new·The score now tells you which way it movedA brain's exam only ever grows: its own material writes questions, and so does every question a real caller asked and did not get answered. The score is a percentage over that growing set, so a brain that learned more could post a smaller number — and this week three did. One of them answered two MORE questions than the week before and showed eighteen points less. Printed as a single percentage, that reads as decline to a reader and as punishment to anyone who contributes material.all news →
mozg.beta
Sign in

Building With AI Agents · all subjects

verification/exam

6 notes, read out of this brain and free to use. Each one was extracted from a source and is re-checked against its exam.

Unchanged exam material keeps its verdict in incremental re-sits

A pass in a category whose notes have not moved since the last run is carried into the new run as a first-class row. Re-judging unchanged material produces only noise. Failures are always re-judged so recovery stays possible. Fresh checks are judged fresh. The log says how many verdicts were carried.

Parent exams must retrieve across the family

A parent's exam retrieves across its family, because that is what asking the parent returns. Scoring it on its own notes alone made the number a lie in the owner's favour: an agent asking stake-engine gets 695 notes, while the score described 29.

Write the exam goal from the specification, not from memory

The parent's goal asked about 'the lifecycle from initialize to shutdown'. The 2026-07-28 specification removed that handshake and made the protocol stateless, so the exam kept asking questions with a false premise. The exam caught a wrong assumption in the goal rather than a gap in the material. Retrieval was fine the whole time; it returned the note saying the handshake was removed.

Brains map to goals, and changed goals must clear their exams

The goal generates the questions a brain is scored against, so editing one without clearing them leaves the brain graded on a promise it stopped making. Three paths write goals now, so it lives in lib/goal.ts and they all go through it. Hand-written checks survive — those are the owner's own questions, not a restatement of the goal.

Exam gains majority-vote judging for robustness

The exam now uses majority-vote judging instead of a single judge, making scores more robust to model variance.

Owner-written checks let experts verify material

Exam gains owner-written checks, so an author can define custom questions alongside the goal-derived ones.

Give your agent this brain