new·Plugin 0.5.0: agents verify and hand offTwo new MCP tools, live for every connected agent right now. brain_verify checks a claim against a brain before you act on it — supported / contradicted / not_covered, with the evidence notes; anything unconfirmed is flagged to the brain's owner as a gap. brain_handoff is the baton between sessions and agents: stop mid-task, leave where you got to, and the next session — this agent tomorrow or a different tool entirely — takes it and continues. brain_list now shows open handoffs, and /mind grows a relay feed.all news →
mozg.betaSign in

Knowledge audit · exam-as-a-service

Your knowledge base claims. Ours examines.

Every RAG and memory tool quotes benchmarks nobody can reproduce. mozg's exam is its own machinery — the same one that grades every brain in the catalogue, in public — and it will happily grade yours: send the corpus, get a dated report of what it actually knows.

What the report says

measured, dated, reproducible
The scoreControl questions are written from your corpus's own stated goal, then answered ONLY from what retrieval returns — the same honest constraint your users live under. Weighted, majority-voted.
Category coverageNot one number but a map: which subjects your base actually answers, which it fails, and which questions exposed the gaps — the fix list comes free.
Judge agreementEvery verdict is voted by independent judge passes; the agreement rate ships in the report. A score without its own error bars is marketing.
Anti-bluffPlausible questions just outside your corpus's scope. A base that confidently answers what it cannot know fails customers quietly — this measures it loudly.
The dateKnowledge rots. The report is dated, and re-audits diff against the previous sitting: what was learned, what was lost.

How it runs

your data stays yours

You export your corpus (JSONL, markdown, or an API dump — we adapt), it is imported as a private brain nobody else can reach, the exam runs, the report is delivered, and the imported copy is deleted on request. The first audits are hands-on with us in the loop — that is deliberate, not a beta apology: the report format is being shaped by real corpora.

First three audits — free, in exchange for a public result.

Your tool's exam score, published with your sign-off, methodology attached. You get an independent number to cite; we get the proof the exam grades anything. Write what your base is and roughly how big — a person answers, usually same day.

Start the conversation

methodology: the same exam every mozg brain sits — question generation from goal + corpus, retrieval-only answering, 3-vote judging. Nothing bespoke, which is the point.