Verify

Done means verified.

The agent that does the work doesn't grade the work. Every deliverable faces its acceptance criteria — tests, evidence, policy, independent review — and can be rejected.

Problem

Self-graded homework isn't quality.

“The AI says it’s done” is the industry’s quietest lie. Self-graded homework isn’t quality; it’s optimism with latency.

Mechanism

Five check lanes.

Deterministic tests → brief conformance → evidence audit → policy/safety → independent model review (a different family than the producer). Verdicts: accepted · conditional · rejected · escalated. Rejection creates remediation with exact findings; re-verification runs the failed checks plus a regression set.

Anatomy

A VerificationResult, real from the Aster run.

VerificationResult · REJECTED

deliverable: landing page v1
verdict: REJECTED
finding: “Claim ‘proven to increase win rate 40%’ — no supporting evidence in company state.”
evidence inspected: 7 items
verifier: independent (family B)
remediation: copy revision task → v2

VerificationResult · ACCEPTED

deliverable: landing page v2
verdict: ACCEPTED
regression: 12 checks re-run

Worked example

Two rejections, two saves.

The second Aster rejection — signup intent lost after auth — shown as a failing test trace, then a fix, then green. One brand save, one product save.

Control

Independence, by rule.

The producer never sole-verifies its own work · high-risk work is checked by a different model family · human override is recorded with rationale and never erases the failed check · Verify has no release authority beyond policy — it gates, humans and contracts release. Customer-side human reviewers slot in wherever you require them.

UnderstandDecideDeliverVerifyLearn

The 'Verified by Brainis' mark means this ran.

Hard question

Can it reject MY work?

Only if you give yourself acceptance criteria — many teams do, because a second set of eyes that never gets tired is the point. Human deliverables use the same evidence contract; your reviewers can be verifiers too. Nothing about Verify is aimed at agents alone — it’s aimed at “done” meaning something.