The first AI coding agents usually produced credible implementations. The biggest improvement came when fresh reviewers challenged the work, combined the best ideas, and tested the finished result through paths the builders missed.
A Verified Receipt means an objective gate executed and every hard criterion of a frozen contract passed, with evidence pinned to the exact commit. Not a summary, not a vibe: a checkable record.
The RunSpec is the one contract the build, jury, integrators, QA, and the receipt share: a versioned spec that freezes the objective, scope, criteria, risks, and proof meaning before agents build. Freeze it before implementation and a changed goal becomes visible instead of convenient.
A candidate can change implementation files, but it cannot redefine the protected checks that decide whether the result passed. Fusion restores gate-feeding inputs from the pinned base, verifies their fingerprints, then runs the objective command against the candidate SHA.
The first jury finds the strongest candidate ideas. The second jury verifies that synthesis actually produced a better final. Integration changes the artifact, so it must not inherit the candidate verdict.
Picking the best candidate leaves value on the table, but combining candidates creates new failure modes: dropped grafts, flattened dissent, new regressions. Fusion makes configured integrators produce competing complete finals, re-gates each one, and returns them to jury pass 2.
The agents closest to a build share its context and its blind spots. Fusion brings in fresh QA after jury pass 2: first a quick ship-readiness filter, then a required independent deep audit that looks for defects the builders, judges, and integrators may all have learned to overlook.
"Failed" is too small a word for a verification run. Fusion records typed stop states so the operator can distinguish bad code, stale evidence, jury contention, infrastructure failure, and a retryable publication problem. Each stop state needs a different recovery route.
A Fusion Receipt gives every frozen criterion its own evidence row spanning six stages: gate, jury pass 1, Integrator disposition, jury pass 2, QA, and final status. The row follows one promise across the whole run without hiding absent proof.
Fusion separates verification into stages because executable tests, comparative judgment, and adversarial release review establish different kinds of evidence. The gate decides what executes, the jury decides which gate-passer best serves the intent, and fresh QA decides whether the selected final is safe to hand off.
Every run produces a manifest: the contract, the verdicts, the gate result, and the decision. Four fields tell you most of what you need before you open a single diff.
Mode, profile, and roster are independent controls. Mode decides who drives the workflow. Profile decides whether one builder or a tournament runs. Roster decides how many candidates, judges, and integrators participate and which engines fill those roles.
Fusion's hosted side coordinates runs, stores receipts, and publishes checks. By default it does not host your code or your keys. The conductor keeps time; the kitchens stay yours.
A Fusion run moves through Discovery, Plan, Build, Review, and Output. Each stage creates a different artifact, owns a different decision, and hands a narrower claim to the next stage.
Tokens are the primary unit for every engine in a run, metered per role: candidate lanes, judge seats, integrator, QA. Dollar figures from subscription CLIs are estimates. The usage record ships on the Receipt.
Fusion prices by verification depth: how many candidates compete, how many judges vote, how deep review goes. The meter is rigor, not tokens.
Model identity comes from one place: the roster assignment the run lead records in the manifest before the run starts. We never reconstruct it from branch names, commit style, or output feel.
Reading a diff is not evidence. In Fusion, every review claim is backed by execution: commands run, exit states decide, and a finding that cannot be reproduced is downgraded or fails closed.
Judges stay blind during a run for fairness. After it closes, an outcome survey unblinds every role, and results accumulate into the win ledger: a per-engine, per-provider, per-model strength record that guides future roster choices.
After judging, an integrator builds one complete final from the winning candidate plus approved grafts from the others, then re-runs the gate. You merge one thing, not a pile.
A Definition of Good is the frozen, testable contract a run must meet: hard criteria that can fail the run, soft criteria that shape it, and a gate a machine can execute.
Posts are MDX files committed to the repo, reviewed in pull requests, and served as static HTML. The frontmatter mirrors a future CMS collection, so migrating later is an import script, not a rewrite.
Fusion runs the same task through competing coding agents in parallel lanes, then keeps only what survives judging. Competition is the quality mechanism, not a gimmick.
AI made code cheap to produce and expensive to trust. The fix is not a better generator; it is a verification layer that proves what survived.