Fusion uses separate verification stages because executable tests, comparative judgment, and adversarial release review establish different kinds of evidence. None should impersonate the others. Each stage asks one question, records one kind of answer, and hands the next stage a boundary it cannot argue past.
The gate asks whether hard claims execute
The objective gate runs deterministic commands against a pinned subject SHA: builds, tests, lint, type checks, and static criteria. It can prove that tests passed under a recorded environment, a schema validated, or a required file remained unchanged. A failed hard command eliminates the candidate from comparative judging unless the contract defines another route.
It cannot prove that the requested workflow is the clearest one for a user. It also cannot cover a condition no one encoded.
Green tests do not choose an architecture
Two candidates may pass every executable criterion and still differ in scope discipline, maintainability, tradeoffs, risk handling, explanation quality, or fidelity to a manual requirement. That is where the jury comes in: blind judges compare candidates that have already met the hard gate, on rubric-bound concerns the test runner cannot settle.
The jury reads the RunSpec, baseline, risk surface, and gate evidence. It can compare those qualities without pretending they came from the test runner. Its rationale is evidence, but it remains model judgment, so Fusion records citations and disagreement so a human can challenge it.
Taste cannot waive a failed gate
Fluent rationale is not counterevidence to a reproducible failure. The jury receives the gate result as a boundary on what it may consider, not as one more opinion to weigh.
Integrators ask whether the strengths can coexist
Candidate comparison often reveals that no single result owns every strong choice. Integrators synthesize complete finals and rerun the gate. Their work is a new implementation, not a clerical merge.
QA asks whether the selected final is safe to hand off
Fresh QA starts after the tournament has a preferred final. The quick review catches obvious ship blockers. The required deep audit challenges production readiness, regression risk, security boundaries, and evidence completeness.
QA does not crown a prettier alternative. It can pass the final, route a bounded repair, or block readiness.
The receipt keeps each claim in its lane
A strong verdict bundle tells the reader which stage produced each fact. "Test passed," "jury preferred," and "QA found no blocker" are useful together because they mean different things.
The stack is deliberately redundant at the handoff point. Software deserves more than one kind of question before a human authorizes shipment.
Source ledger
- Merged strategy, section 2.
- Fusion MVP PRD, sections 6.4, 6.5, and 7.
- Traceable output bundle specification, sections 5 and 6.