Planner / verifier split

A planner extracts requirements and derives failure cases; an adversarial reviewer or evaluator must be convinced and can send the plan back; an implementer builds; a verifier checks. Each context stays small enough to stay reliable, and the boundary between design and execution becomes an explicit gate a human can stand in. The reported failure that motivates the split is blunt: one agent doing all four blew the context window outright.

Sequential checking degrades through context rot — by item five the agent is carrying irrelevant context and gets measurably worse. With ten to twenty requirements per ticket, a single verifier consistently skipped some. The fix that held was twelve to fifteen sub-agents in parallel, one per requirement, under an orchestrator collecting verdicts, with each sub-agent scoped to its own surface so a front-end reviewer is never asked to reason about the backend.

A verifier reading the diff is anchored on the solution the implementer happened to choose, which is how hallucinated requirements creep in. Grounding it in the base branch plus the specification instead forces it to derive what should exist before it looks at what does.

Also: Science Claw, Infinite.

Assay’s desk topology is exactly this shape at repo scale: brief authoring plans, the worker desk implements, the review desk reviews adversarially, the verify desk checks after merge. Four roles, four contexts, four distinct identities, defined in the role skills and the lifecycle. The frontier reports here are peer confirmation rather than a new direction.

A human spec-comment gate. Document-style commenting on the spec before implementation strengthens the case that brief review should be a first-class human step rather than a rubber stamp; Assay has brief review in the lifecycle but no dedicated comment surface. Designed, unbuilt. Per-requirement fan-out. The verify desk runs a brief’s Verify table in a single pass; spawning one sub-agent per Definition-of-Done line with the desk as orchestrator is a small structural change to an existing loop and the highest-confidence steal here, given that the sequential form is reported to skip items. Not built. Base-branch grounding. Both review and verify read diffs — correct for reviewing a change, wrong for spec-fidelity verification. Cheap to A/B inside one repo. Not built. A breaker role. The verify desk checks conformance to the brief; nothing in Assay is chartered to attack the brief itself — to falsify the requirement rather than the implementation. The pre-execution policy check is the mild version and would be a cheap extra layer for briefs on high-consequence surfaces; the breaker agent is the strong version. Neither exists.

• A published precision number on per-requirement fan-out versus sequential verification — the failure is described qualitatively (“consistently skipped some”) with no miss rate attached.
• Whether anyone reports the base-branch-versus-diff comparison with numbers rather than as a design preference.
• Whether a breaker or falsifier role appears outside scientific-discovery settings, where its economics are unusually favourable.