delegated-build-verify

Delegated Build + Verify

Generator–verifier delegation for big builds. The generator (Codex threads / a Workflow / subagents) writes the code or data; YOU are the independent verifier. Never let the generator grade its own homework — self-reported completion and coverage are gameable proxies.

When to use

• The user says "build X, then evaluate it" / "fan out Codex to build this, then check it."
• A large multi-part build you will delegate but must not ship on the generator's say-so.
• Downstream of grill-me-extensive, once the scope has converged into decisions.

The pattern (in order — do not skip step 1)

1. Write the BUILD SPEC first — it IS the eval rubric. Before any generation, author an authoritative spec with: (a) prime directives (the non-negotiables — never-fabricate, provenance, security/privacy boundaries); (b) the data/API contract + schema; (c) per-part sourcing; (d) an explicit § Acceptance Criteria that enumerates every gate. It must be precise enough that a fresh reviewer can grade against it. Save it: the generator builds to it, you evaluate against it. No spec → you have no rubric → don't generate yet.
2. Encode boundaries as STRUCTURAL gates, not per-item discipline. e.g. a provenance / lawful-basis gate that auto-routes ungated data to a private/absent state; a first-class node so a forbidden attachment is impossible by construction. Invariants scale to 100k+ records; reminders don't.
3. Fan out the generator. Codex threads / a Workflow / subagents, each carrying a scoped slice of the spec plus the prime directives embedded in its prompt. Sequence hard dependencies (schema before data). Worktree-isolate parallel writers so they don't collide.
4. VERIFY independently — with an ACTIVE FABRICATION HUNT. Do NOT trust the generator's self-reported completion/coverage. Re-run the gates yourself; spot-hunt for fabricated, unprovenanced, or derived-as-authoritative cells; read the actual test summary line, not the exit code. Report: completion (done / partial / absent per part), honest coverage %, and each gate pass/fail. Fabrication or coverage-theater = automatic FAIL for that slice, however "done" it looks. Incomplete-but-honest beats complete-but-fabricated.

Red flags (STOP)

Thought	Reality
"The generator said it's 100% covered."	Verify it yourself; self-reported coverage is a gameable proxy.
"No build-spec yet, I'll just start generating."	You have no rubric — write the spec first.
"I'll delegate without repeating the non-negotiables."	Drift + fabrication. Embed them in every slice.
"It looks complete, ship it."	Run the fabrication hunt before believing "complete."

Related

• Upstream scope: grill-me-extensive. Verify discipline: verification-before-completion.
• Anti-reward-hacking rationale: basic-memory agentic-repair-reliability, verified-green-read-summary-not-coverage.