Worked example

A full cycle, scored

We ran the stack against a fictional project, end to end, and kept the receipts. One feature shipped with notes. The next one got rejected, was corrected, and shipped — which is the part worth watching, because a quality gate that never says no isn’t a gate.

The project

Pulse — a fictional async feedback inbox for solo SaaS founders. Small enough to build for real, rich enough to exercise the gates: typed data, mandatory loading/empty/error states, a PII constraint, and one genuinely interactive feature. Two features were driven through the cycle.

Run 1 — F-001, a clean T2

The read-only inbox list. Every gate fired in order: BOB produced the direction brief and waited for approval — gate ①; once approved, RAY’s pre-flight passed (vision complete), it declared the tier and wrote a spec with a mandatory ## OUT OF SCOPE block, frozen against the direction — gate ②; BOB then implemented through its 6-step loop. ANALYZER graded it:

The real finding: the “you’re all caught up” empty state only triggered on an empty list — but the spec’s own user story said it should also trigger when every item is archived. A genuine gap, caught by reading the spec against the code, not rubber-stamped.

The logic layer was compiler-verified, not asserted:

$ tsc --strict --noEmit  src/lib/feedback/*.ts
exit 0

CA-2 sort:  new:f2 > new:f1 > triaged:f3 > archived:f4   (newest untriaged first)
CA-7 mask:  maya@acme.io  ->  m•••@acme.io               (PII never rendered raw)

Run 2 — F-002, the reject loop

The interactive layer: change an item’s status in place, list re-sorts. This is where the gate earns its place. BOB’s first pass worked on the happy path but cut corners — and ANALYZER sent it back.

Cycle 1 → ❌ 11/20 — Rejected. The deductions were mechanical, not opinion:

• any used to silence a type — ADR-004 violation, −2
• a debug console.log left in the handler — security check, −1
• the update was fire-and-forget: no rollback on failure, and a success toast that fired even when the write would fail
• no pending state on the control

That score lands on the ladder’s 10–13 → back to BOB rung — a rework, not a re-spec. ANALYZER returned three binary correction criteria (not a vague bug list):

Cycle 2 → ✅ 19/20 — Shipped. BOB re-entered from those criteria and closed each one. Re-verified, not taken on faith:

$ grep -c 'any|console\.' InboxClient.tsx StatusControl.tsx
0        # any
0        # console.log

setItems(snapshot)   # rollback present
toast.error(...)     # actionable failure path
pendingId            # in-flight state
tsc --strict … exit 0

One thing ANALYZER caught that wasn’t BOB’s fault: the spec contradicted itself — one criterion said “single click” while another mandated a dropdown (which needs two). That got routed back to RAY, not BOB. The loop improves the spec author too, not only the builder.

What this exercised

Every mandatory mechanism fired, in order, and did real work:

Honest limits

• Self-run. One session, agents role-played — a demonstration, not a third-party benchmark.
• Pulse is fictional. Representative of a real build, but built to be tested.
• Only the logic layer was compiler-verified. The pure TypeScript (sort, mask, excerpt) passed tsc --strict; the .tsx components weren’t compiled in isolation (no design-system resolution in the sandbox).
• Scores are ANALYZER’s judgment, against the spec and design guide — not an external grade.

What it does show, verifiably: the gates aren’t decoration. They blocked, they scored, they rejected a corner-cut, and they routed the fix to the right owner.

Next: Get Started · Core concepts