SFLO: lights-out only advances when the gate artifact exists

simonasrazm/simon-factory-lights-out (SFLO, Mar 19; still updated mid-Sep) is a gated pipeline protocol for building with agents: six sequential stages, each producing a required artifact. No artifact, no progress. No skipping.

Gate 1 (Discover / PM) writes SCOPE.md. An inner loop then alternates Gate 2 (Build → BUILD-STATUS.md) with Gate 3 (QA → QA-REPORT.md) and Gate 3.5 (Security → SECURITY-REPORT.md), capped at ten rounds, bouncing back to Build when scores fall below threshold or a critical finding appears. Only then does Gate 4 (PM Verify → PM-VERIFY.md) grade the whole; a non-A grade returns to the inner loop. Gate 5 (Ship → SHIP-DECISION.md) is the exit. Scout matches specialized agent directories (BRIEF.md + SOUL.md) to roles from the user prompt; pipeline.yaml owns models, thresholds, skills, and runtime policy across Codex, Cursor, and Claude installs. Multiple factories can run in parallel under .sflo/ without sharing locks or state. Deterministic checks on project-relative deliverables named in SCOPE.md keep Gates 2–5 from treating prose as proof.

The author’s own note points a newer Waydriver variant as stronger for some workloads — SFLO itself remains the explicit lights-out contract: progress is artifact-shaped, not chat-shaped.

Why it matters for code review

Continuous overnight momentum fails when “done” is a model’s last message. SFLO’s review surface is the gate file set: if QA or Security never wrote a report that clears threshold, the factory does not advance — the same fail-closed instinct you want on a merge queue. Steady harnessed throughput compounds past a heroic Claude burst only when each overnight hour leaves auditable artifacts a morning reviewer can open without replaying the session. Bursts still win for inventing scope; the harness wins for refusing to ship without the paperwork.