The executor acceptance test

How to verify that a body of SOPs is actually executable — the owner's method:

Spawn a clean-context agent. Give it (1) the product's documentation and (2) the SOP corpus — nothing else, no web. Ask it to plan and simulate the operation end-to-end, under one hard rule: every action must cite its guideline (file + step). Where the corpus doesn't determine a step, it must not improvise — it logs a structured gap ({needed, where it looked, why uncovered, severity: blocking/degraded/cosmetic}) and continues under a labeled assumption. Then audit: verify the citations are real, and treat the gap log as the defect list.

Why it works:

• The corpus is the system under test; the agent is the test harness. A clean context can't paper over holes with conversation memory the way the author can — it reads only what's written, which is exactly what a future executor (human or agent) will have.
• Citation-or-gap forces the distinction between "the corpus decided this" and "I decided this" — the second category, made explicit, is the theory-practice gap.
• The severity triage prioritizes the fill: blocking gaps are missing assembly (do the work), degraded gaps are missing rules (find tested theory/cases/norms — never invent), cosmetic gaps are tolerable assumptions.
• The fill discipline mirrors the corpus's own sourcing law: gaps get filled with established theory, documented cases, or norms that survived testing — a gap is not a license to make something up.

First run (2026-07-16, StandMeet × the awareness SOPs): the executor completed a fully-cited 30-day plan and returned 9 gaps — 1 blocking (the per-project channel kit didn't exist for the test product), 6 degraded (no hour-budget rule, no arena-sequencing trigger, no pre-launch genre, no pre-funnel measurement, a promised-but-unwritten fit table, a missing profile step in one SOP), 2 cosmetic — plus one structural observation (the decision layer had no slot for thesis–arena resonance). All six degraded gaps were filled the same day from tested sources (Bullseye/Traction for sequencing; the corpus's own pre-launch precedents for the rest); the blocking gap is per-project assembly work.

Second run (2026-07-16, FlexMesh × the now-thickened corpus): the distance to application, measured. After a full theory round (matrix, context-manifest, engagement dial, being-seen ladder, video pipeline), the executor reported a markedly better verdict: the theory layer executed cleanly and cited to real steps — fit table, order-of-operations, engagement dial, labor+time-to-signal table, Bullseye, and the context-manifest compilation all ran deterministically; the concentration/probe split fell out without invention. The residual gaps were dominated by (a) per-project assembly, not (b) theory holes: the DemoForge 9:16 build, an unassembled FlexMesh voice corpus and product-facts pack, the two unbuilt instruments (per-post log, winners-scan), and store-console/attribution wiring — all owner work, not corpus work. Exactly one genuine rule-gap surfaced: the Reddit/persona doctrine assumed founder ∈ users and gave no procedure for a non-user maker — filled the same day from documented norms and cases (outsider-founder-earns-voice). The distinction the second run sharpened: the theory is close to applicable; the distance that remains is per-project assembly, which the test correctly refuses to let the corpus fake.

Kin: industrialization-needs-a-mechanical-arbiter (the test converts "is the SOP complete?" from opinion into a mechanical verdict) · the vault's pre-commit lints (same philosophy at the note level) · trust-follows-provenance (both refuse assent-as-evidence: verify by re-derivation from the primary source, not by a fluent claim).