Package, load, gate — a machine must operationalize its own theory
Parent: disciplines. # A machine must package, load, and gate its own theory — or it won't apply it
The sharpest failure in utter-manual's build: the awareness column had a thick craft theory (the short-video SOP, retention/hook/mute-proof, the communication delivery craft, the persuasion spine) — and the machine produced bare, theory-free content anyway (a silent text-on-black "video" that applied none of it). Having the theory was not the same as the machine using it.
Why the theory didn't get used (three machine gaps, all real)
- It was never packaged as a loadable artifact. It lived as prose analysis under
awareness/, not as a SKILL.md a generator sub-agent could load. There was nothing to load. - It was never in the generation context. The context pack assembled L0/L1/L2/L3 (target, winners, facts, voice) — craft (hook, primacy, three-appeals, retention beats, cadence) was in none of those layers, so the drafter never saw it.
- No craft gate existed on the audit side. The lint hard-gated compliance (answer-shaped, disclosure, provenance, situated-move) and softly judged slop/machine-feel/mid — none covered delivery/video craft. A flat-but-compliant, on-voice, non-mid artifact passed every gate. Craft had zero coverage, generation or audit.
The fix (package → load → gate)
- Package: the theory distilled into two dual-purpose skills — [[um-as-skills|
video-craft]] (the "CPM gate": 3s mute-proof hook, design-backwards-from-retention, hook+final-first, Hook→Problem→Demo→Proof→CTA, prosody, anchor, CTA-by-name) anddelivery-craft(ELM route-match, three appeals, narrative/concreteness, primacy-recency, concision, inoculation, non-mid). Each is simultaneously a generation checklist and an audit rubric, cited by principle id. - Load: the agent loads the craft skill for the artifact kind before drafting (design to the checklist), not only at audit — loading it audit-first is exactly what produced bare first drafts the auditor then had to wave through.
- Gate: a hard
check.craft_floor(K) — the load-bearing subset (text P4/P5/P10/P11; video P1/P2/P5/P8 + the shipped cut carrying VO + music + real footage) is a mechanical floor; a miss FAILs the lint, not just the soft judge. Craft is now enforced like anti-slop, not merely advised.
The loop that got there (generate → audit-against-theory → diagnose → harden → re-audit)
Run agents to generate per platform with the theory loaded, then an audit agent scores each artifact against the theory's own rubric and lists what wasn't applied; a diagnosis agent names why the machine missed it and the fix; harden; re-audit. Passing round, 2026-07-17: every active platform — video, X, Reddit, LinkedIn — applied_all = true, no load-bearing miss, text and video alike. The generalizable move: to know whether a machine actually uses its theory, make the theory an executable rubric and audit the output against it — an unpackaged theory is decorative.