docs/creator-v10/editing-discovery-engine.md

Editing Discovery Engine

Status: structural core shipped (src/creator-intelligence/discovery/), gated. Idea: an always-on lab that discovers recurring editing techniques and turns them into a reusable playbook — the honest realization of the "Editing Discovery Engine" loop (DISCOVER → SEGMENT → HIGHLIGHT → PATTERN → PLAYBOOK → VERIFY).

What it is

Composes the measured detectors already in the dashboard into named, reusable editing techniques:

  • SEGMENT (segment.js): a clip's analysis → role-labeled beats in the spec

time buckets (0-1 / 1-3 / 3-7 / 7-15 / 15s+). Roles — hook, build, surprise, payoff, cta — come ONLY from measured signals (A1 cuts, A2 novel tag, A3 speech hook/CTA, A4 retention). Consecutive runs are collapsed so a sequence reflects transitions.

  • PATTERN DISCOVERY (pattern-mining.js): mines recurring role sub-sequences

(length 2–4) across clips, counts how many clips share each, and names them ("hook → surprise → payoff" = Quick Payoff; "hook → build → surprise → payoff" = Delayed Reveal).

  • PLAYBOOK (playbook.js): patterns → techniques with plain-language steps.
  • Orchestrator (index.js): discoverFromClips(clips, opts) (pure) +

discoverFromDisk() (loads analyzed entries + outcomes from the calibration store).

The honesty contract (the VERIFY stage)

This is the whole point, and it is enforced in code + tests:

  • A sequence is a "pattern" only when ≥3 clips share it (structural frequency).
  • A pattern's performance is insufficient_data until ≥5 of its example

clips carry a real first-party outcome → then directional; ≥20 → calibrated. Frequency ≠ performance; they are never conflated.

  • A technique is never promoted to "verified" without that data. With zero

outcomes (today), the playbook is purely structural and every technique reads insufficient_data.

What it does NOT do (the wall, made explicit)

It does not scrape other creators' top Shorts/TikToks/Reels for replay rate / shares / engagement-per-second. No platform API exposes those per-video private metrics, bulk-downloading the videos is ToS + copyright, and inventing the numbers is exactly what the no-fabrication rule forbids. So DISCOVER operates on clips the operator actually has — own renders/uploads + a curated reference set — and "performance" comes from the operator's own analytics (the calibration store, docs/creator-v10/learning-pipeline-research.md).

Dependency / when it becomes powerful

The engine is the consumer of the calibration loop. It can discover and cluster patterns now (exploratory), but its headline output — verified techniques — only lights up once real outcomes are imported. Until then it honestly reports insufficient_data. Run it after analytics flow in, not "forever" before.

Tests

tests/test_discovery_engine.js (10 checks): buckets, role labeling (incl. measured-only CTA), template naming, pattern discovery (≥3-clip floor), the insufficient_data → directional gate, and that the playbook never claims a verified technique without data.

Not yet built

Persistence (editingPlaybook.json / highlightLibrary.jsonl), an HTTP route + UI for the playbook, segment-by-context "best for", and the experimentation loop (reuses the B1 variant engine). All deferred until there's a reason to surface them.