VariantEngineV2

Generates up to five edit variants (A–E) from one analyzed upload, each differing along controlled dimensions, then ranks them. Ranking is honest: it reports prediction confidence as insufficient_data until the dataset crosses thresholds.

Variation dimensions

Each variant is a deterministic recipe over real analysis output (HighlightTimeline, SafeZones, events):

Dimension	Options
hook	instant_payoff · cold_open · countdown · question_text
caption style	(any style id from CaptionEngineV3)
cut timing	tight · medium · loose (derived from highlight density)
music timing	beat_sync · ambient · none
zoom timing	on_event · steady · punch_in
ending	hard_cut · loop_back · cta

Default A–E presets cover a spread (e.g. A = tight/instant_payoff/beat_sync, E = loose/question/ambient). Recipes are stored so a chosen variant is reproducible.

Variant object


{

  "id": "variantA",

  "recipe": { "hook":"instant_payoff", "captionStyle":"mrbeast_bold", "cut":"tight",

              "music":"beat_sync", "zoom":"on_event", "ending":"hard_cut" },

  "renderPath": null,

  "prediction": { "status": "insufficient_data", "have": 0, "need": 500 }

}

renderPath is null until the variant is actually rendered (ffmpeg). The dashboard shows "Not rendered" — not a fake thumbnail.

Ranking & prediction

Ranking calls scoring/score-engine.js:

Dataset ≥ MIN_ROWS_FOR_SCORING: each variant gets { status:"ok", value, basis } for completionRate / viralScore / shareRate, computed from how similar real high-performing rows were edited. Variants sort by viralScore.
Dataset below threshold: every variant returns { status:"insufficient_data", have, need }. The UI ranks by deterministic heuristic order (e.g. tighter cuts first for gaming) and clearly labels it "heuristic order — no performance data yet", not a predicted percentage.

This is the crux of the honesty rule for variants: we will happily generate five edits with zero data, but we will not claim "Variant B: 91% completion" unless a real dataset supports it.

Continuous learning loop

When the operator keeps a variant, training/learning-store.js appends a variant_selected EditEvent. These first-party signals accumulate locally and, once past threshold, let the score engine personalize ranking to the operator's own results — still data-backed, still auditable.

Feature flag

Behind variantEngineV2 (LANTERN_CI_VARIANT_V2, default off). When off, the editor produces the single highlight render plus the existing A/B/C render slots.

Implementation steps

Recipe presets A–E + reproducible recipe storage.
Renderer that applies a recipe via ffmpeg (cut list, zoom keyframes, music mix, caption burn).
Ranking via score-engine with the insufficient_data path wired to the UI.
variant_selected learning hook.