# Totally Spies validation shot plan ## Purpose This validation package adapts the Radial PoC structure to the `Totally Spies x Cultshot` marketing system. The goal is not to prove a long continuous scene. The goal is to prove that an owned-asset image and short-motion workflow can lock the franchise's most useful campaign families: - trio-led setup - gadget reveal - HQ or location beauty - villain or threat beat - hero or dialogue closeup ## Goal Prove that a Totally Spies-specific workflow built from owned assets can generate reviewable stills and short motion previews that: - hold character identity across the trio - preserve costume and gadget continuity - stay inside approved locations and brand language - remain editable enough for real campaign finishing - outperform a generic baseline in reviewability and retake efficiency ## Required inputs before the run Minimum inputs for the first validation cycle: - the `docs/internal/training-asset-manifest.csv` rows marked `P0` - 5 owned episodes as the minimum episode pool - official or rights-holder-cleared brand guidelines - approved reference frames for each chosen shot family - a live campaign brief with platform, aspect ratio, duration, and call to action ## Workflow variants to compare For each validation shot, track two variants: 1. `generic-baseline` - generic workflow without the final owned-asset weight pack - used only as an internal baseline for quality and effort comparison 2. `owned-asset-tuned` - workflow driven by approved Totally Spies-specific asset packs and weights - this is the production candidate ## Validation set | Shot ID | Shot family | Primary test | Required owned inputs | Deliverable bundle | Pass focus | | --- | --- | --- | --- | --- | --- | | `TSV-001` | Trio-led setup | Lock Sam, Clover, and Alex in the same frame with readable posing and no face drift | Trio character packs, core mission outfit pack, mission location pack A | still variation board, hero still, short motion preview, metadata package | identity lock, pose readability, no invention | | `TSV-002` | Gadget reveal | Keep the hero gadget readable and correctly attached to the action | gadget pack core, lead character pack, mission location pack A | still variation board, hero still, short motion preview, metadata package | gadget fidelity, hand/object continuity, editability | | `TSV-003` | HQ beauty | Establish WOOHP as a clean, reusable campaign environment | WOOHP HQ pack, logo/title pack, platform ratio pack | still variation board, hero still, optional motion preview, metadata package | location fidelity, brand alignment, clean reuse | | `TSV-004` | Villain or threat beat | Show controlled contrast between hero language and threat language without drift | villain pack A, villain lair pack A, lead character pack if heroes are present | still variation board, hero still, short motion preview, metadata package | villain readability, environment consistency, controllability | | `TSV-005` | Hero or dialogue closeup | Hold a hero face cleanly enough for review and finishing | one lead character pack, core mission outfit pack, approved dialogue or expression reference | still variation board, hero still, short motion preview, metadata package | face lock, expression control, finishing load | ## First-cycle execution brief Use `docs/strategy/top-3-validation-shot-brief.md` as the operating spec for the first three priority validations: - `TSV-001` - `TSV-002` - `TSV-005` ## Step-by-step validation process ### 0) Lock the owned asset pool Before generation starts: - confirm provenance for every `P0` row in `docs/internal/training-asset-manifest.csv` - mark which rows are cleared for project weights - separate reference-only materials from trainable materials Output: - a validated asset manifest with clear rights status ### 1) Build deterministic source packages For each validation shot: - choose the source episode references - record exact episode IDs, timecodes, and asset rows - normalize aspect ratio, color expectations, and naming Output: - one deterministic input package per shot family ### 2) Run both workflow variants For each shot ID: - generate a `generic-baseline` pass - generate an `owned-asset-tuned` pass - keep prompt recipe, seed logic, weight versions, and selection notes Output: - one comparable pair of deliverable bundles per shot ### 3) QC gates Every shot bundle must be reviewed against: - identity lock - costume lock - gadget or prop lock where relevant - location lock - temporal stability for motion previews - brand alignment with the official campaign grammar - no-invention rule - finishing and editability load Any shot that fails a critical gate should either: - move to retake with a documented change path, or - fall back to manual finishing without blocking the whole package ### 4) Measure what matters Track for each shot and variant: - time to first reviewable output - number of variations generated and shortlisted - rounds of notes to reach approval or rejection - which outputs were actually usable - whether the tuned path reduced retake noise and improved approval quality ## Definition of success The validation package succeeds if it demonstrates all of the following: - at least 3 of the 5 shot families reach reviewable quality in the `owned-asset-tuned` path - the tuned path beats the generic baseline on identity and brand alignment - no critical provenance or no-invention failures remain unresolved - the finishing load stays concentrated on hero picks rather than the whole set - the resulting package is strong enough to support a live monthly campaign brief ## Immediate next actions 1. Fill the `P0` rows in `docs/internal/training-asset-manifest.csv`. 2. Request the first 5 owned episodes and ingest them under `materials/source-episodes/`. 3. Fill `docs/strategy/top-3-validation-shot-brief.md` with the real episode IDs, timecodes, and chosen owned reference files for `TSV-001`, `TSV-002`, and `TSV-005`. 4. Select the actual episode references for `TSV-001` through `TSV-005`. 5. Fill `docs/internal/validation-metrics-tracker.csv` as soon as the first baseline and tuned outputs exist.