# Higgsfield Character Pack Contract v0.3

Status: canonical production-readiness contract for all active Luminautas character versions.
Supersedes `HIGGSFIELD_CHARACTER_PACK_v0_2.md` for new work. Historical files remain readable for provenance.

## Principle

A poster, model sheet or generated character board is a **candidate source**, not production approval. Every exact character version needs a reusable reference pack, canonical voice/TOV binding when speaking, provider Element/reference binding where supported, Golden Character Test and QA record.

All seven active mascots are first-class production identities. **Lumi remains the franchise protagonist and exploration leader.** No pair is a mandatory production gateway.

Meaningful visual/voice/personality change creates a new version. Never silently overwrite an Element/reference set or canonical voice identity used by published content.

## Production phases

`character_rnd → candidate → extracted_reference_pack → visual_QA → voice_audition → provider_binding → golden_test → production_approved → active → deprecated`

Production use is phase-aware:
- broad identity readiness is maintained for all seven;
- deep readiness follows the exact cast required by the next approved story;
- Lumi receives continuous deep-readiness priority because she anchors the flagship franchise.

## Final character gate

Required for speaking/visible final character generation:
- exact `character_id/version`;
- approved `visual_reference_set_version`;
- approved turnaround slots;
- approved expressions needed by the scene;
- approved pose/action slot or reviewed scene-specific candidate;
- `higgsfield_element_id` or validated equivalent identity binding when used;
- exact canonical `voice_identity_id` when speaking;
- exact provider `voice_type + voice_id` only after audition/approval;
- approved TOV version;
- voice identity/pronunciation/emotion/lipsync QA when speaking;
- passed Golden Character Test;
- identity >=95;
- anatomy/shape >=95;
- signature-feature consistency >=95;
- expression readability >=90;
- child-safety PASS;
- IP-distance PASS;
- zero unresolved P0 continuity defects.

Do not fabricate scores or provider IDs. Keep values `pending/null` until measured or explicitly approved.

## Required reference slots for every active mascot

### Turnaround
1. front
2. back
3. left profile
4. right profile
5. three-quarter left
6. three-quarter right
7. close neutral face
8. silhouette test
9. scale chart

### Expressions
Minimum:
- neutral
- happy
- delighted
- curious
- thinking
- uncertain/confused
- surprised
- worried/concerned
- focused/determined
- relieved
- laughing
- character-specific signature expression

### Poses/actions
Minimum:
- standing/resting
- locomotion appropriate to anatomy
- observing
- listening
- indicating a clue where anatomically appropriate
- explaining/responding
- canonical prop interaction
- safe problem solving
- setback/wise pause
- determined action
- celebration/relief
- friend interaction

### Format/context tests
- 16:9 solo medium/full
- 9:16 solo hook
- one story-selected pair test when the character is scheduled in a pair
- one justified specialist-pair test where relevant
- one complex-environment identity test

No single historical duo is required for every mascot approval.

## Character-specific locks

### CHAR-LUMI-v1 — franchise protagonist
Lock:
- canonical child age/proportions;
- stable face/hair silhouette;
- canonical explorer outfit/backpack/accessories;
- warm expressive face;
- brave-curious leadership body language;
- Curiosity Graph interaction language without superhero/magic drift.

Must test:
- brave uncertainty → thoughtful determination;
- child-facing question;
- Curiosity Graph clue collection / zoom-out / missing-link reasoning;
- map/console or real-world decision;
- natural movement;
- story-selected friend conversation;
- specialist listening/asking-help beat.

Reject:
- adult/teen drift;
- passive presenter treatment when designated lead;
- unsafe stunt framing;
- superhero invulnerability;
- Curiosity Graph as magic answer machine;
- outfit/hair/backpack drift.

### CHAR-TIKO-v1
Lock:
- fictional axolotl-inspired silhouette;
- canonical external gill count/shape/placement;
- stable tail/limbs;
- warm large eyes;
- approved explorer accessory state.

Must test:
- `¿Y si…?` hypothesis;
- clue discovery;
- safe comic reaction;
- hop/turn/swim as appropriate;
- story-selected friend interaction.

Reject gill/anatomy drift, fish/reptile transformation, missing tail, graphic regeneration imagery or humiliation comedy.

### CHAR-NOVA-v2
Lock rounded pearl-white child-safe robot, dark face screen, stable cyan expression grammar, luminous core and stable blue accents.

Must test uncertainty/Confidence Meter, verification/modeling, hand/object interaction, locomotion and story-selected collaboration.

Reject armor/weapons, uncanny human skin, screen/core relocation, extra limbs or omniscient staging.

### CHAR-MARINA-v2
Lock unmistakable sea-turtle silhouette, stable blue/teal shell, coherent flippers, kind intelligent eyes and approved navigator detail.

Must test swim views, calm observation, focused courage in marine settings, Ocean Memory Map and story-selected marine collaboration.

Reject tortoise drift, shell/flipper drift, frail/sleepy stereotype, horror framing or terrestrial volcano/geology/history authority. Marina's expertise remains ocean/marine only.

### CHAR-RIN-v1
Lock red-panda identity, facial mask, tail pattern, explorer/nature details and warm observant eyes.

Must test habitat observation, empathy → practical action, trail movement, ecosystem interaction and story-selected friendship beats.

Reject fox/raccoon drift, tail changes, unsafe wildlife handling or guilt-heavy eco-doom.

### CHAR-ARQUEO-v1
Lock young archaeology/time-detective identity, hat/backpack, approved field tools and respectful evidence-focused styling.

Must test clue comparison, field notes, competing hypotheses, careful site observation and story-selected mystery collaboration.

Reject looting, trap/stunt framing, adult-adventurer drift or archaeology/paleontology confusion.

### CHAR-PEPE-v2
Lock urraca/magpie silhouette, black-white plumage, blue-iridescent wing accents, beak/crest/eyes and compact agile anatomy.

Must test head tilt/perch, short flight/landing, Pista Brillante, quiz/reaction pacing and story-selected collaboration.

Reject penguin/crow drift, wings-as-human-arms, extra wings/feet, manic default expression or random shiny-object distraction.

## Canonical voice requirement

Each speaking mascot has one active versioned voice identity:
- `VOICE-LUMI-v1`
- `VOICE-TIKO-v1`
- `VOICE-NOVA-v1`
- `VOICE-MARINA-v1`
- `VOICE-RIN-v1`
- `VOICE-ARQUEO-v1`
- `VOICE-PEPE-v1`

Exact provider voice bindings remain `audition_required` until real samples pass `VOICE_CANON_AND_QA_v0_1`.

Provider/TTS engine is an implementation detail; changing engine without identity regression QA is a voice change.

## Golden Character Test v3

Every exact character version runs:
1. neutral full body;
2. happy/curious close-up;
3. uncertain/worried close-up;
4. signature expression;
5. approved prop/mechanic interaction;
6. simple movement 5–8 sec;
7. canonical voice sample when speaking;
8. difficult pronunciation/emotion sample when speaking;
9. complex-environment identity test;
10. 16:9 composition;
11. 9:16 adaptation.

### Relationship benchmark
Run only for actual scheduled relationships:
- one current story-selected pair/collaboration;
- one specialist pair if the character is likely to use one;
- one 3-character mutual-quest establishing shot only when near-term long-form requires it.

Do not require a seven-character ensemble generation for individual approval.

## Relationship/crossover QA

Record:
- exact cast versions;
- canonical voice identities when speaking;
- scale consistency;
- duplicate/merged-character defects;
- accessory swapping;
- eye-line coherence;
- emotional-role readability;
- domain-authority correctness;
- protagonist hierarchy when Lumi leads;
- composition crowding;
- first-pass acceptance.

## Image-model benchmark policy

Reference asset generation is model-routed, not tied permanently to one image model.

Benchmark matched prompts/reference inputs across eligible image models such as the Google/Gemini Nano Banana family, GPT Image, Seedream, FLUX and other live candidates.

Character-asset routing requires identity >=95 before cost can influence selection. Optimize `credits_per_accepted_image`, not nominal request price.

## Video-model benchmark policy

For Golden motion/dialogue tests compare matched exact character/version, shot brief, duration and quality gate. Optimize `credits_per_accepted_second` with identity, continuity and first-pass acceptance hard gates.

## Asset lifecycle and storage

- original PNG master keeps checksum/provenance;
- Google Drive may be used as temporary staging/recovery;
- reviewed reference masters should move through binary-capable Git/LFS or canonical object storage;
- repository previews never replace original masters;
- raster-in-SVG is not true vector;
- native 4K != upscaled 4K.

## Canonical references

- identity/domain/TOV: `MASCOT_BIBLE_v0_4.md`
- canonical voices: `VOICE_CANON_AND_QA_v0_1.md`
- machine voice bindings: `data/catalog/voice-binding-registry-v0_1.json`
- friendships/crossovers: `RELATIONSHIP_AND_CROSSOVER_BIBLE_v0_4.md`
- relationship graph: `data/catalog/character-relationship-graph-v0_3.json`
- asset generation prompts: `HIGGSFIELD_ASSET_GENERATION_PROMPTS_v0_2.md`
- model selection: `docs/analytics/HIGGSFIELD_IMAGE_VIDEO_VOICE_MODEL_SELECTION_v0_1.md`
- binary storage/transfer: `docs/operations/VISUAL_ASSET_BINARY_TRANSFER_AND_STORAGE_v0_1.md`
- emotional rules: `docs/agents/EMOTIONAL_ADVENTURE_ENGINE_v0_1.md`
- scene/asset schema: `docs/operations/SCENE_AND_ASSET_CONTRACT_v0_1.md`
- operating workflow: `docs/operations/MASCOT_PRODUCTION_OS_v0_1.md`