%%{init: {"theme": "base", "flowchart": {"htmlLabels": false}, "themeVariables": {"primaryColor": "#eef8f6", "primaryTextColor": "#17211f", "primaryBorderColor": "#0f766e", "secondaryColor": "#fff8e6", "tertiaryColor": "#ffffff", "lineColor": "#17211f", "fontFamily": "Inter, ui-sans-serif, system-ui, sans-serif", "fontSize": "14px"}}}%% flowchart TD Mic["Microphone audio"] --> ASR["Cohere Transcribe
03-2026
16 kHz en"] Typed["Typed prompt"] --> Prompt["Prompt text"] ASR --> Prompt Prompt --> Mellum["Mellum2 Thinking 12B
via Modal Cloud
structured JSON schemas"] Mellum --> Parse["PromptDocument
extract + repair/verify"] Parse --> Canon["BGE enum canonicalizer
bge-small-en-v1.5
threshold 0.62"] Canon --> Locks["Pre-run locks
thresholds
IQA >= 0.55
Alignment >= 0.25
Human IQA >= 0.70"] Locks --> Coord["LHS coordinate
Thompson + GP
Bayesian score"] Coord --> Seeds["Bonsai Ternary 4B
5-seed batch
steps=4, 512x512"] Seeds --> IQA{"JoyQuality SigLIP2
score >= 0.55?"} IQA -- "fail" --> Persist["Persist candidate + failure evidence"] IQA -- "pass" --> VLM{"MiniCPM-V-4.6
alignment >= 0.25?"} VLM -- "fail" --> Persist VLM -- "pass" --> Impact{"TRIBE v2 lite-qv
disabled by default"} Impact -- "disabled or pass" --> Curated["Curated catalog
fragile: 1 promoted
viable: 2 promoted
strong: >= 3 promoted"] Curated --> Feedback["Accept / reject / shred feedback"] Feedback --> Priors["Update priors
enum arms alpha/beta
enum-combo GP affinity"] Priors --> Persist Persist --> Stop{"Stop rule?"} Stop -- "Gradio runtime cap: 15 minutes" --> End["Stop run"] Stop -- "stall: fewer than 10 curated after 10 minutes" --> End Stop -- "backend/requested stop" --> End Stop -- "continue with updated priors" --> Coord classDef step fill:#eef8f6,stroke:#0f766e,color:#17211f classDef gate fill:#fff8e6,stroke:#a16207,color:#17211f class Mic,ASR,Typed,Prompt,Mellum,Parse,Canon,Locks,Coord,Seeds,Persist,Curated,Feedback,Priors,End step class IQA,VLM,Impact,Stop gate

Workflow Steps

  1. ASR / text input accepts typed prompts or microphone audio. The local ASR path uses CohereLabs/cohere-transcribe-03-2026, normalizes to 16 kHz, defaults to English punctuation, and decodes with max_new_tokens=256.
  2. Decomposition extracts objects, relations, constraints, and cinematography lanes from typed text or ASR output.
  3. Mellum2 Thinking 12B via Modal Cloud supplies the structured JSON reasoning path for prompt extraction, repair, and verification through an OpenAI-compatible endpoint using temperature=0.0, strict schema output, and max_completion_tokens=2048.
  4. Repair/Verify checks blocking issues, unresolved targets, and threshold readiness before generation is allowed.
  5. Canonicalization maps raw prompt values to project enums with BAAI/bge-small-en-v1.5, match_threshold=0.62, and LLM fallback for uncertain matches.
  6. LHS proposes coverage-oriented coordinate rows across unlocked enum arms while preserving locked prompt evidence.
  7. Thompson Sampling/GP ranks sampled arms with Bayesian feedback state, GP-style coordinate scoring, and compatibility priors as evaluation evidence accumulates.
  8. Quick generation uses prism-ml/bonsai-image-ternary-4B-gemlite-2bit through Bonsai local or HTTP adapters, with the fast preview path defaulting to steps=4, 512x512, and seeds [7, 42, 156, 8888, 42069].
  9. IQA filters each 5-seed batch with fancyfeast/joyquality-siglip2-so400m-512-16-05k047vn against the configured quality cutoff.
  10. VLM alignment scores surviving images with openbmb/MiniCPM-V-4.6, max_new_tokens=128, and the compiled prompt/target manifest.
  11. TRIBE v2 uses Jessylg27/tribev2-lite-qv as an optional metacognitive impact adapter; it is disabled by default and kept downstream of IQA and VLM gates.
  12. Prior updates write evaluation and feedback evidence back into enum-arm alpha/beta priors and enum-combination GP affinity for the next coordinate.