AI Generation Lanes — Diagrams

Mermaid companions to ../technical/ai-generation-strategy.md (council-approved v2, 2026-07-31). Mermaid inline so they render natively on GitHub, per the diagrams toolchain.

1. Lanes → provider → indemnity coverage

Who generates what, and whose legal cover it sits under (decision log 2026-07-31: Claude PRIMARY for cloud text as a documented exception; indemnity-first governs media lanes).

flowchart LR
    subgraph Lanes
        L1[Flashcard diagrams]
        L2[Characters / mascots]
        L3[Study songs - BLOCKED]
        L4[Ambient / white noise]
        L5[Narration TTS]
        L6[Study Cast episodes]
        L7[Project workspaces chat]
        L8[Core inference / PageAgent]
        L10[Resume / letters / outreach]
        L11[Mock interview voice loop]
        L12[Grounded research]
        L13[Ingestion / embeddings]
    end

    subgraph Providers
        P0[Procedural renderer]
        PG[Google paid APIs<br/>Nano Banana 2 / Gemini TTS / Lyria 3]
        PA[Claude Sonnet 5<br/>Anthropic commercial terms]
        PGF[Gemini 3.6 Flash]
        PC[CC0 field recordings]
    end

    subgraph Coverage
        C0[$0 procedural - no model output]
        CG[Google indemnity<br/>paid APIs only]
        CA[Anthropic-terms indemnity<br/>documented exception, verify at onboarding]
        CC[CC0 - public domain, per-file audited]
        CU[UNVERIFIED - songs-lane blocker]
    end

    L1 --> P0 --> C0
    L2 --> PG --> CG
    L3 --> PG
    PG -. Lyria indemnity-list status .-> CU
    L4 --> PC --> CC
    L5 --> PG
    L6 --> PG
    L7 --> PA --> CA
    L8 --> PA
    L8 --> PGF --> CG
    L10 --> PGF
    L10 --> PA
    L11 --> PG
    L12 --> PGF
    L13 --> PGF

Not shown: video (lane 9, not offered v1) and grok CLI (internal docs/tutorials only — its output never ships in product). Every lane stamps provenance metadata (WING-298).

2. Cap → meter → top-up purchase flow (WING-299 / WING-300)

flowchart TD
    A[User requests a generation] --> B{Lane cap reached?<br/>WING-299}
    B -- no --> C[Generate]
    C --> D{Succeeded?}
    D -- yes --> E[Deliver artifact + provenance stamp<br/>WING-298; captions if audio]
    E --> F[Increment visible per-lane meter]
    D -- no --> G[No quota consumed - retry free]
    B -- yes --> H[Show meter + offer top-up pack<br/>WING-300, ~50% margin]
    H -- buys pack --> I[Pack credits added to lane] --> C
    H -- declines --> J[Lane paused until next month<br/>never silent throttling]

3. Cost per user by tier (generation only, $/month)

The stacked view reads best as a table (figures from the per-user table in the technical doc):

Lane	Light	Moderate	Heavy uncapped	Heavy capped
Project chat	1.00	3.75	15.00	7.50
Mock interviews	0	2.00	10.00	6.00
Narration	0.34	1.35	6.75	4.05
Study Cast	0.35	1.40	7.00	3.50
Resume/docs	0.30	2.00	6.00	3.60
PageAgent	0.50	2.25	6.00	4.00
Songs	0.15	0.60	3.00	1.50
Images	0	0.34	1.68	1.68
Grounded research	0.10	0.50	2.00	1.00
Ambient	~0	0.05	0.25	0.25
Total	≈ 2.75	≈ 14.25	≈ 57.70	≈ 33.10

pie showData title Heavy-capped user (~$33.10/mo) - where the money goes
    "Project chat" : 7.50
    "Mock interviews" : 6.00
    "Narration" : 4.05
    "PageAgent" : 4.00
    "Resume/docs" : 3.60
    "Study Cast" : 3.50
    "Songs" : 1.50
    "Images" : 1.68
    "Grounded research" : 1.00
    "Ambient" : 0.25

Chat, interviews, and PageAgent dominate — which is why caps cover them, not just audio.