BOSTON, JUNE 2, 2026 — anima, active
Anima
◐ ACT 1 SHIPPED · ACT 2 IN FLIGHT
─ OPENER ─
I drew a pencil sketch of myself on the iPad: me at 33, slightly stooped, stylus in the right hand. Then I taught Gemini Nano Banana 2 to redraw the same guy 220 times in a row, in 16:9, on cream paper, with the same facial scruff in every frame. The model has a list of ways it can fail me now. They have numbers. SF02 is “wrong jaw.”
It could’ve been Midjourney. Midjourney would’ve been faster. Midjourney would’ve also been gauzy, vertical, photo-real, and someone else’s. The pipeline is a pencil test on purpose: construction lines visible, paper grain, hole-punch marks down the left edge. It’s an animator’s idiom, not a content-farm idiom. The discipline is the proof that a human still owns the timing.
I block the motion in Procreate Dreams. The agents fill the frames. Seedance 2.0 interpolates between the approved keys. A critic agent flags identity drift before the render burns budget. The pipeline is called anima. Act 1 plays clean. Act 2 is on the board.
─ INVESTIGATION BOARD ─
A-1 Pipeline rev 3: final cleanup chain
220-frame cycle locks. Move to case-study hero media (not portfolio hero).
Rev 3 is consistent enough to ship; portfolio hero stays with the 94-frame loop per hero-spec §7.5.
Render time per frame (seconds)
the alpha channel is the rub. vp9 with alpha_mode=1, not h264.
tried that yesterday, re-running with the corrected flag.
— anonymized, static svg —
Production board: first cleaned cycle
A1-014
Land 146-frame raw cycle with consistent line
─ METHODS ─
| TASK | AGENT / TOOL | MODEL / COST |
|---|---|---|
| keyframe stills | Gemini Nano Banana 2 | ~$0.04/frame |
| motion interpolation | Seedance 2.0 | ~$0.40/clip (Fast tier) |
| orchestration | Code Brain | Claude Sonnet 4.6 (HybridRouter) |
| vision critic (T2) | Gemini 3.1 Pro via Anti-Gravity CLI | $0 incremental (subscription) |
| multi-CLI critic (T3) | Codex CLI + Anti-Gravity CLI in parallel | $0 incremental (subscriptions absorb) |
| planner | Opus 4.7 (Maya persona) + Sonnet 4.6 adversarial | per-token billing |
─ 4Q ─
A-1.Q1 What is this?
anima is a 10-phase pipeline for shipping 2D animated stories made by a human and a fleet of named agents working together. A brief becomes a plan. A plan becomes an animatic. An animatic constrains motion. Motion gets cleaned to the aesthetic. The fleet runs the volume work (a line producer, a character designer, a scriptwriter, a storyboard artist, a frame generator, a script-supervisor critic, a museum writer) while the human owns timing, taste, and the decision to ship. The Pencil Test short is the first reference implementation; anima is the system.
A-1.Q2 Why this approach?
The decision wasn't which model to use. The decision was which working method: a studio budget and a four-year project timeline (out of reach), AI-only generation (fast but someone else's idiom), or a human-author plus AI-fleet partnership. The third one is anima: production-company speed at solo-creator cost, without losing four years of life to a single piece. The model layer is replaceable; the architecture and the human role are not.
A-1.Q3 What would break?
Three named failure modes the architecture guards against, not three encoder bugs. Correlated blind spots when the orchestrator and the vision critic share a model family: 'valid output' and 'acceptable frame' silently align, and bad frames slip through. anima pairs a Sonnet orchestrator with a Gemini vision critic at the busiest checkpoint by construction. Local-optimization drift when every phase ships 'better' output that no longer matches the approved brief. Phase 0's planner emits an immutable acceptance_criteria.json; every downstream critic cites criteria IDs when it blocks. Cheap-judge failures with documented base rates: 58% sycophancy, up to 90% self-preference bias, length and position bias. T3 runs three peers from three vendors (Codex CLI + Anti-Gravity CLI + Claude SDK), with a separate Opus chairman that never grades its own work.
A-1.Q4 What did I learn?
Validators cannot recover taste that was absent at generation time. The pipeline is the artifact. The character is the test the pipeline has to pass.