Status: approved, with one amendment for Elikem's approval. Elikem approved
revision 6 on 2026-09-29 and the family opened the same day. He approved revision
14 the same day, which adds what sections 12 to 14 list: a field review of the newest techniques with
corrected facts (12), a role-by-role audit of how top studios crew a film or
series (13), and the Blender feature adoption map with a full studio workspace
inside Blender (14): 31 tasks, 2 epics and 56 task amendments. Author: Claude (claude-mac-worker), 2026-09-29, from
Elikem's brief the same day. Board family: lilith-v1-5. Revision 20, approved
the same day, added section 15: what Arcane, The Clone Wars, The Expanse, Altered
Carbon, Altered Carbon: Resleeved, Cyberpunk: Edgerunners, Pixar and DreamWorks add
(7 tasks), and Elikem's correction of the audience to a mature, Rockstar-level
standard. Revision 24, approved the same day, added section 16: a second,
independent role audit (15 tasks). Revision 28, approved the same day, added section
17: four targeted passes (the pilot walk-through, the animation control catalogue,
the autonomy readiness matrix and the short's pre-production pack; 3 tasks).
Revision 32, approved the same day, added section 18: a feature-gap review of what
surrounds the films (release windows, short-form versions, editions, incentives and
sales, controlled previews, background voices, ensemble improvisation and the studio
from the personal agent; 9 tasks and Decision 20). Revision 36, approved the same day, added section 19: Elikem's language decision (English for a good while), which revises Decision 12 and defers dubbing. Revision 45, approved on 2026-10-05, added section 20: performance mining and the open face and body stack, with 12 tasks (R21, N21–N28, Q14, C18, C19) and Decision 21, which Elikem decided as "build own". The amendment now awaiting his approval is section 21: text-to-motion on cleared data (2026-10-06), which gives N12 the route Decision 21 left it without. It adds no Lilith task (the four new tasks are in the Isis family: T.01.21 and T.12.08–T.12.10); N12, N14, R05 and R21 were amended on the board before this revision, and the evidence document (revision 13) already carries them. Release name in code:
LILITH_STUDIO_RELEASE in libs/oshun/navigation/src/release-scope.ts.
Two independent read-only reviews were folded in before filing, both through
codex exec on the ChatGPT subscription with web search: GPT-6 Sol at xhigh
effort (35 findings) and GPT-6 Astra at xhigh effort (35 findings, at Elikem's
request). They found acceptance-level dependency cycles, milestone exits no
task delivered, overlaps with the V1.4 and content-workbenches plans, a missing
per-revision publication authorisation, and legal overstatements (UK s9(3)
proposal stated as decided, the Academy's writing-category rule generalised,
C2PA treated as the whole of EU Art. 50 disclosure, folklore assumed public
domain although Ghana's Copyright Act 2005 protects it in perpetuity). All 70
findings are addressed in this revision; 21 tasks were added (G11, S16, S17,
A12, A13, B13, N13, H15, Q11, Q12, R13–R19, Y10, D11, K13, M09).
0. The brief (Elikem, 2026-09-29, verbatim)#
"is there an oshun v1.5? which one is lilith tv"
"i would like to add a v1.5 that gives me all the tools and agents and workspace features to create compelling animated movies and tv series using blender, that would enable me elikem as a solo builder to use the latest advancements in ai and ai agents to create compelling and high quality content. i want to make amazing films and shows and using ai not because of laziness so want to have all possible features and avenues for control and granulairty in all aspects of the film making process end to end while at the same time building agentic systems so powerful that they could technically produce this content on their own as well. including tools for world building, ideation, charcter building and everything else. lookup the absolute latest in ai in this area and agentic systems as well as what pople are doing with gpt 6 astra and claude opus 5.5, also study what we already have in this monorepo and then create prd and comprehensively cover everything requrired and sequence all required tasks in the eve task board as well."
"use gpt 6 astra xhigh for this codex review. you can let the current one complete and then do anotehr with astra"
"go with your recommendations for the open decisions"
"analyze the recent v1.5 and ensure it captures the best and latest industry techniques plus new effective techniques to produce the best results possible" (2026-09-29, after approval; answered by revision 14, section 12)
"for v1.5 go through all possible roles when teams make these things now in hollywood and the best studios around the world … ensure every single role is accounted for … and for every role and perspective ensure we have full features … the most granular creative control for me … and the best autonomous systems everywhere … in addition look at all the latest blender features that apply and how we could adopt them also consider building a comprehensive workspace within blender as well such that like everything should be controllable by me through the web app but at the same time … a level of jumping down to the details … with a very full features set of tools and views and workspaces in blender" (2026-09-29, abridged; answered by sections 13 and 14 and the O and U pillars)
"are there any roles, tasks, and such that would be involved in the making an arcane or star wars clone wars or altered carbon or the expanse or pixar or dreamworks or altered carbon resleeved or cyberpunk edgerunners in granular / deeply technical detail" (2026-09-29; answered by section 15)
"only correction is not 12+ that must have been a mistake. everything in oshun synthetics is kinda like rock star games level of audience so please rework all things that were made based on this mistake" (2026-09-29; Decision 10 corrected, R06, R18, R19, Y01, Y02, Q06 and K07 reworked; section 15)
The role request above, sent again the same evening, answered with "Hunt for missing roles": a second independent audit against more productions (2026-09-29; answered by section 16)
"what narrower more targeted pass do u have in mind" — Elikem then chose all four offered: the pilot walk-through, a control-depth catalogue, autonomy readiness and the short-film prep list (2026-09-29; answered by section 17)
"review the recently added v1.5 prd and tasks. do you see any feature missing", then "proceed" after the findings were listed (2026-09-29; answered by section 18)
"language will be english for a good while before we explore others", then "defer them too" for translated subtitles and the M3 dubs (2026-09-29; answered by section 19)
The answer given before the brief: there was no V1.5, and "Lilith TV" is the working name of V1.4's own streaming platform. The last line settles every decision in section 9 as recommended.
1. Summary#
V1.5 gives Elikem a film and series studio: one workspace, built around Blender, in which a single person can take an idea to a finished animated film or a season of a series, with a crew of AI agents that can do any job in the pipeline and, when he chooses, the whole pipeline on their own.
The studio has two promises. They are equal, and neither may be traded for the other:
- Total control. Every decision a film makes, from a world's history to a character's eyelid on frame 1,042, is visible, editable and lockable by Elikem, at the granularity a professional in that department would expect. Agents propose; he disposes. Nothing an agent does is hidden, final or irreversible.
- Real autonomy. The same pipeline, driven by the agent crew, can run a department, a shot, a scene, an episode or a whole film on its own, to a measured quality bar, within a budget, and report honestly what it did and where it fell short.
The two promises share one mechanism: one set of operations, two kinds of client. Every action in the studio (write a scene, move a camera, retime a key, relight a shot, trim a cut) is a typed, versioned operation. Elikem calls it from the interface or from Blender itself; an agent calls it through the same API. The autonomy dial per department (section 5) decides only who proposes and who approves.
Blender is the ground truth of every frame. Cameras, blocking, performance, lighting and timing live in real 3D scenes. Generative models are used where they are strong (ideation, concept art, reference, texture, look, restyling a rendered shot, sound, music, voice) and are always measured against the 3D truth rather than replacing it. That is what keeps characters on model across a season, cuts continuous across shots (the weakest dimension across the workflows DirectorBench evaluated, section 3.2), and every decision editable after the fact.
V1.5 is built on what exists rather than beside it. The Isis film track
(F.00–F.20, 130 items, family isis-chroma-runpod) already specifies the
Blender craft primitives: Fountain screenplay, breakdown, boards, previs,
cinematography, sets, casting, performance, lighting, simulation, render,
compositing, generative finish, OpenTimelineIO editorial, sound, grade,
delivery and a film benchmark. The Isis 3D track (T) builds assets, rigs and
motion. The content workbenches' creative studio (content-workbenches:CR)
builds the character forge, ideation board, series bible, model sheets,
storyboards and script studio. V1.4 (lilith-v1-4) builds the universe bible,
canon registry, writers' room, voice and music lanes, and the home platform
where the films will premiere. V1.5 adds what none of them has: long-form
development (world building, story, series), the director's workspace that
unifies every department, the agent crew that can run it all, the taste and
quality system that tells good from merely finished, and the rights record that
makes the films Elikem's.
The release ladder is measured by films, not features: the 40-second pilot the film track already defined (M1), a festival-grade short (M2), a series season premiering on Lilith TV (M3), then a feature and the studio's opening to others (M4 horizon).
2. What exists today (code and board, 2026-09-29)#
An Explore survey of the repository on 2026-09-29 (read-only, excerpts, spec
files counted) and a read-only export of the board. The film studio is
specified in detail and very little of it is built. The Isis tracker moved
into Eve at d092109ae6; of its 130 film items (F.00–F.20) only F.00.02,
F.00.03 and F.03.01 are shipped, F.03.05 (Kitsu sync) is parked, and F.00.01
and F.02.05 have commits but are still ready.
2.1 Real and tested — build on these#
| Area | Where | What it does | Evidence |
|---|---|---|---|
| Headless Blender | libs/isis/3d-post-pipeline/src/blender/ (runner.ts, jobs.ts, job-tools.ts, live-session.ts, ops/python.ts, round-trip.ts) |
Hardened runner (--factory-startup, --disable-autoexec, RSS cap, no network); jobs for animation bake, simulation cache, render and mixdown with submit/poll/cancel/resume (F.02.05); a live Blender Lab MCP session; oshun_blender_ops 1.6.0 with 19 versioned operations (inspect, transform, mesh, UV, material, rig, keyframe, sample, import/export, capture, checkpoint, scene). Output kinds are only glb/png/blend/json — no EXR, video or audio yet. |
T.20 8/8 shipped; 9 specs in the folder |
| GPU worker | apps/isis/gpu-worker (16 executors; Dockerfile.dcc-cad pins Blender 5.2.1 by sha256) |
Real Blender scripts: retarget, render motion frames, mesh clean-up, QuadriFlow remesh, bake, delivery export | 16 specs that run real Blender |
| Scene assembly | libs/isis/3d-scene-assembly |
Semantic layout, keyframes, orbit/dolly/hold cameras on a frame grid, shot checks, natural-language scene edits, scene renders; the F.03.01 production manifest | 39 specs; apps/isis/web /scenes/:sceneId |
| Rigging and motion | libs/isis/universal-rigging |
Anatomy classification, skeleton mapping, deformation diagnostics, bounded repair, capability manifest (T.11.02–06); parametric gaits, narrative motion planner, motion catalogue (T.12.03/05/06) | 55 spec files |
| 3D workspace | apps/isis/web |
Scene tree, inspector, locks, undo, rig and motion panels, workflow graph editor, jobs, outputs, RunPod and provider pages | 157 unit specs, 30 e2e |
| Generation service | apps/isis/generation-api |
Jobs, RunPod routing, model cache, cost ledger and kill switch, output store, OpenRouter and Meshy lanes; services/three-d-director with 13 typed tools and role outputs; controller admission (libs/isis/workflows/.../controller-admission.ts) |
241 specs; live OpenRouter proofs |
| Image and video lanes | libs/isis/workflows (Chroma, Z-Image, HiDream-O1, Qwen-Image, SAM 3, BiRefNet; Wan 2.2 t2v/i2v, VACE edit, Animate-2, LTX AV, SCAIL-2, SeedVR2, FlashVSR, GIMM-VFI) |
Catalogued, validated ComfyUI workflows on RunPod | 90 specs; 8 live clips (V.11) |
| Ideation and story | libs/hathor/ideation, apps/hathor/narrative-api, apps/hathor/workbench |
Premises, what-ifs, tournaments; logline → synopsis → treatment → beats ladder with lineage; a writers'-room protocol; one-pager, act template, A/B/C episode grid, series bible, casting sheet with likeness consent, table read. The model tier is an unconfigured seam (503 until a provider is wired). | 60 + 7 specs; workbench /ideation |
| Worlds | libs/hathor/world, apps/hathor/world-api, apps/hathor/workbench |
World editor: worlds, cultures, factions, timeline, relationship web, story graphs (CRUD) | 3 + 3 + 62 specs |
| Canon | libs/yemaya/canon-graph, libs/yemaya/canon-enforcement |
Typed property graph, retrieval, consistency checks, write-back | 8 + 2 specs |
| Screenplay formats | libs/yemaya/pre-production (parseFountain, export to Fountain/HTML/FDX/RTF, and PDF page preparation — 55-line page data, no PDF typography); libs/hathor/lore-compiler (screenplay compiler to Fountain/FDX); libs/hathor/pre-production/src/scriptwriting |
Two Fountain parsers and FDX export | small spec counts |
| Voice, alignment, loudness | apps/oshun/bff/src/generation/ (phase 182: local-kokoro-voice-provider.ts, comfy-voice-provider.ts Chatterbox, comfy-music-provider.ts ACE-Step 1.5 Turbo, human-video-loudness.ts, human-video-mux.ts, caption-dub-executor); tools/human-video/forced-align.py (CTC, ≤81 ms onset error on its fixture); libs/euterpe/master (BS.1770) |
TTS, forced alignment, EBU R128 loudness, muxing | 294 specs in the BFF generation tree |
| Visemes | libs/psyche/avatar-lipsync, viseme-generator |
Phonemes to visemes with coarticulation | narrow |
| Agent lanes | tools/eve-codex-agent.mjs (Codex exec, never full access); tools/workbench-mcp/{server,eve-server}.mjs |
The Codex worker pattern and MCP servers to copy; no film tools yet | specs |
| Orchestration | libs/oshun/creative-orchestrator (brief → plan → route → critique/revise, fail loud); libs/yemaya/orchestration pipeline runner (human approval gates, checkpoints) |
Candidate foundations for the crew runtime | 13 + few specs |
| Other Blender bridges | libs/bellona/blender-agent (50 typed actions, transactions), apps/bellona/remote-host (8,488-line adapter), libs/bellona/mcp-gateway (real MCP stdio server) |
A second, remote-control Blender stack; ADR-0008 and T.21.07 audit the overlap | 36 + 69 + 25 specs |
| Colour | libs/oshun/color-science |
Partial: emits OCIO v2 colour-space and look metadata without the transforms; almost unused. F.14.03 owns the working OCIO pipeline. | 10 specs |
| Distribution | libs/shared/live-media |
Player, publisher, MediaMTX edge, HLS/DASH packaging | 57 specs |
Owner decisions already recorded for the film track (F.00.03, receipt in
docs/audits/isis-film-reuse-audit/2026-09-24/): the pilot film is 40 seconds,
8 shots, 2 characters, one interior set, 1080p at 24 fps, safe for work only (for V1.5, Elikem superseded this on 2026-09-29: the pilot follows the mature standard);
colour uses the AgX view with the OCIO configuration bundled in Blender 5.2.1;
the Codex lane may use at most 20% of each usage window; the cheap-model
bake-off is capped at 5 USD; no Kitsu sync.
2.2 Traps — look real, are not (never wire these as if they worked)#
- The generation API's "pack" route families (
character-turnaround-sheet,character-expression-sheet,character-costume-variant-pack,storyboard-frame-set,animatic-shot-set,cinematic-shot-sequence,audio-sfx-pack,voice-line-pack,music-cue-pack, and theirapps/isis/web/src/pages/*-ui-utils.tshelpers) are contract, budget and quality-gate wrappers whose workflow ids no catalog workflow implements. libs/hathor/pre-production/src/planning: about 244 "director", "drone" and script-supervision modules that compute readiness scores and never perform the operation.- About 1,588 files under
apps/oshun/web/src/components/studio: admin panels mostly on fixtures and planners; their Blender panels only validate or describe. - Isis
video-edit-studio,relight-studio,motion-camera-control,foley-studio,av-narrative-studio,portrait-studio, andaudio-generation's analysis: labelled catalogue or fabricated by the repository's own reuse audit, or the same shape unaudited.voice-cloningandface-synthesisare planners;3d-video-diffusionand3dgs-relighting-pbrare catalogues. libs/isis/universal-rigging'sml-skeleton-prediction.tsandhumanoid-animation-library.ts(a catalogue of clips it does not hold).libs/yemaya/dailies-reviewand the Yemayafilm.tspipeline config (planners);libs/oshun/render-farm(scheduling only).libs/oshun/c2pa-exportis an honest seam:{configured:false}without an X.509 signing certificate; the BFF's audio "C2PA" signer is an in-house Ed25519 manifest, not embedded C2PA.- About 600
libs/neith/*Rust crates re-implement Blender, the sequencer (OTIO, EDL, conform), grading, delivery (DCP, IMF), audio and 2D/NPR animation. They are real and tested but wired to nothing; the ownership map (G02) gives each one a reuse-or-ignore verdict before any task imports it. - Two Blender pins: 5.2.1 in
Dockerfile.dcc-cadand a stale 4.0.2 inDockerfile.blender-renderthat tool admission refuses (B01 retires it).
2.3 Absent — what V1.5 must build#
A screen for the film itself (script, breakdown, shot tracker, boards, cut,
dailies); Fountain to breakdown to a typed shot list tied to the manifest; any
series or season level in production (it exists only inside Hathor ideation);
Grease Pencil boards and the animatic; an OpenTimelineIO or sequencer bridge in
TypeScript or Python; EXR, Cryptomatte, the compositor and OCIO wiring in the
Isis runner; camera, light, simulation, Geometry Nodes and render-pass
operations in the versioned op set; a GPU Blender render farm; lip sync driving
shape keys and any facial performance; text-to-motion, ML auto-rigging and
self-hosted neural 3D; the film-skills MCP, the Codex director
(tools/isis-codex-director.mjs), the token governor and the film benchmark;
film sound spotting and mixing. On the machines: Blender 5.2.1 is installed on
the Mac (not on PATH); the local Python environments have no diffusers,
transformers, kokoro, OTIO, OCIO or loudness packages; Kokoro and Chatterbox
weights are cached; ffmpeg 8.1.1 and codex-cli 0.157.1 are present.
2.4 The board around V1.5#
| Family | Open | What V1.5 takes from it |
|---|---|---|
isis-chroma-runpod |
225 ready, 27 parked | F.01–F.20 film craft (the Blender primitives V1.5's rooms and agents sit on) and T asset/rig/motion work. V1.5 tasks depend on the F items they integrate. |
content-workbenches |
323 ready | CR: character bible, ideation board, divergence and novelty, series bible and beat board, model sheets, on-model and style checkers, storyboard/animatic, script studio with table read, performance direction, preference lab, likeness screen, revision impact. |
lilith-v1-4 |
265 ready | Universe bible and canon (its U pillar), writers' room and promos (W), casting and ensembles (C), voice and music gates (V, M), provenance and rights registry (lilith-v1-4:F06, lilith-v1-4:S03), audience screenings (lilith-v1-4:W12), the crew runtime (lilith-v1-4:H01) and the films-and-series channel (lilith-v1-4:Z02) where V1.5 titles premiere. |
phase-182-human-video |
6 ready, 24 parked | The shipped voice, alignment, loudness and mux stack in the BFF. |
3. The field in September 2026#
Sources were read on 2026-09-29. P marks a primary source (the vendor's own page, the paper's abstract, or a live API call); S marks a secondary source (news, a blog, an aggregator) that a task must re-verify before binding a decision to it (Q10 keeps this list current).
3.1 Agent models: what people are doing with GPT-6 Astra and Claude Opus 5.5#
| Fact | Source | Consequence for V1.5 |
|---|---|---|
Claude Opus 5.5 shipped 2026-09-22 (claude-opus-5-5): $4/$20 per million input/output tokens, cache reads $0.20, fast mode $8/$40; Terminal-Bench 4.0 66.4% (Opus 5: 52.3%), OSWorld 2.1 81.8%, GDPval-AA 1846 Elo; output over 30% faster than Opus 5; "much less likely than recent models to take hard-to-reverse actions". |
P: anthropic.com/claude-opus-5-5 | A strong candidate for the long-horizon roles (showrunner, story editor, pipeline operator). Opt-in profile with a cap (Decision 3); the bake-off (H04) decides per role on measured results. |
| People make video with Opus 5.5 as programs: it writes a storyboard document, then per-chapter painter code (p5.js with a watercolour library), renders ~3,760 frames in headless Chrome and assembles them with ffmpeg; every frame is a pure function of its timestamp. A 157-second music video, a 280 KB demoscene intro after about 5 hours and 697M tokens, an animated music video in 45 minutes for about $200. It uses about 1.6x the output tokens of Opus 5 on such tasks. | S (anecdotal: no primary record of these costs, token multiples or times, so nothing plans on them): OrcaRouter blog; 80aj.com | Code-as-film is a real lane for motion graphics, title sequences, stylised inserts and procedural effects. The studio's operation set treats a deterministic program as a legitimate shot source (Y06), with the same review and lineage as a render. |
GPT-6 Astra shipped 2026-09-03. It operates Blender itself: writes real bpy, renders frames, inspects them and edits until the geometry matches the brief; it also drives GUIs by computer use. Reported $10/$50 per million tokens; OSWorld 2.0 72.6 and BenchCAD 95.9 are secondary figures. No published Blender success rate; weaker on organic forms than on geometric ones. Astra takes no video or audio input, and its knowledge ends before Blender 5.2 LTS. |
P: OpenAI's practitioner account (developers.openai.com, 2026-09-04) shows Astra building and revising an editable Blender scene and exporting it to Unreal; it gives no Blender success rate and no evidence about character animation or film direction. S: neural4d, atoms.dev | Astra is the reference for the Blender operator role, reachable only through Codex on the ChatGPT subscription (CLAUDE.md), through the lane F.01.08 builds. Dailies review must route to a video-input model. Version-matched Blender API docs are served to every controller (F.02.02). |
| What people build with Astra: a house in Blender exported to a walkable Unreal level; an overnight reconstruction of San Francisco's Palace of Fine Arts from archival photos; a cityscape over a week; a museum scene with venue layout, cast plan, shot list and blocking for a film shoot (Higgsfield); an Ableton arrangement; playable games in about 90 minutes. | S: atoms.dev, neural4d | Long unattended runs on a single scene are practical. The studio needs overnight runs with checkpoints, budgets and a morning review (H07, S05). |
| A filmmaker remade an AI short in about a day and a half (a week in 2025) with Astra in the Codex desktop app driving Runway through MCP. The setup: a "studio skill" folder (instructions, references, cast and location assets, shot lists, stills, prior output) and a one-file HTML dashboard over the project (shots with prompts, boards, references, locations, session logs). Lessons: driving Premiere by computer use was slow and fought the human's manual edits, so manual editing was faster; generation and file latency created dead time; direct API or MCP integrations beat GUI automation; large asset archives need a dashboard. | S: MindStudio | Validates three V1.5 choices: typed operations over GUI automation (principle 2), the studio workspace as the dashboard over one production graph (S pillar), and an editorial room where the human edits directly while agents propose (K pillar). |
| Anthropic's own guidance on long-running agents: checkpoints, agent teams, context compaction, generator–evaluator loops, subjective-quality rubrics, persistent artifacts, and less scaffolding as models improve. | S: AI Engineer talk (Anthropic applied AI) | The crew runtime (H) is built on these: persistent production records instead of chat memory, a generator and an evaluator per department, and rubric-scored critique. |
| Newer controllers since this PRD was first written: GPT-6 Sol and Luna (2026-09-22; OpenAI reported an image-encoding fix on 2026-09-25 that changes visual results in the API and Codex), Claude Sonnet 5.5 (2026-09-28), Gemini 3.8 Flash (GA 2026-09-02; 3.8 Live on 2026-09-15), and open models Kimi K3, GLM-5.3-Flash and DeepSeek-V4.1-Flash (2026-09-10, with native vision). | P: OpenAI API changelog; anthropic.com/claude-sonnet-5-5; Gemini API changelog; DeepSeek release log; model cards | Candidates for the per-role bake-off (H04), which now records each model's exact build and date; an open model is not the same as a cheap one to run. |
MCP specification 2026-07-28, the largest revision since launch: stateless requests (no initialize handshake; server/discover), long-running work moved to the official tasks extension (tasks/get polling, tasks/update), Multi Round-Trip Requests (input_required) replacing server-initiated elicitation and sampling, OpenTelemetry traceparent in _meta, deterministic tools/list order for prompt caching, and Roots, Sampling and Logging deprecated. |
P: modelcontextprotocol.io/specification/2026-07-28/changelog | The studio MCP server (A04) targets it: renders and bakes return task handles, a call that needs Elikem's decision returns input_required for the propose-mode protocol, and traces join across calls (H11). |
Agent Skills (SKILL.md, an open standard Anthropic published in December 2025) is loaded by Claude Code, the Codex CLI, Gemini CLI and most coding agents; a skill is a folder of instructions and files loaded progressively. |
P: agentskills.io, github.com/agentskills/agentskills; S: adoption write-ups | Each crew role's craft knowledge ships as a versioned skill directory (H02), so one role runs unchanged on the Astra lane, an Opus 5.5 profile or a cheap binding. |
| GEPA (ICLR 2026 oral) evolves prompts by reflecting on full execution traces and keeping a Pareto frontier; it beat GRPO by 6% on average and by up to 20% with up to 35x fewer rollouts, and MIPROv2 by over 10%, on its tasks. | P: arXiv 2507.19457 | One of the two ways H10 proposes prompt and skill changes; promotion still needs a replayed gain on held-out work. |
3.2 Agentic film research, 2026#
| Work | Source | What V1.5 takes from it |
|---|---|---|
| FilmWorld (2026-07-21): novel-to-film as dynamic cinematic world modelling. A construction phase turns prose into stateful world entities with visual grounding and shot plans; an evolution phase propagates state across shots and verifies it in a closed loop. Largest gains in narrative fidelity and cross-scene consistency. | P: arXiv 2607.19038 | Typed story state (F.04.05) extended into a world-state model (W03, T06) that every shot reads and every change updates; closed-loop verification after each shot (Q03). |
| CineForge (2026-08-30): typed narrative, character, spatial and cinematic states; records every production trajectory and learns policies from recurring failures (case → pattern → policy), validated by replay; +8.8% on its benchmark and 37% fewer review calls on new stories. | P: arXiv 2608.29621 | Self-improvement from recorded trajectories with replay-validated promotion (H10), which also matches the SkillsBench lesson recorded for V1.3: self-authored skills add nothing unless promoted on a replayed gain. |
| DirectorBench (2026-05-28): minute-long multi-shot evaluation across script, visual, audio, cross-modal sync and stability; 40 checkpoints, 7 viewer profiles. Across the workflows it evaluated, transition quality averaged 0.256 while prompt fulfilment reached 0.71. | P: arXiv 2605.30090 | Q02 scores transitions explicitly; the 3D-truth pipeline (principle 3) exists largely to win on cuts. |
| FRAMEWORKERS (2026-08-30): a Director agent edits a task stack; an assistant grounds work in a shared workspace; sub-agents register by descriptor and join without redesign; it recovers from runtime failures. | P: arXiv 2608.29814 | The crew runtime borrows its shape (H01–H03): a task graph over the production graph, roles registered by capability descriptor. FRAMEWORKERS' measured gains come from its own trained director and do not transfer to this crew without local held-out evidence (Q12). |
| Also: Co-Director (2604.24842), ViMax (2606.07649), "The Script is All You Need" (2601.17737), AniMaker (SIGGRAPH Asia 2025, MCTS clip search). | P: arXiv listings | Read by H02 during the crew design. |
| CineCrew (2026-09-07, ECCV 2026): FilmDSL, a typed film language between script and generator (shot and camera directives, asset and continuity requirements, persona cues); a generation agent builds asset packages and storyboard keyframes before clip synthesis; a critic agent emits structured QA and triggers targeted repair. | P: arXiv 2609.07720 | Confirms the typed shot record (F.04.04, A01) as the contract between story and pictures, and critique that repairs one element rather than regenerating everything (Y10, H16). |
| MVAgent (2026-09-24): spatial-grounding and transition agents build consistent conditions; shot-level selection keeps the best candidate per shot as the trunk the next shot continues from. GroundShot (2026-06, revised 07-20): an entity-level visual memory built from accepted shots, verified before entry, retrieved as references before each later shot, with shots scheduled by how useful they are as references; training-free. | P: arXiv 2609.30609, 2606.20799 | Candidate search with a per-shot trunk and a verified entity memory (H16); the production graph already holds the accepted shots such a memory needs. |
| Also July–September 2026: VIGA and Thinking in Blender (vision-as-inverse-graphics agents that alternate looking and editing an editable scene); Blender-VideoBench (2609.15478, separates factual reconstruction from perceptual similarity); AnimationBench (2604.15299, the twelve principles and character preservation instead of realism); SurpriseForcing and TetherMem (long-video memory that keeps a subject and lets the background change); MemoryCard (video memory as events, frames and utterances); CRePE (lens-aware camera control; NeurIPS 2026 acceptance reported by its author); CachedSearch (cheaper candidate search); LynnReal-Omni (a unified controllable video model). ICCV is not held in 2026. | P: arXiv listings | Q01 scores factual and perceptual outcomes separately; C11 calibrates stylised identity; Y05 and H09 gain revision-aware visual memory; Y04 checks lens projection; H16 compares cached search with best-of-N. None is adopted without local evidence (Q12). |
| Test-time search for video: best-of-N with a mixture of verifiers (Video-T1) and Temporal Backtracking Search (ICML 2026), which restarts generation from the last verified prefix instead of from scratch because errors cluster early. | P: arXiv 2503.18942, 2606.13861 | Every generative step is a search with verifiers, with N set from a measured curve (principle 12, H16). |
| Story-quality judges are weak: on LitBench the best off-the-shelf judge agrees with human preference 73% of the time, trained reward models 78% (results for that benchmark's literary task and models, not a general ceiling). VLM judges have seen lossy fixed-rate frame samples; since 2026-09-01 Gemini's agentic video understanding (3.7 Flash, 3.6 Flash, 3.5 Flash-Lite) chooses what to watch and re-reads sections at a higher rate (Google reports up to 88% fewer tokens and 7% higher accuracy). | S: LitBench (EACL 2026); VLM-judge surveys; P: blog.google, 2026-09-01 | Story and visual critics are advisors, never the gate on taste; Elikem's ratings and screenings are (Q04, Q06). Critic panels use several models plus rule checks, and a dailies critic reports which frames and at what rate it examined (Q02). |
3.3 Generative media available now#
| Fact | Source | Consequence |
|---|---|---|
| OpenRouter serves video generation. A live call on 2026-09-29 listed 29 video models, including Seedance 2.5 and 2.0 (text, image, video and audio input), Kling v3.0 Pro/Std and Kling o1, Veo 3.1 (and Fast, Lite), Wan 3.0 (and Prime), Wan 2.7/2.6, MiniMax Hailuo 3 (and Max), Runway Gen-4.5 and Runway Aleph 2 (video-to-video), FLUX.3 video, FLUX video edit and upscale, Grok Imagine Video 1.5 and HeyGen Avatar IV. It also lists 57 image models and Lyria 3 music. | P: GET openrouter.ai/api/v1/models?output_modalities=video |
The note that OpenRouter had no video model (Isis, 2026-09-16) is out of date. A hosted video lane on the preferred route is possible (Y02), under the terms register (F.01.01) and per-production caps. Paid avatar products (HeyGen) stay off the default path, as V1.4 decided. |
| Seedance 2.5 (GA 2026-07-31): multi-shot generation that holds characters and style across cuts; up to 50 references (30 images, 10 clips, 10 audio); 30 seconds per request; lip-synced dialogue and stereo sound generated with the picture. Seedance and Kling lead the public video arenas; Veo 3.1 is often chosen for cinematic 4K (vendor and arena claims; Y09 measures each lane per shot class). A secondary source says Sora 2's API was scheduled to end on 2026-09-24. | S: fal, Hedra, arena write-ups | Candidates for inserts, establishing shots, effects plates and restyling. Never the source of a lead character's performance unless it passes the identity and fidelity checks (Y04). |
| Open weights: LTX-2.3 (audio and video in one pass, IC-LoRA and camera LoRAs, 30 s), Wan 2.2 (strong open motion; newer Wan releases are closed), HunyuanVideo keyframe LoRA (not usable here: the Tencent Hunyuan Community License's territory excludes the UK, the EU and South Korea, which Isis T.01.03 already applies to Hunyuan3D and HY-Motion; R05 refuses it). LTX-2.5 (August 2026) is the newest open audio-video release. The repository's RunPod catalog runs Wan 2.2, VACE, Animate-2, SCAIL-2 and LTX-2.5 AV (which replaced LTX-2.3 in the catalog); Uni3C is only in the RunPod inventory so far (E.03.06 open). | S: LTX, Magic Hour; repository | Render-to-video restyling from Blender control passes (F.16) runs on our own workers by default; hosted lanes are the second choice. |
| 2D and anime: ToonComposer (Tencent ARC, ICLR 2026) does inbetweening and colourisation in one model, succeeding ToonCrafter. | S: Hugging Face card | The 2D look lane (V07, N09) pairs Grease Pencil keys with a measured inbetween model. |
| 3D assets: TRELLIS.2 (MIT, 4B, PBR, GLB) is a leading downloadable image-to-3D model; Pixal3D (MIT) added multi-view in September; Hunyuan3D 3.1 is hosted only and its open predecessors carry the same UK exclusion; Tripo P1.0/H3.1, Rodin Gen-2.5 and Meshy 7 are hosted. UniRig (MIT) learns skeletons and skinning; Autodesk Flow Studio added AI rigging (2026-04-28). | S: Cinevva timeline, Autodesk | The T track owns asset generation (T.05) and rigging (T.11); V1.5 consumes it through the character pipeline (C). |
| Motion: OpenT2M (2,800 hours, March 2026) and Text2BFM (ICML 2026) advance text-to-motion; creature and stylised motion remain weak. Single-camera markerless capture (Move AI Gen 2) exports FBX/USD for Blender. | S: arXiv, CVPR, Move AI | Elikem can act a performance to a phone camera (N04); text-to-motion is admitted only on a model whose code, weights and training data all clear the register: the studio's own any-skeleton model trained on cleared data (Isis T.12.10, section 21), or the SMPL family under the T.12.07 licence that Decision 21 declined. |
| Face: NVIDIA open-sourced Audio2Face (models, SDK, training framework, Audio2Emotion). ARKit's 52 blendshapes stream from an iPhone into Blender through open add-ons (LiveLinkFace receiver, Niram); a non-iPhone path exists (Face Landmark Link). | S: NVIDIA, AWN, Blender community | Two face paths: audio-driven lip sync and emotion (N06) and Elikem's own facial performance (N05), both writing to the character's shape keys through its capability manifest (T.11.06). |
| World models: World Labs Marble generates persistent, exportable 3D worlds as Gaussian splats from text, images, video or panoramas (World API since January 2026); Genie 3 worlds cannot be exported. Blender 5.3 imports and renders splats natively (PLY, SPZ, USD) but cannot export them yet. | S: Radiance Fields, Creative Bloq, Digital Production | Splat worlds as set backgrounds and location scouting (W09, F.08.04), gated on 5.3's release and the vendor's terms. |
| Voice: Chatterbox (emotion-exaggeration control, 23 languages), Orpheus (emotion tags), Sesame CSM, Dia (multi-speaker with non-verbal sounds), Kokoro; OpenRouter's speech endpoint returns no timings. Music: ACE-Step XL 4B (open, April 2026), Lyria 3 on OpenRouter, Suno v6 (licensed-data training; label litigation not fully settled), AIVA for orchestral work. | S: tool guides; P: OpenRouter listing | Voice acting (D02) and score (D05) choose by bake-off under the licence gate; no cloned real voices; Suno stays out until its litigation and terms clear the rights register (R05). |
| Voice performance control: IndexTTS2 (AAAI 2026) describes an exact-duration mode (a set token count) and takes timbre and emotion from separate prompts, built for dubbing; the released IndexTTS2.5 (2026-08-10) exposes a speaking-duration factor but does not enable the precise mode, so frame-exact delivery is measured, not assumed (Bilibili Model License: commercial use below user and revenue thresholds); Eleven v4 (2026-09-28, paid, opt-in only) and Fish S2 (open, March 2026) join the bake-off; Qwen3-TTS (January 2026, Apache-2.0) designs a voice from a text description; zero-shot speech-to-speech conversion (Seed-VC and its successors) carries a spoken performance onto another timbre. | P: arXiv 2506.21619, the IndexTTS repository, Qwen3-TTS model cards, elevenlabs.io and fish.audio announcements; S: licence summaries | Lines generated to a target length for locked animation and dubbing (D02, K09); Elikem directs a line by acting it, converted to the character's original designed voice (D12), never to a real person's voice. |
| Sound from picture: video-to-audio foley models generate synchronised effects and ambience from footage (Foley-Omni, arXiv 2606.03672; LTX's video-to-audio mode; FoleyBench scores sync and semantics). HunyuanVideo-Foley is excluded here by the same Tencent territory clause. Seed Audio 1.0 (ByteDance, 2026-06-23, on OpenRouter) returns one mixed scene track of voice, music, effects and ambience up to 2 minutes, which suits scratch audio but not stems. | P: arXiv, Hugging Face licence, OpenRouter listing | A video-to-audio lane proposes layers beside the 3D-event sync, scored before use, stems only (D04); Seed Audio for temp tracks at most. |
| Animation in-betweening: Animaj's Adaptive Interpolation-Synthesis (SIGGRAPH 2026; in production on Pocoyo and Maya the Bee; 3.5x faster in-betweening; outputs rig-control keys in Maya or Blender) and NVIDIA's ARDY (SIGGRAPH 2026; keyframe, path and sparse-joint constraints with about 33 ms latency for its 4-step model, code and checkpoints released). Code and weight licences are not yet verified. | P: arXiv 2605.02742, 2607.08741; Animaj and NVIDIA project pages | Learned in-betweens seed the spline pass as editable keys, admitted only after licence clearance and a held-out comparison on Oshun's rigs (N14). |
| Real-time video-to-video: StreamDiffusionV2 (MLSys 2026 best paper, open; 58 fps with a 14B model and 65 fps with 1.3B on four H100s) and Decart's realtime restyle API (Lucy 2.0, WebRTC). | P: arXiv 2511.07399; decart.ai docs | A live look preview over the Blender viewport, advisory only and never in the cut, kept only if it changes decisions (X09). |
| Identity adapters from a character's own renders: studios and tool vendors train a character LoRA on 20–50 rendered views (angles, expressions, shot sizes, lighting) so the image and video models hold the character across shots. | S: RunComfy LTX-2.3 character-LoRA trainer, Reallusion AI-film guidance, practitioner guides | Oshun renders its certified 3D characters and trains identity adapters on those renders only (Y11), which is rights-clean and ties the generated finish to the 3D truth. |
| Images: GPT Image 2.5 (2026-09-08) is the newest hosted image model; Qwen-Image-2.1 (September 2026) ships under a research and non-commercial licence with separate commercial terms, so downloadable weights are not a production admission. Music: Lyria 3.5 (2026-09-03). | P: OpenAI API changelog; Hugging Face licence file; Gemini API changelog | Refresh Y07's and D05's candidates through the rights register (R05). |
| Google's Gemini Omni (I/O 2026) makes video from text, images, video and audio and edits conversationally with a reference video; it ships through Google's own Flow app and YouTube, not OpenRouter. OpenAI discontinued Sora: the apps closed in April 2026 and the API ended on 2026-09-24. | S: Google I/O coverage, Wikipedia | No lane change; the field watch (Q10) tracks both. |
| Previs tools converge on the same shape: lock likeness, choose camera and lens, re-render one panel, drop the result into a timeline (LTX Studio, FrameForge's 3D blocking with real lenses). | S: vendor pages | The boards-to-layout path (F.05, F.06) plus studio UI (P pillar) matches the market and adds 3D truth. |
3.4 Blender#
- Blender 5.2 LTS (2026-07-14, supported to July 2028): experimental Geometry Nodes XPBD physics for cloth and hair, bundles and lists, a Bevel node, audio sampling; Grease Pencil gets an exact Delaunay fill with gap detection and new line materials; EEVEE's screen-space ray tracing is overhauled; Cycles adds a texture cache; online asset libraries. The repository pins 5.2.1; the bug-fix release 5.2.2 shipped on 2026-09-15 (B01 moves a title only after its golden shots re-render, Q09).
- Blender 5.3 is in alpha until 2026-09-30, with release expected in November. In development and subject to release: native Gaussian splats (import and render, no export; USD import of the Gaussian-splat schema), and Material Lighting nodes (Light Accumulation, Light Info, Light Evaluation, Shadow Raycast, a light-mode Attribute node and light-space Vector Transform) that let a material decide how it reacts to each light, which is the per-light control toon shading lacked. Native OpenTimelineIO in the sequencer is a 2026 GSoC project, not merged at the time of writing. Blender 5.4 is proposed for 2027-05-11 (release-cycle proposal, not a commitment).
- Blender Lab ships an official MCP server (the repository pinned 1.0.3 in T.20.01) plus Clay Pencil and XR work. Flamenco 3 is Blender's official render-farm manager and has a Cycles GPU job type.
- NPR has moved: Blender 5.1 added the Shader Raycast node (EEVEE, screen-space) for lines, rims, x-ray and other stylisation integrated with scene shading, so the pinned 5.2 LTS has it; 5.3 adds the Material Lighting nodes above. Toon kits (V05) use Shader-to-RGB, ramps and Grease Pencil line art together with the Raycast node now, and adopt the lighting nodes with 5.3 (B10).
- Colour since 5.0: ACES 1.3 and 2.0 workflows (ACEScg working space, ACES 2.0 views), Rec.2100 PQ and HLG and AgX HDR displays, EXR in ACES2065-1 or ACEScg. The pilot's AgX decision (F.00.03) stands; an HDR master for Lilith TV is a per-title option F.14.03 and F.19.04 can take without new tooling.
3.5 What the industry has shown#
- Flow (Gints Zilbalodis) won the 2025 Academy Award for Best Animated Feature, the first winner made entirely in Blender. It was rendered with EEVEE at 0.5 to 10 seconds per 4K frame on the director's own computer with no render farm and no compositing (colour was set in the shaders), and the director laid out every scene himself before set dressing (S: blender.org interview, Creative Bloq). Il Baracchino, Italy's first adult animated series, also finished in EEVEE (P: blender.org user story, 2026-02-04). For a stylised title, EEVEE is a credible final engine; V05 measures it against Cycles and B09 plans capacity from that measurement.
- Critterz, OpenAI-backed and produced with Vertigo Films and Native Foreign, set out to make an animated feature in nine months for under $30 million, with a human-written script, human voice actors and artists' sketches feeding the tools; a Cannes premiere was planned for May 2026 (S; whether it took place is unverified). The model is human authorship with machine labour, which is principle 1.
- Showrunner (Fable, public since August 2026) generates 10-to-20-minute animated episodes from a short prompt inside two house shows, and chose animation to cut compute (S, unverified). It shows demand and the autonomous extreme; it also shows why V1.5 keeps authorship, canon and control first.
- Solo creators are shipping series: DRAGGLiNS, a 3D children's series made by one person with an AI pipeline (four episodes on YouTube from 2026-04-05); ASHES, a 15-episode series made for under $1,000 (S, unverified). Reported multi-agent pipelines produce shorts in 2 to 5 days at $315 to $750 per finished minute (S, unverified; nothing plans on these figures). The bar V1.5 aims at is above these: festival-grade, not merely finished.
- Runway's fourth AI Film Festival (2026) gave its Grand Prix to "A Face Only a Mother Could Love" (S). Festival entry needs a clean chain of title and an honest account of the tools, which R pillar produces.
3.6 Rules that bind the films#
| Rule | Source | Consequence |
|---|---|---|
| US: prompting alone is not authorship; human selection, arrangement and modification of AI output are protectable. The Supreme Court denied certiorari in Thaler on 2026-03-02. | S: Morgan Lewis, Holland & Knight | The authorship ledger (R01) records every human decision with its artifact, so any film can show its human authorship. |
| UK: the government's March 2026 Report on Copyright and AI proposed removing CDPA s9(3)'s protection for computer-generated works (a provisional view; the law is unchanged). Films, sound recordings and broadcasts keep their own entrepreneurial protection, separate from the authorial works inside them. No text-and-data-mining reform for now. | P: gov.uk report; S: HSF Kramer, IPKat | Oshun Synthetics LTD should not plan on s9(3) lasting for autonomous films. Directed films rest on Elikem's authorship; autonomous films are released knowing their protection may become thin (Decision 7); counsel reviews each title class before release (R14). |
| EU AI Act Art. 50 applies from 2026-08-02: generative providers mark output machine-readably; deployers disclose deepfakes, with a lighter duty for evidently artistic or fictional works. The Commission's Art. 50 code of practice was published 2026-06-10: providers mark output with signed metadata (C2PA meets the signed-metadata criterion) and imperceptible watermarks and offer public detectors by 2027-02-02; deployers disclose visibly with the EU AI Office's common icons; CAWG metadata assertions are encouraged; the IPTC digital source type vocabulary names how each item was made. Current versions: C2PA 2.4 (April 2026) and the CAWG Identity Assertion 1.3 (ratified 2026-08-17). A C2PA credential proves who signed what was asserted, not that the assertions are true or that viewers saw a disclosure. | P: Regulation (EU) 2024/1689; European Commission announcement 2026-06-10; C2PA 2.4 specification; cawg.io; IPTC NewsCodes; S: Greenberg Traurig | Every delivered film is C2PA-signed, its credential verified separately, and it carries a visible disclosure tested after transcode on each destination, per a destination-by-jurisdiction matrix (R04), with per-ingredient IPTC digital source types in the manifest and the common EU icon where the code applies; no film depicts a real person. |
| The Academy: generative AI "neither helps nor harms"; the branches judge how far a human was at the heart of the authorship; screenplays must be human-authored for its writing categories; the 99th Awards rules (2026-05-01) reinforce human writing and performance and let the Academy ask about AI use. Festival rules vary. | P: Academy press releases (including 2026-05-01); S: coverage | A submission pack (R09) cites the authorship ledger and checks the destination's own rules; T13's strict human-authored screenplay mode is an Oshun option, switched on where a destination requires it. |
| SAG-AFTRA's TV animation agreements (ratified 2024, running to 2026-06-30) require consent before a performer's name is used as a prompt, govern digital replicas, and require notice and bargaining when AI voices replace actors. SAG-AFTRA and the AMPTP announced a tentative successor animation agreement on 2026-09-17, with terms not yet public; the WGA's 2026 MBA was ratified on 2026-04-24 and SAG-AFTRA's 2026 TV/Theatrical contracts on 2026-06-04. | P: sagaftra.org (2026-09-17, 2026-06-04), wga.org (2026-04-24); S: Engadget, IP Law Blog | Original synthetic voices, or a hired performer's own recorded performance under R17 (Decision 6); consent to record a performer never licenses a replica of that performer (principle 8); R17's performer pack names the agreement, version and status that apply. |
| Netflix's partner guidelines (2025; a dated summary, since Netflix now points partners to a newer hub whose deal-specific requirements apply per sale): no replication of unowned IP, tools must not train on production data, enterprise-secured tools where possible, generated material temporary unless approved, no replacing talent without consent. | S: Netflix via trade press | The studio is designed toward these principles (production-data rules R10, rights R02/R16, performer permissions R17), but a streamer's own approval process for final-use AI material still applies to any sale. |
4. Principles (binding on every task in this family)#
- Human at the heart of authorship. Elikem is the author. Agents write, draw, animate and cut as proposals he can accept, edit, reject or lock. This is a creative promise with legal weight: US copyright protects human selection, arrangement and modification but not prompts alone (the Supreme Court left Thaler standing on 2026-03-02); the UK government's March 2026 report proposed removing the computer-generated-works protection in CDPA s9(3), though the law has not changed; the Academy judges "the degree to which a human was at the heart of the creative authorship" and requires human-authored screenplays in its writing categories. Every artifact records who authored, selected and changed it (R01). The record is evidence for a claim, not the claim itself; a person reviews any registration (R14).
- One operation set, two kinds of client. No capability exists only for agents or only for the interface. Every studio action is a typed, versioned, undoable operation with a schema, callable by Elikem (studio or Blender) and by an agent (MCP/API) with the same permissions model (A03).
- Blender is the source of truth; generative output is measured against it. A generated finish is labelled, carries its control passes and passes a fidelity check against the rendered depth, pose and identity before it can replace a render in the cut (Y pillar, F.16.04).
- Everything is data. Worlds, characters, scripts, boards, shots, takes, cuts, mixes and grades are versioned records with diffs, branches and lineage. A film can be rebuilt from its records (F.19.06). Nothing important lives only in a chat transcript or a model's context.
- The autonomy dial is per department and per scope. Manual, assist, propose, auto-with-review and auto are settings, not modes of the product (section 5). Raising the dial requires benchmark evidence for that department (Q pillar), and lowering it is always one click.
- Cheap by default, frontier by choice. The default binding of every agent role is the cheapest model that passes its bake-off on OpenRouter; frontier profiles (GPT-6 Astra through the Codex subscription only; Claude Opus 5.5 through a key Elikem approves) are opt-in per production, with a spend cap, and never a dependency. Tests, harnesses and benchmark runs in CI always use CLAUDE.md's test binding; a frontier comparison is capped production spend Elikem enables, never a test (H04).
- Honest measurement. Rule-based checks run before any model judge; model judgements are labelled as such; human preference (Elikem's ratings, then audience screenings) is the ground truth for taste. Nothing reports success it did not compute.
- Original work only. No synthetic replica of any real person's likeness or voice, ever; a recorded human performance (a hired actor's takes, Elikem's own motion or face capture on an original character) only under a scoped permission record (R17). No recognisable third-party character, style-of-a-living-artist prompt or unlicensed input; every model, LoRA, voice, font, sample and reference carries its licence in the chain of title (R pillar). Folklore is not automatically free to use: Ghana's Copyright Act 2005 protects expressions of folklore in perpetuity and requires the National Folklore Board's permission for many uses, so folklore sources are cleared (R15) as well as reviewed by consultants (R08).
- Solo-builder ergonomics. One person runs the studio. The workspace always answers "what needs me now", batches decisions, works from a phone for review, and never makes Elikem babysit a queue.
- Build on the existing tracks, never beside them. Where an Isis F/T, content-workbenches CR or V1.4 task already owns a capability, V1.5 depends on it and integrates it; the ownership map (G02) settles every overlap before code.
- The machines have limits. Rendering, simulation, alignment and model inference run on RunPod, the Linux dev server or a render farm, never stacked on the Mac (CLAUDE.md machine limits). Estimates come before spend (F.14.05).
- Search, then choose. Every generative step (a board, a design, a restyle, a voice take, a cue, a cut) is a search with verifiers: several candidates within a reserved budget, rule checks first, then the critic panel, then Elikem's pick or a measured automatic pick. How many candidates a shot class gets comes from a measured quality-against-cost curve (H16), never from habit.
5. The autonomy dial#
| Setting | Who proposes | Who approves | Example |
|---|---|---|---|
| Manual | Elikem | Elikem | He keys the performance himself in Blender; agents only run checks. |
| Assist | Elikem, with agent suggestions inline | Elikem | Dialogue pass with alternatives offered per line. |
| Propose | Agent, as a diff | Elikem, per item | The cinematographer proposes coverage for a scene; he accepts two setups and edits one. |
| Auto with review | Agent, applied | Elikem, in batches at a gate | Overnight: lighting passes on 40 shots, reviewed as a contact sheet in the morning. |
| Auto | Agent | The quality gates (Q), then Elikem at the release gate only | An autonomous episode from the series bible, labelled as such (R07). |
The dial is set per department (story, research and consultation, world, characters, costume, colour design, graphics and motion graphics, props, art, boards, layout, camera, animation, choreography, vehicle action, face, lighting, effects, render, finish, edit, sound, music, grade, delivery) and can be overridden per film, sequence or shot. Locks (A05) always win: an agent at any setting cannot change a locked record.
6. Scope by pillar#
6.1 Programme, capability gate and production office [G]#
The V1.5 capability gate, the ownership map against the Isis F/T tracks, the content-workbenches creative studio and V1.4, the studio architecture ADR, the decision register, the production office (slate, greenlight, calendar, budgets), the cost ledger with per-production caps, the milestone exit runner and the studio handbook.
Tasks: G01, G02, G03, G04, G05, G06, G07, G08, G09, G10, G11, G12, G13, G14 (board map, section 11).
6.2 Production graph and the operation core [A]#
The data model every department shares: the production graph from project to take, versioning with branches and diffs, the single typed operation API used by the interface and the agents alike, the MCP server over it, locks and approvals, lineage and selective regeneration, the media store, events, interchange formats, search and archive.
The core idea of V1.5 is that the studio has one set of operations and two kinds of client. The same typed operation moves a camera whether Elikem drags it in Blender, clicks it in the layout room, or an agent proposes it overnight. The operation carries the permission check (locks, the dial, the capability gate), the undo group, the cost estimate and the authorship row. That is what makes "total control" and "real autonomy" the same system rather than two products.
This picture is wider than it can be shown here. Move across it, or read it as text.
Drawn from its Mermaid source.
As text
Parts: - Elikem: studio rooms - Operation API: schema, locks, dial, gate, undo, cost, authorship - Elikem: desktop Blender live link - Agent crew via MCP - Production graph: versions, branches, lineage (data) - Jobs: headless Blender, render farm, generation lanes - Media store (data) - Events - What needs me / agent console / notifications - Authorship ledger + chain of title + cost ledger (data) Connections: - Elikem: studio rooms to Operation API: schema, locks, dial, gate, undo, cost, authorship - Elikem: desktop Blender live link to Operation API: schema, locks, dial, gate, undo, cost, authorship - Agent crew via MCP to Operation API: schema, locks, dial, gate, undo, cost, authorship - Operation API: schema, locks, dial, gate, undo, cost, authorship to Production graph: versions, branches, lineage - Operation API: schema, locks, dial, gate, undo, cost, authorship to Jobs: headless Blender, render farm, generation lanes - Jobs: headless Blender, render farm, generation lanes to Media store - Production graph: versions, branches, lineage to Events - Events to What needs me / agent console / notifications - Production graph: versions, branches, lineage to Authorship ledger + chain of title + cost ledger
Show source
flowchart LR
E[Elikem: studio rooms] --> OPS
BL[Elikem: desktop Blender live link] --> OPS
AG[Agent crew via MCP] --> OPS
OPS[Operation API: schema, locks, dial, gate, undo, cost, authorship] --> G[(Production graph: versions, branches, lineage)]
OPS --> JOBS[Jobs: headless Blender, render farm, generation lanes]
JOBS --> MS[(Media store)]
G --> EV[Events] --> HOME[What needs me / agent console / notifications]
G --> LED[(Authorship ledger + chain of title + cost ledger)]
Tasks: A01, A02, A03, A04, A05, A06, A07, A08, A09, A10, A11, A12, A13, A14 (board map, section 11).
6.3 The studio workspace (the director's chair) [S]#
The one place Elikem works: shell and navigation, the "what needs me" home, the production tracker, the review room, the overnight report, the agent console, director chat at any scope, the Blender live link, mobile review, the autonomy and spend controls, guest collaborators, a craft mentor, accessibility and end-to-end coverage. Department rooms live in their own pillars.
The rooms, and where each lives:
| Room | Pillar | Built over |
|---|---|---|
| Home ("what needs me"), tracker, review, agent console, director chat | S | the production graph and events |
| Development (ideas, slate, pitch) | I | Hathor ideation |
| World atlas (bible, maps, timeline, cultures, systems) | W | Hathor world editor, Yemaya canon graph |
| Character forge | C | CR07/CR08, the T track |
| Story room (structure, outline, scene design, screenplay, table read) | T | Hathor ladder, Fountain parsers, F.04 |
| Series room | E | Hathor episode grid and series bible |
| Art room | V | F.04.06, CR09, image lanes |
| Boards, animatic, layout, camera | P | F.05–F.07 |
| Animation | N | F.10, F.11, T.11–T.12 |
| Sets, lighting, effects, render, comp | X | F.08, F.12–F.15 |
| Generative finish | Y | F.16, OpenRouter and RunPod lanes |
| Sound, voice and music | D | F.11, F.18, the phase-182 stack |
| Edit, grade, delivery, localisation | K | F.17, F.19 |
Tasks: S01, S02, S03, S04, S05, S06, S07, S08, S09, S10, S11, S12, S13, S14, S15, S16, S17, S18 (board map, section 11).
6.4 Blender service, render farm and compute [B]#
Blender as the studio's engine: version policy, the headless service, the film skills as operations, live sessions, the contained script lane, .blend versioning, the Flamenco render farm on RunPod and the Linux box, estimates and budget holds, the asset library, capacity planning, Blender 5.3 adoption and a Mac-safe local mode.
Heavy work never runs on the Mac. The Linux dev server hosts the Flamenco manager, CPU jobs (playblasts, alignment, simulation caches that fit) and the headless service; RunPod GPU pods run Cycles and EEVEE renders and model inference; the Mac runs desktop Blender on proxies through the live link.
Tasks: B01, B02, B03, B04, B05, B06, B07, B08, B09, B10, B11, B12, B13, B14 (board map, section 11).
6.5 Ideation and the development slate [I]#
Where films start: an idea inbox, guided brainstorming, a premise and logline lab, comparables and collision checks, pitch documents, the development slate and greenlight, the format designer, Elikem's taste and influences board, and an autonomous development run.
Tasks: I01, I02, I03, I04, I05, I06, I07, I08, I09 (board map, section 11).
6.6 World building [W]#
Worlds as structured, canon-checked data: the world bible model, the canon and continuity engine, world state across story time, maps and terrain, timelines, cultures and societies, languages and naming, systems with rules and costs, world visualisation including splat worlds, lore Q&A, bible export and an autonomous world-building run.
Tasks: W01, W02, W03, W04, W05, W06, W07, W08, W09, W10, W11, W12 (board map, section 11).
6.7 Characters [C]#
Characters from psychology to rig: the character record, relationships, interviews in voice, design exploration, model sheets, the 3D build, the face rig standard, wardrobe and hair, acting profiles, voice casting, the on-model checker, screen tests and an autonomous character run. From section 20: the studio's own expression basis (C18) and a per-character basis learned from 4D capture (C19).
Tasks: C01, C02, C03, C04, C05, C06, C07, C08, C09, C10, C11, C12, C13, C14, C15, C16, C17, C18, C19 (board map, section 11).
6.8 Story and screenplay [T]#
From structure to locked script: story structure templates, outline to treatment to step outline, a scene designer, the screenplay editor with revisions and per-line authorship, dialogue passes, story state, coverage by a critic panel, voiced table reads, breakdown handoff, script lock, themes and setups, a research desk, an autonomous writers' room, and a strict human-authored mode.
Tasks: T01, T02, T03, T04, T05, T06, T07, T08, T09, T10, T11, T12, T13, T14, T15 (board map, section 11).
6.9 Series and episodic production [E]#
What a series needs beyond a film: the show bible, season arcs, episode breaking, continuity across episodes, the episode pipeline at scale, recurring elements, release cadence, a series dashboard and an autonomous episode from the bible.
Tasks: E01, E02, E03, E04, E05, E06, E07, E08, E09 (board map, section 11).
6.10 Visual development and art direction [V]#
The look of a title: the art direction brief and style bible, the concept art room, colour scripts, style frames and key art, look-development kits (toon, painterly, stylised PBR), show-specific style models trained only on Oshun's own art, the 2D and anime look lane, the style conformance scorer and art department review.
Tasks: V01, V02, V03, V04, V05, V06, V07, V08, V09, V10, V11, V12, V13, V14, V15 (board map, section 11).
6.11 Boards, previs, layout and cinematography [P]#
Studio rooms and agents over the Isis F.05–F.07 craft: the boards room, the animatic editor, the layout room, the film's camera language, the coverage and shot-design agent, previs review, a phone or tablet virtual camera and the boards-to-layout handoff.
Tasks: P01, P02, P03, P04, P05, P06, P07, P08, P09 (board map, section 11).
6.12 Animation and performance [N]#
Performance in all its forms: the performance plan editor, blocking to spline with pose libraries, the motion library and retargeting, video capture of Elikem's own acting, phone facial capture, audio-driven lip sync and emotion, stylised animation principles, interaction and contact, Grease Pencil 2D animation, crowds, an animation agent that learns from corrections, and text-to-motion from a model that clears the register (section 21: the studio's own any-skeleton model trained on cleared data, Isis T.12.10, for human and non-human cast alike). From section 20: an expression library and a motion-phrase library mined from generated reference (N21, N27), an open-stack face solver with a re-render gate and a body solver with a physics-tracking gate (N22, N26), a gait grammar per species (N28), an inverse-rendering solve in the character's own basis (N23), audio-and-video fusion for dialogue (N24) and an acting-language stylisation layer (N25).
Tasks: N01, N02, N03, N04, N05, N06, N07, N08, N09, N10, N11, N12, N13, N14, N15, N16, N17, N18, N19, N20, N21, N22, N23, N24, N25, N26, N27, N28 (board map, section 11).
6.13 Sets, lighting, effects, rendering and compositing [X]#
Studio rooms and agents over Isis F.08, F.12–F.15: set building, a reusable location library, the lighting room, the effects room, render management, the compositing room, splat and matte worlds, and a lighting agent that works from the colour script.
Tasks: X01, X02, X03, X04, X05, X06, X07, X08, X09, X10, X11, X12 (board map, section 11).
6.14 Generative finish and the hosted media lanes [Y]#
Where generated pixels may enter a film and how: the generative-lane policy, the OpenRouter video lane, render-to-video restyling from Blender control passes, the fidelity and identity gate, long shots, a code-as-shot lane, image lanes for design, upscaling and interpolation, and a bake-off that picks a lane per shot class.
| Shot class | Generated pixels allowed | Gate |
|---|---|---|
| Establishing shots, skies, distant backgrounds | Yes | style scorer, render QA |
| Inserts and cutaways without a lead character | Yes | style scorer, continuity |
| Effects plates (fire, water, crowds at distance) | Yes | comp QA |
| Restyle of a locked 3D shot from its control passes | Yes | fidelity against 3D truth and on-model (Y04) |
| A lead character's performance | Only as a restyle that passes Y04 | Y04 plus Elikem's approval |
| Titles and motion graphics | Code-as-shot (Y06) | deterministic render check |
Tasks: Y01, Y02, Y03, Y04, Y05, Y06, Y07, Y08, Y09, Y10, Y11 (board map, section 11).
6.15 Voice, music and sound [D]#
Everything the audience hears: the voice lane with acting range, voice direction and take comping, alignment, a recording booth, the sound design room, original score with themes and human-editable MIDI, songs, the mix, audio QA and captions, and a sound agent.
Tasks: D01, D02, D03, D04, D05, D06, D07, D08, D09, D10, D11, D12, D13, D14, D15, D16, D17 (board map, section 11).
6.16 Editorial, finishing, localisation and delivery [K]#
From assembly to deliverables: the edit room on OpenTimelineIO, assembly and alternate cuts, cut review, conform, titles and credits with the AI disclosure, the grade room, deliverables and QC, subtitles, dubbing with re-keyed lip sync, accessibility, trailers and promos, and posters and key art.
Tasks: K01, K02, K03, K04, K05, K06, K07, K08, K09, K10, K11, K12, K13, K14, K15, K16, K17, K18 (board map, section 11).
6.17 The agent crew and autonomy [H]#
The crew that can do any job and, when allowed, all of them: the crew runtime, role design, the propose-mode protocol, the controller bake-off, the Codex/Astra lane, the Opus 5.5 profile, overnight and long-horizon runs, the autonomy dial engine, production memory, self-improvement from trajectories, traces and replay, failure recovery, crew safety, and end-to-end autonomous production.
The crew is a set of roles over one operation set, not a chain of prompts:
| Role | Default dial | Writes | Reads |
|---|---|---|---|
| Showrunner / director agent | propose | the task stack, cross-department proposals | everything |
| Producer | auto with review | schedules, budgets, job plans | ledgers, estimates |
| Writer, story editor, punch-up | assist | outlines, scenes, dialogue proposals | world, characters, story state |
| World-builder | propose | world records | canon |
| Character designer, art director | propose | designs, sheets, looks | style bible |
| Board artist, layout artist, cinematographer | propose | boards, blocking, cameras | shot list, camera language |
| Animator | propose | keys, layers | performance plan, acting profile |
| Lighter, effects, compositor | auto with review | rigs, bakes, comps | colour script, style frames |
| Editor | propose (alternate cuts only) | alternate timelines | takes, script |
| Sound designer, composer | propose | cues, stems | spotting, themes |
| Colourist | propose | grades | looks |
| Production manager | propose | schedules, bids, quotas (G12) | shot list, estimates, actuals |
| Script coordinator, researcher | assist | revision sets, cited research | script, sources |
| Subject-matter experts (history, psychology, medicine, weapons and martial arts, science, law, religion, music, dance, animals, child development, representation, textiles, architecture) | propose (advisory) | notes on exact records, never a required human sign-off | their domain |
| Production designer, costume designer, colour designer, graphic designer, prop master, character TD | propose | sets, costumes and plots, colour models, graphics, props, rigs | style bible, story state |
| Fight choreographer, dance choreographer | propose | phrases in notation, stunt-vis | fight bible, beat grid |
| VFX supervisor, post supervisor | propose | lane choices, delivery plans | fidelity reports, QC |
| Voice director, casting director, dialogue editor, music supervisor, mixer | propose | reads, casting proposals, comps, clearance requests, mixes | script, lexicon, stems |
| Localisation lead, marketing | propose | dubs and subtitles, trailers and key art | locked cut |
| Script supervisor, QC | auto | checks and notes only | everything |
| The full roster, with every role a top studio staffs mapped to its tools, Elikem's controls and its agent, is section 13. |
The defaults are starting points; Elikem sets the dial per department and scope (section 5) and the evidence gate (Q05) limits how high it can go.
Tasks: H01, H02, H03, H04, H05, H06, H07, H08, H09, H10, H11, H12, H13, H14, H15, H16, H17 (board map, section 11).
6.18 Quality, evaluation and taste [Q]#
How "compelling and high quality" is measured: the film benchmark ladder, critic panels, continuity verification after each shot, Elikem's taste model, the autonomy evidence gate, blind screenings, directed-versus-autonomous comparisons, the capability matrix, a quality regression suite and a field watch that keeps section 3 true.
The benchmark ladder rises with the milestones: shot and scene rungs at M0 and M1, the pilot at M1, the short at M2, the episode at M3. Every rung runs rule checks first, then critic panels, then human preference, and every report separates the three. Section 20 adds a face rung (Q14) and a body rung (defined by N26) at M2, so every face and body producer reports against a number before it claims to be better.
Tasks: Q01, Q02, Q03, Q04, Q05, Q06, Q07, Q08, Q09, Q10, Q11, Q12, Q13, Q14 (board map, section 11).
6.19 Rights, authorship, provenance and standards [R]#
What makes the films Elikem's and safe to release: the authorship ledger, chain of title, the originality and likeness screen, provenance and disclosure, the rights register, film standards and ratings, labelling of autonomous work, cultural consultation, festival and copyright packs, production data handling, music rights and title clearance.
The authorship ledger is the studio's legal backbone. Every operation already carries its actor, so the ledger costs nothing extra to keep, and it lets any film state exactly which decisions a human made. The chain of title is its counterpart for inputs: no title can be published while any model, voice, music, font or reference in it has an unknown licence. Section 20 adds the face and motion stack admission (R21): every solver, body model, physics engine and dataset the performance lanes may use is admitted or refused in the register before it is wired, and the Max Planck family (FLAME, SMPL, SMAL, AMASS) stays refused under Decision 21 (build the studio's own basis and skeletons). Section 21 adds what UniMate taught the register (R05): code, weights, training data and the training data's own provenance are checked separately; a disputed provenance refuses a source until resolved, even a pack bought commercially; and each row pins the terms revision it read, with a re-read date.
Tasks: R01, R02, R03, R04, R05, R06, R07, R08, R09, R10, R11, R12, R13, R14, R15, R16, R17, R18, R19, R20, R21 (board map, section 11).
6.20 Distribution, premieres and audience [L]#
Getting films to people: the films-and-series channel on Lilith TV, the on-demand library in the watch area, premiere events, syndication, festival submissions, audience analytics fed back to development, and press kits.
Tasks: L01, L02, L03, L04, L05, L06, L07, L08, L09, L10 (board map, section 11).
6.21 Productions: the slate that proves the studio [M]#
The films that are V1.5's acceptance tests: the directed pilot, its autonomous remake, the short, its autonomous twin, the series pilot episode, season one, an autonomous episode, and a post-mortem after each.
The films are the acceptance tests. Each production is directed by Elikem first and then attempted by the crew alone from the same brief, so every milestone yields both a film and a measurement of how close autonomy has come.
Tasks: M01, M02, M03, M04, M05, M06, M07, M08, M09, M10 (board map, section 11).
6.22 Horizon [Z]#
After V1.5 GA: a feature film, opening the studio to Lilith members and outside creators, interactive and live formats, game tie-ins, immersive worlds, and character licensing.
The horizon is not part of V1.5 GA. It exists so the studio is built to scale to a feature and to other creators without a rebuild.
Tasks: Z01, Z02, Z03, Z04, Z05, Z06, Z07, Z08, Z09 (board map, section 11).
6.23 Consultants and subject-matter experts [O]#
The experts a top production consults, as records and as agents: the expert network and consultation records for every domain (generalising R08 beyond culture), an AI expert panel grounded in cited research, period research with an anachronism check, and psychology and portrayal review. An expert agent is an advisor: domains marked for human sign-off (real cultures, religion, mental health, disability) still wait for a person, and every agent note is labelled as model judgement.
Tasks: O01, O02, O03, O04, O05 (board map, section 11).
6.24 The studio inside Blender [U]#
Everything is controllable from the web studio; the Blender workspace is the second, deeper door into the same production. S08's add-on grows into a Blender extension with an app template, a workspace for every department, the same typed operations (so a control in Blender, a control in the web room and an agent's MCP tool produce the same diff, which U03 proves with a parity table), agent proposals shown as ghosts in the viewport with item-by-item acceptance, co-driving with the agent through the live-session lease, review notes drawn on their frames, context panels for the shot in hand, the craft instruments a department head reads (composition, screen direction, exposure, arcs and spacing, an exposure sheet, the phoneme track, palette and on-model overlays), and asset shelves fed from the studio library.
| Workspace | What it adds to stock Blender |
|---|---|
| Story and boards | Script and beat panel beside Grease Pencil panels; animatic timing from the shot list |
| Layout and camera | Camera rules, lens kit, composition and screen-direction overlays, coverage proposals as ghost cameras |
| Animation | Performance plan, acting profile, pose and motion shelves, human and agent layers, arcs, spacing and X-sheet, learned in-betweens |
| Face and lip sync | Dialogue waveform and phoneme track, viseme keys, close-up review sheets |
| Costume and props | Costume plot and prop state for the shot, fabric presets, continuity warnings |
| Choreography | Fight and dance phrases in notation beside the stunt-vis |
| Lighting | Colour script swatch, colour models, light rigs, false colour, light groups |
| Effects | Simulation stages with their reservations and caches |
| Compositing | Comp templates, Cryptomatte mattes, comp QA |
| Edit, sound and grade | Sequencer on the OpenTimelineIO cut, spotting markers, stems, the look |
| Asset departments | Modelling and sculpt, retopology and UV, texture paint and surfacing, rigging and weights, grooming and procedural nodes with studio context, over the T-track tools |
| Review | Notes and drawovers on their frames, approvals, playblast to review |
The extension's manifest permissions are declarations, not a sandbox: the broker authorises every production, networking stays off Blender's main thread, native edits are explicit checkpointed transactions, and taking the lease from an agent is a quiescent, fenced handoff (U01, U03, U04). Costume, props, choreography and research workspaces arrive with their departments (C14, X10, N15, O03).
Tasks: U01, U02, U03, U04, U05, U06, U07, U08, U09, U10 (board map, section 11).
7. Milestones#
| Milestone | Exit (evidence in G08's exit audit) |
|---|---|
| M0 Foundations | Capability gate live; ownership map and architecture ADR approved; the production graph with versions, branches and locks, the operation API (failing closed for agents until the dial exists), the intake gate, publication authorisation, resource admission and durable job reconciliation live on Dev; the evaluation protocol frozen; the Blender service renders a shot headless on the Linux box and on a RunPod worker; the agent crew runtime registers its roles, plans and returns dry-run proposals for one department on a fixed fixture with a cost ledger (live agent writes follow the dial engine, H08, at M1); the authorship ledger records every operation; the film benchmark (Q01) has its first baseline. |
| M1 First film | The development rooms (ideation, world, characters, story) and every room the pilot needs (art, boards and animatic, layout and camera, animation with lip sync and layers, sets, lighting, the cloth effect, render, edit, voice, sound design, mix, grade, delivery) have studio views with browser coverage, and the Blender workspace (U01–U06, U08, U09) opens every department the pilot needs, with every admitted operation reachable from the web room, from Blender and as an MCP tool (U03); the colour design, props and expert-panel departments serve the pilot (V11, X10, O01, O02, O04); every generative step runs under the candidate-search contract (H16); the animation control catalogue is built into the studio (N20); every cost row records its incentive fields from the first spend (G07, G14); the pilot's evidence checklist records real use of each of these (M01); propose-mode agents run for development, story, camera coverage and layout; the pilot (M01, the same film as F.20.05: 40 seconds, 8 shots, 2 characters) is made through the studio, directed by Elikem, with its authorship ledger and chain of title complete. |
| M2 The short | The crew remakes the pilot on its own and the two are compared with the F-track's agent-made pilots (M02, Q07); a 3-to-8-minute short (M03) is delivered, screened under the Q06 protocol and submitted (M09), with its result reported whatever it is; the department autonomy matrix (Q11) shows each department passing its auto-with-review scenario or the measured gap; the remaining agents (animation, lighting, sound), the costume, graphics, fight-choreography, character-effects, music-editorial and period-research departments (C14, V10, N15, N17, D13, O03), sequence and temporal search for generated shots (H17), character variants, environment physics, the cel-look anime toolkit, in-world motion graphics and field research (C15, X11, V12, V13, O05), authored linework, drawn lighting, art-directed effects, sound-design recipes, dramatic mix control, secondary grading, formative animatic screenings, preservation, the campaign plan and sustainability accounting (V14, V15, X12, D14, D16, K16, Q13, A14, L08, G13), control catalogues for every department (U10), the short's pre-production pack (M10), release windows and premiere status (L09), controlled previews and piracy response (G01, R20), vertical and square versions and rating editions (K17, K18), walla and efforts (D17), ensemble improvisation (C17), the studio through the personal agent (S18), the craft instruments in Blender (U07), the effects-to-comp rooms, the score, the 2D/anime look lane and the hosted video lanes are admitted; the face and body solvers with their gates, the reference-mining libraries, the gait grammar and the face benchmark rung ship (N22, N26, N21, N27, N28, Q14, section 20). |
| M3 The series | Series tooling (bible, season arcs, episode pipeline, continuity across episodes, the adaptation desk and vehicle action, T15 and N18; feathered characters, transformations, the M&E and textless deliverables for buyers, C16, N19, D15 and K15; translated subtitles and dubs wait for Decision 12) ships; the sales desk (L10) is ready for any title Elikem chooses to sell; a pilot episode and a first season (M05, M06) premiere on Lilith TV on demand (L02), and on the V1.4 films-and-series channel (L01) once V1.4's Z02 lands (Decision 18), after Elikem's authorisation; an autonomous episode from the bible passes the same QC, is labelled as autonomous, and is screened beside a directed one (E08, M07); the studio's own face basis, the inverse-rendering solve, dialogue fusion and the stylisation layer ship (C18, N23, N24, N25, section 20). |
| M4 Horizon | A feature, opening the studio to Lilith members and creators, interactive and live formats, game tie-ins, a per-character face basis learned from 4D capture (C19). Not part of V1.5 GA. |
8. The quality bar#
"Compelling and high quality" is measured, not asserted:
- Rule-based craft checks in every department (the F-track QC items: continuity line, exposure, render QA, comp QA, loudness, captions, photosensitivity, final QC) must pass before any judgement is asked.
- Critic panels (Q02) score story, visual, performance, continuity, sound and transition quality, each labelled as model judgement, with DirectorBench-style checkpoints rather than one number. Transition quality between shots is tracked explicitly because it was the weakest dimension across the workflows DirectorBench evaluated (0.256 against 0.71 for prompt fulfilment). Thresholds, held-out cases and comparison rules are fixed before results are seen (Q12).
- Elikem's taste model (Q04) learns from every accept, edit and reject, and predicts his call before he makes it; its agreement rate is reported.
- Blind screenings under a reproducible protocol (Q06: sample size and
power, recruitment, consent, matching criteria, screening rights for the
comparison films, aggregation and uncertainty), reusing V1.4's audience
screenings (
lilith-v1-4:W12): a film reaches the festival bar when a blind panel rates it at least as highly as a matched set of festival-selected human-made shorts of the same length and genre, within the protocol's confidence bounds. That is the target for M2 and M3; a film that falls short is reported as measured, and the exit review decides what follows.
9. Decisions for Elikem#
Elikem accepted every recommendation on 2026-09-29 ("go with your recommendations for the open decisions"). Each row's recommendation below is now the decision; G10 copies them into the decision register. What still waits for him is marked in the task that needs it: amounts (GPU caps, collaborator and counsel budgets, the Opus 5.5 cap), purchases (the C2PA certificate) and engagements (consultants, counsel, reviewers, screening panels). Nothing in this family becomes claimable until the PRD itself is approved. Decision 20 was added by section 18 and approved with revision 32. Elikem revised Decision 12 on 2026-09-29 (section 19). Decision 21 was added by section 20 on 2026-10-05 and decided by Elikem the same day: build the studio's own expression basis and skeletons; no Meshcapade licence. Section 21 (2026-10-06) applies Decision 21 to text-to-motion: the studio trains its own any-skeleton model on cleared data instead of waiting on the licence.
| # | Decision | Options | Decision (recommendation accepted 2026-09-29) | Tasks |
|---|---|---|---|---|
| 1 | Where the studio lives | (a) the Isis studio surface in apps/isis/web, film libraries as libs/isis/film-*, Hathor's ideation and world domains joined as rooms or libraries; (b) a new Lilith Studio app; (c) the Oshun workbenches |
(a). Isis already has the 3D workspace, jobs, lanes, cost ledger and the Blender layer; Hathor already has real ideation and world editing. A new app would duplicate both. G03 confirms it in an ADR. | G03, S01 |
| 2 | Release naming | V1.5 in the Lilith line (LILITH_STUDIO_RELEASE = 'V1.5'), films premiering on Lilith TV; or a builder-side product outside the Lilith line |
V1.5 in the Lilith line, as this PRD's commit already names it. The studio is Elikem's tool; its films are Lilith's programming. | G01, G04 |
| 3 | Frontier crew profiles | Cheap OpenRouter bindings only; plus GPT-6 Astra through Codex on the ChatGPT subscription (already sanctioned; F.00.03 caps it at 20% of each usage window); plus Claude Opus 5.5 through an API key | All three as profiles, cheap by default. Astra through the Codex lane only (never metered). Opus 5.5 default off, enabled per production with a monthly cap Elikem sets, never on a test path. The bake-off (H04) binds each role on evidence. | H04, H05, H06 |
| 4 | Hosted generative video | Admit OpenRouter's video models (Seedance 2.5, Kling 3, Veo 3.1, Wan 3.0, Runway Aleph 2, FLUX video) or stay on self-hosted RunPod workflows only | Admit through OpenRouter only, per production caps, for the shot classes Y01 allows (inserts, establishing shots, effects, backgrounds, restyles of locked 3D shots) and never for a lead's performance unless Y04 passes. Paid avatar products stay off. | Y02, Y03 |
| 5 | The first productions' world | An original world rooted in West African folklore (cleared with Ghana's National Folklore Board where Act 690 requires, R15, and reviewed by paid cultural consultants, R08); the House of Lilith universe from V1.4 (lilith-v1-4:U01); a wholly invented world |
The pilot inside the House of Lilith universe (it already has a bible and a host, and it feeds Lilith TV); the short in an original West-African-rooted world, with paid consultants (R08) and folklore clearance (R15) before its canon is locked, which is distinctive at festivals. | W06, R08, M01, M03 |
| 6 | Voices | Original synthetic voices only; hire human actors for leads under written consent; a mix | Synthetic original voices for M1 and M2; decide on human leads for the series at M3 with a budget, each under a performer permission record (R17). Never a synthetic replica of a real person. | C10, D01, D10 |
| 7 | Autonomous films | Release them labelled on Lilith TV; do not release them; release them without a label | Release labelled as made by Oshun's autonomous studio, on Lilith TV only; never registered for copyright, submitted to festivals or awards as human-authored. This is a policy choice, not a legal necessity (R07 enforces it; Elikem can change it). The UK government has proposed removing s9(3), so their protection may become thin; counsel reviews each title class first (R14). | R07, E08, M02, M04, M07 |
| 8 | Compute | RunPod GPU pods with Flamenco; buy a local GPU workstation; a commercial render farm | RunPod plus Flamenco; no hardware purchase until B09 measures the hours per finished minute. Monthly cap to be set by Elikem. | B06, B09 |
| 9 | Look of the first productions | Stylised 3D (toon or painterly NPR); 2D or anime; photoreal | Stylised 3D first; the 2D/anime lane at M2; no photoreal. Stylised looks hold identity across a season, cost less to render and avoid the uncanny valley. Disclosure duties still apply (R04). | V01, V05, V07 |
| 10 | Audience and rating | Mature (Rockstar-level); general or teen (12+); family or children | Corrected by Elikem on 2026-09-29 (the 12+ recommendation he had accepted was a mistake): a mature audience, like Rockstar Games' titles. 18 by default, 15 where a title's content fits; strong violence, strong language, drug use and dark themes are in scope as craft choices; sexual content follows his recorded boundary (artistic nudity allowed, no explicit sex acts) unless he changes it. Hard lines whatever the rating: nothing sexual involving a minor or a character depicted as one, and no real person's likeness. 18-rated titles sit behind age assurance on Lilith TV and are age-restricted on every outside destination; no title is directed at children. Adult-flagged assets run only on self-hosted workers. | R06, R18, R19, Y01, Y02, Q06, K07, I07 |
| 11 | Series format | 11-minute episodes; 22-minute episodes; other | 11-minute episodes, a 6-episode first season. It is animation's standard short format and keeps the first season affordable. | E02, M06 |
| 12 | Languages | English only; English with subtitles; English plus dubs | Revised by Elikem on 2026-09-29: the films are English for a good while before other languages are explored. English dialogue, English captions and English subtitles for deaf and hard-of-hearing viewers (D08, K08); walla English by default (D17); the M&E and textless elements every buyer needs are still made (D15, K15). Translated subtitles and dubs (the earlier plan: Twi, Ga, Ewe, Hausa and French at M3) wait until he chooses a language; K09 and K13 stay later until then, and any language he opens is reviewed by native speakers and aligned with V1.4's language desks. |
K08, K09, K13, D15, K15, D17 |
| 13 | Blender version | 5.2 LTS pinned per title; follow each release | 5.2 LTS pinned per title; 5.3 features behind flags after its release; retire the stale 4.0.2 pin. | B01, B10 |
| 14 | Human collaborators | None; a small budget line for cultural consultants, a script consultant and a composer review; more | A small budget line, from M1 for cultural consultants if the first world draws on a real culture (Decision 5), otherwise from M2, amounts decided later. No real-culture-inspired record becomes canon without a consultant (R08). | S14, W06, R08 |
| 15 | Opening the studio to others | Horizon after M3 evidence; earlier | Horizon, after M3. | Z02, Z03 |
| 16 | Screenplay authorship on festival titles | Default assist mode with the ledger; the strict human-authored mode (T13) where a destination requires it; T13 on every festival title | Assist by default; T13 switched on where a destination's rules require a human-authored screenplay (the Academy's writing categories do; festival rules vary and are checked per submission in R09). | T13, R09 |
| 17 | C2PA signing certificate | Buy an X.509 signing certificate for C2PA; sign with an in-house key only | Buy a certificate before M2 deliveries, because libs/oshun/c2pa-export is an honest seam without one and in-house manifests are not C2PA. |
R04 |
| 18 | Where the first season premieres | Ask V1.4 to bring its films-and-series channel (lilith-v1-4:Z02, now in V1.4's M4 horizon) forward so it exists before V1.5's M3; premiere through the on-demand library (lilith-v1-4:T08) instead; an outside destination |
Bring lilith-v1-4:Z02 forward, agreed through the ownership map (G02); if it cannot land in time, premiere on demand through T08. The season (M06) and its release schedule (E07) depend only on the on-demand route (L02), so the channel (L01) adds a premiere when it lands and never blocks the season. |
L01, L02, E07, M06 |
| 19 | Legal review | Engage counsel before the first release or registration of each title class; release without review | Engage counsel before the first release of a directed and of an autonomous title (R14), amounts decided later. | R14, R09 |
| 20 | Revenue and incentives | (a) keep money coming in out of V1.5, as V1.4 keeps monetisation out; (b) record incentive fields on every cost from the first spend, keep an income record per title, and add a sales desk at M3; (c) a commercial studio now (financing plans, pre-sales, a sales agent from M1) | Recommended, for Elikem's approval with section 18: (b). The fields cost little now and cannot be rebuilt later. Whether the UK Audio-Visual Expenditure Credit (39% of qualifying expenditure for animation, where qualifying expenditure is the lower of UK core expenditure and 80% of total core expenditure; British certification; at least 10% UK core spend; at least 51% animation spend) covers this kind of production is for a tax adviser and the BFI, so nothing is planned on it until they answer. Selling stays optional per title, and every deal is his. | G07, G14, L10 |
| 21 | Face and body models under licence | (a) buy the Meshcapade commercial licence for FLAME and the SMPL family (Isis T.12.07) now; (b) build without it: the studio's own expression basis on ICT FaceKit or an original template (C18) and the studio's own skeletons (T.11.03) for the body solver (N26), buying only on a named failure of that path; (c) stay on the 52-shape set and library skeletons and skip C18 | (b), decided by Elikem on 2026-10-05 ("build own"). The 52-shape set and the viseme set stay as export projections, so nothing built before C18 is lost; FLAME and the SMPL family are refused under R21 by decision, not by default, and the licence is reconsidered only against a named failure of the C18 or N26 path, by revising this decision. | R21, C18, N26, N23, N12 |
10. Risks#
| Risk | Why it is real | Mitigation in this plan |
|---|---|---|
| Scope against one person | 318 tasks, many waiting on Isis F and T items that are not built. | Films as milestones (the pilot proves the minimum set); V1.5 depends on the F-track rather than duplicating it, which moves the right F items up the queue; the dial lets agents carry departments Elikem is not working in. |
| The F-track is the critical path | Only 3 of 130 F items are shipped. Many V1.5 rooms are interfaces over F primitives. | G02 names the F items V1.5 needs first; the board dependencies make those the only way forward, so product workers pick them before V1.5 rooms. |
| Generative quality ceiling | Identity drift and weak transitions between shots are the documented failure of 2026 video systems (DirectorBench: transitions 0.256). | Blender as ground truth; generated finishes gated on fidelity and identity (Y04); transitions scored explicitly (Q02). |
| Taste cannot be automated | On LitBench's literary task, story judges agreed with human preference 73–78% of the time; no benchmark shows a model judge can stand in for taste on film. | Critics advise, Elikem decides; the taste model ranks but never gates alone (Q04); blind screenings are the bar (Q06). |
| Cost overruns | GPU renders, hosted video at $0.04–0.13 per second, long agent runs (Opus 5.5 used ~1.6x output tokens of Opus 5 on film tasks). | Estimates before spend (B07), caps per production, department and run (G07), the meter (S11), cheap bindings by default. |
| Vendor churn | A secondary source reports Sora 2's API was scheduled to end on 2026-09-24 (OpenRouter still listed it on 2026-09-29); licence and training terms change. | The terms register (R05), multi-lane routing (Y09), the field watch (Q10), self-hosted defaults for restyling. |
| Legal uncertainty | US and UK authorship law is unsettled for AI-assisted works; EU Art. 50 applies to deployers. | The authorship ledger (R01), strict human-authored screenplay mode (T13), labelling of autonomous work (R07), C2PA and disclosure (R04). |
| Cultural misrepresentation | A West-African-rooted world made without the people it draws on would be both wrong and harmful. | The consultation workflow (R08) blocks canon until a consultant has reviewed it; Decision 14 funds it. |
| Agent-run code in Blender | Agent-written bpy can reach the filesystem or network. | The contained script lane (B05), crew safety tests (H14), the hardened runner flags. |
| The Mac | The machine has frozen under stacked load (three diagnosed freezes in CLAUDE.md). | Heavy work runs on RunPod and the Linux box; the live link enforces the Mac-safe mode (B11). |
| Astra allowance | The Codex lane may use 20% of each window (F.00.03) and stops when the allowance ends. | The governor (F.01.09) and cheap fallbacks that never lower a gate (H12). |
| Traps in the repository | Many modules look real and are planners or scaffolds (section 2.2). | Paths in task bodies name the traps; workers park rather than wire an unmeasured component. |
| Losing a festival's eligibility | One public upload, a long trailer or an on-demand release before a festival or qualifying run can end the short's festival and awards route | The release plan and its exhibition journal checked at every publication (L09); screeners only as previews through R20 |
| Leaks of unreleased cuts | Screeners go to festivals, consultants, freelancers, panels, press and buyers | Per-recipient marks, expiring links and access logs (R20) |
11. Board map#
Keys are lilith-v1-5:<id>. Dependencies in other families are written in full.
G — Programme, capability gate and production office (lilith-v1-5:G)
| Task | M | Effort | Depends on |
|---|---|---|---|
G01 V1.5 studio capability gate |
M0 | S | — |
G02 Ownership map across the film, 3D, creative-studio and channel families |
M0 | M | — |
G03 Studio architecture ADR |
M0 | M | G02 · outside: Elikem approves the ADR |
G04 V1.5 rows in the release tables and feature map |
M0 | S | — · outside: this PRD approved |
G05 V1.5 page inventory and walkthrough plan |
M0 | S | G01, G03 |
G06 Production office: slate, greenlight, calendar and budgets |
M1 | S | A01, lilith-v1-4:G06, G02 |
G07 Cost ledger and per-production spend caps |
M0 | M | isis-chroma-runpod:F.01.04, A03 |
G08 Milestone exit review runner |
M0 | S | G01 |
G09 Studio handbook for one person |
M1 | S | S01 |
G10 V1.5 decision register |
M0 | S | — |
G11 Publication authorisation |
M0 | M | G01, A02, A05, G10 |
G12 Production schedule, bids and throughput |
M1 | M | G06, S03, B07, H01, G07 |
G13 Production sustainability accounting |
M2 | S | G07, B09, G12, R10 |
G14 Incentives and revenue record |
M1 | S | G07, G06, G10 · outside: a tax adviser Elikem engages, and BFI certification, before any claim is planned on (Decision 20) |
A — Production graph and the operation core (lilith-v1-5:A)
| Task | M | Effort | Depends on |
|---|---|---|---|
A01 Production graph schema |
M0 | L | G03 |
A02 Versions, branches and diffs for every record |
M0 | M | A01 |
A03 Typed operation API — one operation set for people and agents |
M0 | L | A01, A02, isis-chroma-runpod:F.02.05 |
A04 MCP server over the operation API |
M0 | M | A03, A05 |
A05 Locks and approvals |
M0 | M | A03, isis-chroma-runpod:F.03.02 |
A06 Lineage, staleness and selective regeneration |
M1 | M | A02, isis-chroma-runpod:F.03.06, content-workbenches:CR38, G02 |
A07 Media store and proxies |
M0 | M | G03 |
A08 Event stream and notifications |
M1 | S | A03 |
A09 Interchange: Fountain, FDX, OTIO, USD, glTF, FBX, Alembic |
M2 | M | A01, isis-chroma-runpod:F.17.07 |
A10 Search across the production |
M1 | S | A01 |
A11 Archive and rebuild drill |
M2 | S | A07, isis-chroma-runpod:F.19.06 |
A12 Native-edit coverage |
M1 | M | A03, A05, B04, B12, S08 |
A13 Production restore drill |
M1 | M | A02, A07, B12 · outside: platform-ops backs up the studio's database and media stores |
A14 Preservation maintenance |
M2 | S | A07, A11, A13, B01, R05, R10 |
S — The studio workspace (the director's chair) (lilith-v1-5:S)
| Task | M | Effort | Depends on |
|---|---|---|---|
S01 Studio shell and navigation |
M0 | M | G03, G05 |
S02 "What needs me" home |
M1 | M | A08, S01 |
S03 Production tracker |
M1 | M | A01, S01, isis-chroma-runpod:F.20.03 |
S04 Review room with frame-accurate notes |
M1 | M | A07, isis-chroma-runpod:F.03.04, content-workbenches:CR13, G02 |
S05 Overnight run report |
M1 | S | H07, S02 |
S06 Agent console |
M1 | M | H01, H11, S01 |
S07 Director chat at any scope |
M1 | M | A02, isis-chroma-runpod:F.20.04 |
S08 Blender live link |
M1 | L | B04, A05, A03 |
S09 Mobile and tablet review |
M2 | M | S04 |
S10 Autonomy dial controls |
M1 | S | H08, Q05 |
S11 Spend and compute meter |
M1 | S | G07, B07 |
S12 Studio accessibility and performance budget |
M2 | S | S01 |
S13 Studio end-to-end and walkthrough coverage |
M2 | M | S03, S04, K01 |
S14 Guest collaborators |
M2 | S | G01, A05, R10, R13, R20 · outside: collaborator amounts (Decision 14, accepted 2026-09-29; amounts decided later) |
S15 Craft mentor |
M2 | S | H09, Q02 |
S16 Live-link offline recovery |
M1 | M | S08, B12, A02, A05 |
S17 Pilot journey in the browser |
M1 | M | S03, S04, T04, P02, K01 |
S18 The studio through the personal agent |
M2 | S | A04, S02, S05, S07, G01, R01, I01, lilith-v1-3:C09, lilith-v1-3:T09 |
B — Blender service, render farm and compute (lilith-v1-5:B)
| Task | M | Effort | Depends on |
|---|---|---|---|
B01 Blender version policy and version-matched knowledge |
M0 | S | isis-chroma-runpod:F.02.02, G03 |
B02 Headless Blender service |
M0 | M | G03, isis-chroma-runpod:F.02.05, B13 |
B03 Film skills as studio operations |
M0 | M | A03, isis-chroma-runpod:F.02.01 |
B04 Live session broker |
M1 | M | B02, A05 |
B05 Contained script lane |
M1 | S | isis-chroma-runpod:F.02.06, isis-chroma-runpod:F.02.08, G03 |
B06 Render farm on Flamenco |
M1 | M | B02, isis-chroma-runpod:F.14.04 · outside: platform-ops installs and runs the Flamenco manager service on the Linux dev server (filed by this task; product work consumes its evidence) and Decision 8 sets the GPU cap |
B07 Render estimates and budget holds |
M1 | S | B06, G07, isis-chroma-runpod:F.14.05, Q12 |
B08 Asset library and file layout |
M0 | M | A07, isis-chroma-runpod:F.03.03 |
B09 Compute capacity plan |
M1 | S | B06, M01 · outside: Elikem sets the monthly GPU cap (Decision 8, accepted 2026-09-29: RunPod and Flamenco, no hardware purchase until this plan measures) |
B10 Blender 5.3 adoption |
M2 | S | B01 · outside: Blender 5.3 released (expected November 2026) |
B11 Mac-safe local mode |
M1 | S | B04 |
B12 .blend versioning and scene diff |
M1 | M | A02, B02 |
B13 Resource admission |
M0 | M | G03, G07 · outside: numeric GPU and spend limits set by Elikem before any paid probe (Decision 8, accepted 2026-09-29) |
B14 Blender feature adoption map and 5.2 API migration |
M1 | S | B01, Q10, B03, A12 |
I — Ideation and the development slate (lilith-v1-5:I)
| Task | M | Effort | Depends on |
|---|---|---|---|
I01 Idea inbox |
M1 | S | A01 |
I02 Brainstorm sessions |
M1 | M | I01, content-workbenches:CR04, content-workbenches:CR05, G02 |
I03 Premise and logline lab |
M1 | S | I02, content-workbenches:CR06, G02 |
I04 Comparables and collision check |
M1 | S | I03, content-workbenches:CR19, G02 |
I05 Pitch documents and sizzle |
M1 | S | I03, V02 |
I06 Development slate and greenlight |
M1 | S | G06, I05 |
I07 Format designer |
M1 | S | I06 |
I08 Taste and influences board |
M1 | S | I01 |
I09 Autonomous development run |
M2 | M | I03, I05, H07 |
W — World building (lilith-v1-5:W)
| Task | M | Effort | Depends on |
|---|---|---|---|
W01 World bible model |
M1 | M | A01, lilith-v1-4:U02, G02 |
W02 Canon and contradiction engine |
M1 | M | W01, lilith-v1-4:U03, lilith-v1-4:U05, G02 |
W03 World state across story time |
M1 | M | W02, T06 |
W04 Maps and terrain |
M1 | M | W01, B03 |
W05 Timeline editor |
M1 | S | W01 |
W06 Cultures and societies designer |
M1 | M | W01, R08, R15 · outside: consultants engaged (Decision 14, accepted 2026-09-29: from M1 for a real-culture-inspired world) and folklore permissions where R15 requires them |
W07 Languages, naming and pronunciation |
M2 | M | W06, lilith-v1-4:V05, G02 |
W08 Systems designer: magic, technology and economy |
M2 | S | W02 |
W09 World visualisation and splat worlds |
M2 | M | W04, V02, Y07 · outside: a splat-world vendor clears the rights register, and Blender 5.3 for native splats |
W10 Lore Q&A with citations |
M1 | S | W02, A10 |
W11 World and series bible export |
M2 | S | W01 |
W12 Autonomous world-building run |
M2 | M | W02, H07 |
C — Characters (lilith-v1-5:C)
| Task | M | Effort | Depends on |
|---|---|---|---|
C01 Character record |
M1 | M | W01, content-workbenches:CR07, G02 |
C02 Relationships and ensemble dynamics |
M1 | S | C01, lilith-v1-4:C04, G02 |
C03 Character interview |
M1 | S | C01 |
C04 Design exploration |
M1 | M | C01, V01, Y07 |
C05 Model sheet and turnaround pack |
M1 | M | C04, content-workbenches:CR08, G02 |
C06 3D character build |
M1 | L | C05, B03 · outside: Isis T.05.01 (TRELLIS.2 adapter) and T.11.01 (learned skinning) admitted, or a hand-built path |
C07 Face rig standard |
M1 | M | C06 |
C08 Wardrobe, hair and props per character |
M2 | M | C06, isis-chroma-runpod:F.09.03, isis-chroma-runpod:F.09.04 |
C09 Acting profile |
M1 | S | C01, C06 |
C10 Voice casting |
M1 | M | C01, D01, lilith-v1-4:V02, lilith-v1-4:C03, R17, G02 |
C11 On-model checker |
M1 | M | C05, content-workbenches:CR10, Q12, G02 |
C12 Screen test |
M1 | S | C07, C10, N02, isis-chroma-runpod:F.09.02 |
C13 Autonomous character run |
M2 | M | C05, H07 |
C14 Costume, hair and makeup department |
M2 | L | C08, C05, V11, R15, O01, isis-chroma-runpod:F.13.02, U02, Q12 |
C15 Character variants across time, bodies and forms |
M2 | M | C06, C07, C09, C10, C11, T06, A02 |
C16 Feathered-character qualification |
M3 | M | C06, N03, N17, V05, isis-chroma-runpod:F.09.04, isis-chroma-runpod:T.11.02 |
C17 Ensemble improvisation |
M2 | S | C02, C03, C09, T03, T06, H03 |
C18 Oshun face basis: the studio's own expression set |
M3 | L | C07, R21, C06, Q12 · outside: an original template head (a commissioned sculpt) if R21 refuses ICT FaceKit |
C19 Per-character learned basis from 4D capture |
M4 | L | C18, N23, R17, R03 · outside: a multi-view capture setup or a performer session, if a human source is chosen over the character's own renders |
T — Story and screenplay (lilith-v1-5:T)
| Task | M | Effort | Depends on |
|---|---|---|---|
T01 Story structure toolkit |
M1 | M | A01, content-workbenches:CR25, G02 |
T02 Outline, treatment and step outline |
M1 | M | T01 |
T03 Scene designer |
M1 | S | T02 |
T04 Screenplay editor |
M1 | L | isis-chroma-runpod:F.04.01, R01, content-workbenches:CR31, G02 |
T05 Dialogue and voice pass |
M1 | M | T04, C01 |
T06 Story state tracking |
M1 | M | isis-chroma-runpod:F.04.05, G03 |
T07 Script coverage by a critic panel |
M1 | M | T04, Q02 |
T08 Voiced table read |
M1 | S | T04, D02, content-workbenches:CR31, G02 |
T09 Breakdown and shot list handoff |
M1 | S | isis-chroma-runpod:F.04.02, isis-chroma-runpod:F.04.04, G03, A02, A03, A06, T06, T10 |
T10 Script lock and change control |
M1 | S | A05, A06, T04 |
T11 Themes, motifs, setups and payoffs |
M2 | S | T06 |
T12 Autonomous writers' room |
M2 | M | T07, H03, lilith-v1-4:W06, G02 |
T13 Human-authored screenplay mode |
M1 | S | T04, R01 |
T14 Research desk |
M1 | S | A10 |
T15 Adaptation desk |
M3 | M | W02, C01, T02, E02, R02, R16, A06 |
E — Series and episodic production (lilith-v1-5:E)
| Task | M | Effort | Depends on |
|---|---|---|---|
E01 Show bible |
M3 | M | W11, content-workbenches:CR25, lilith-v1-4:W04, G02 |
E02 Season arc planner |
M3 | M | E01, T01, lilith-v1-4:W07, G02 |
E03 Episode breaking board |
M3 | M | E02 |
E04 Continuity across episodes |
M3 | M | T06, W03 |
E05 Episode pipeline at scale |
M3 | L | A01, B06, H07, B13 |
E06 Recurring elements |
M3 | S | K05 |
E07 Release cadence and scheduling |
M3 | S | L02 |
E08 Autonomous episode from the bible |
M3 | M | E03, E05, Q05, H13, K07, R07 |
E09 Series dashboard |
M3 | S | S03 |
V — Visual development and art direction (lilith-v1-5:V)
| Task | M | Effort | Depends on |
|---|---|---|---|
V01 Art direction brief and style bible |
M1 | M | isis-chroma-runpod:F.04.06, content-workbenches:CR09, G03, G02 |
V02 Concept art room |
M1 | M | V01, Y07, content-workbenches:CR26, G02 |
V03 Colour script |
M1 | S | V01, isis-chroma-runpod:F.12.05 |
V04 Style frames and key art |
M1 | S | V02, B03 |
V05 Look-development kits |
M1 | M | isis-chroma-runpod:F.12.04, G03 |
V06 Show-specific style models |
M2 | M | V02, R05 |
V07 2D and anime look lane |
M2 | L | V05, N09 |
V08 Style conformance scorer |
M2 | S | V01, content-workbenches:CR29, G02 |
V09 Art department review |
M1 | S | S04 |
V10 In-world graphic design |
M2 | M | V01, W06, W07, R05, B08, O03 |
V11 Colour design |
M1 | M | V01, V03, C05, isis-chroma-runpod:F.14.03 |
V12 Cel-look 3D anime toolkit |
M2 | L | V05, V11, C07, N07 |
V13 In-world motion graphics, holograms and HUDs |
M2 | M | V10, Y06, X06, W06, W07 |
V14 Authored animated linework |
M2 | M | V05, C07, N07, N13, A12 |
V15 Drawn lighting and texture tracking |
M2 | M | N09, V11, X06, A06, U07 |
P — Boards, previs, layout and cinematography (lilith-v1-5:P)
| Task | M | Effort | Depends on |
|---|---|---|---|
P01 Boards room |
M1 | M | isis-chroma-runpod:F.05.01, isis-chroma-runpod:F.05.02, S04, T09 |
P02 Animatic editor |
M1 | M | isis-chroma-runpod:F.05.03, content-workbenches:CR32, T08, G02 |
P03 Layout room |
M1 | M | isis-chroma-runpod:F.06.02, isis-chroma-runpod:F.06.03, G03 |
P04 The film's camera language |
M1 | S | isis-chroma-runpod:F.07.02, isis-chroma-runpod:F.07.05, G03 |
P05 Coverage and shot-design agent |
M1 | M | isis-chroma-runpod:F.07.06, isis-chroma-runpod:F.06.04, H03, P04 |
P06 Previs review |
M1 | S | isis-chroma-runpod:F.06.05, S04 |
P07 Virtual camera |
M2 | M | B04 · outside: a phone or tablet (Elikem's device) |
P08 Boards-to-layout handoff |
M1 | S | isis-chroma-runpod:F.05.05, G03, P01, P03 |
P09 Stereoscopic direction (opt-in per title) |
M4 | M | P04, X05, B07, H03, Q12, K07 |
N — Animation and performance (lilith-v1-5:N)
| Task | M | Effort | Depends on |
|---|---|---|---|
N01 Performance plan editor |
M1 | M | isis-chroma-runpod:F.10.01, G03 |
N02 Blocking to spline with pose libraries |
M1 | M | isis-chroma-runpod:F.10.02, C09 |
N03 Motion library and retargeting |
M1 | S | C06 |
N04 Video performance capture |
M2 | M | N03, R10, R17, N13, D03, S04, isis-chroma-runpod:T.13.06 · outside: an admitted single-camera body-capture model or vendor that clears the rights register; a phone and a capture space (Elikem) |
N05 Facial capture from a phone |
M2 | M | C07, R10, R17, N13, D03 · outside: an iPhone with a TrueDepth camera, or the non-iPhone landmark path (Elikem's device) |
N06 Audio-driven lip sync and emotion |
M1 | M | isis-chroma-runpod:F.11.03, C07, isis-chroma-runpod:F.11.04, isis-chroma-runpod:F.11.05 |
N07 Stylised animation principles kit |
M2 | M | N02, isis-chroma-runpod:F.10.05 |
N08 Interaction and contact |
M2 | S | isis-chroma-runpod:F.10.04, G03 |
N09 Grease Pencil 2D animation |
M2 | L | V05, V11, A12, U07 · optional: an admitted inbetween and colour model for assisted ranges |
N10 Crowds and background action |
M2 | S | isis-chroma-runpod:F.09.05, G03, N03, N08, X01, H03, B13, C15 |
N11 Animation agent |
M2 | M | N02, H03, Q04 |
N12 Text-to-motion on a cleared model |
M3 | M | N03 · outside: Isis T.12.10 admitted (section 21), or Isis T.12.01/T.12.02 admitted and the T.12.07 licence obtained |
N13 Animation layers and handoff |
M1 | M | N02, C07, A02, S08 |
N14 Learned in-betweening from blocking |
M2 | M | N02, N13, R05, Q12 |
N15 Fight and action choreography |
M2 | L | N02, N08, N03, P05, X10, O01, R05, U02 |
N16 Dance and musical staging |
M3 | M | D06, D11, N02, O01, R08 |
N17 Character-effects shot finaling |
M2 | M | N13, C14, X04, A12, H03 |
N18 Vehicle and ship action |
M3 | L | X10, N03, P05, X04, X11, H03, B13, Q12 |
N19 Transformation performance |
M3 | M | C15, C07, N08, N13, X04, A12 |
N20 Animation control parity |
M1 | M | U03, N13, N01, N02, isis-chroma-runpod:F.10.04, A12 |
N21 Expression library mined from generated reference |
M2 | M | N22, Y02, Y11, N02, C09, N01, R03, Y01, U08 |
N22 Face solver on the open stack, with a re-render gate |
M2 | M | R21, C07, N13, B02, Q12, R10, R17 |
N23 Inverse-rendering solve in the character's own basis |
M3 | L | C18, N22, Q14, B13, B09 |
N24 Audio and video fusion for dialogue takes |
M3 | M | N06, N22, D03, N13 |
N25 Acting-language stylisation layer |
M3 | M | N07, N21, C18, C09, N13, H10 |
N26 Body solver with a physics-tracking gate |
M2 | L | R21, N04, N03, N13, X11, B13, Q12, Q01, R10, R17 |
N27 Motion-phrase library mined from generated reference |
M2 | M | N26, N28, Y02, Y11, N02, N03, N01, R03, Y01, U08 |
N28 Gait-grammar gate per species |
M2 | S | N03, O01, Q12, N07 |
X — Sets, lighting, effects, rendering and compositing (lilith-v1-5:X)
| Task | M | Effort | Depends on |
|---|---|---|---|
X01 Set building room |
M1 | M | isis-chroma-runpod:F.08.01, isis-chroma-runpod:F.08.02, G03 |
X02 Location library |
M1 | S | X01, B08 |
X03 Lighting room |
M1 | M | isis-chroma-runpod:F.12.01, isis-chroma-runpod:F.12.03, isis-chroma-runpod:F.15.02, isis-chroma-runpod:F.12.02, G03 |
X04 Effects room |
M1 | M | isis-chroma-runpod:F.13.01, isis-chroma-runpod:F.13.06, G03 |
X05 Render management |
M1 | M | isis-chroma-runpod:F.14.06, isis-chroma-runpod:F.14.07, B06, isis-chroma-runpod:F.14.03, B07 |
X06 Compositing room |
M2 | M | isis-chroma-runpod:F.15.01, isis-chroma-runpod:F.15.05, G03 |
X07 Splat and matte worlds |
M2 | M | isis-chroma-runpod:F.08.06, W09 · outside: Blender 5.3 released for native splats |
X08 Lighting agent |
M2 | M | X03, V03, H03 |
X09 Live look preview (optional) |
M2 | S | S08, V04, B13, R05, G07, R10 |
X10 Props, vehicles and weapons department |
M1 | M | X01, T09, isis-chroma-runpod:F.04.03, B08, T06 |
X11 Environment physics profiles |
M2 | M | W08, X04, T06, isis-chroma-runpod:F.13.01 |
X12 Art-directed effects animation |
M2 | M | X04, V05, N07, X06, A12 |
Y — Generative finish and the hosted media lanes (lilith-v1-5:Y)
| Task | M | Effort | Depends on |
|---|---|---|---|
Y01 Generative-lane policy |
M1 | S | R01, R06 |
Y02 OpenRouter video lane |
M1 | M | G07, R05, R10, R13, H15 |
Y03 Render-to-video restyle |
M2 | M | isis-chroma-runpod:F.16.01, isis-chroma-runpod:F.16.02, Y02 |
Y04 Fidelity and identity gate |
M2 | M | isis-chroma-runpod:F.16.04, C11, Q12, P04 |
Y05 Long shots |
M2 | S | isis-chroma-runpod:F.16.05, G03, H09, A06, Y04 |
Y06 Code-as-shot lane |
M2 | M | A03, B02, B05, H14, K01, A12 |
Y07 Image lanes for design |
M1 | S | G07, R05, R10, R13, H15 |
Y08 Upscale and frame interpolation |
M2 | S | Y02, N07 |
Y09 Generative-lane bake-off |
M2 | M | Y02, Y03, Q02 |
Y10 Local generative repair |
M2 | M | Y03, Y04, A05 |
Y11 Character identity adapters from the character's own renders |
M2 | M | C06, isis-chroma-runpod:F.09.02, V06, Y03, C11, Q12, R05, Y04 |
D — Voice, music and sound (lilith-v1-5:D)
| Task | M | Effort | Depends on |
|---|---|---|---|
D01 Voice lane with acting range |
M1 | M | lilith-v1-4:V01, G03, G02 |
D02 Voice direction and take comping |
M1 | M | D01, isis-chroma-runpod:F.11.01, content-workbenches:CR28, G02 |
D03 Alignment and timing |
M1 | S | isis-chroma-runpod:F.11.02, G03 |
D04 Sound design room |
M1 | M | isis-chroma-runpod:F.18.01, isis-chroma-runpod:F.18.03, G03 |
D05 Score and themes |
M2 | L | lilith-v1-4:M01, content-workbenches:CR30, G03, G02 · outside: music lanes admitted by lilith-v1-4:M01's licence gate; a human composer review if Decision 14 funds one |
D06 Songs |
M3 | M | D05, D11, D13, D03, N06, C07, R11 |
D07 Mix and loudness |
M1 | M | isis-chroma-runpod:F.18.04, G03 |
D08 Audio QA and captions |
M1 | S | isis-chroma-runpod:F.18.05, isis-chroma-runpod:F.18.06, G03 |
D09 Sound agent |
M2 | M | D04, H03 |
D10 Recording booth |
M2 | S | S14, R13, R10, R17 |
D11 Editable score representation |
M2 | M | D05, A03, K01 |
D12 Direct a line by acting it: performance-to-voice |
M2 | M | D01, D02, D10, C10, R17, R10, R03 |
D13 Music editorial |
M2 | M | D11, K04, R11, H03 |
D14 Editable sound-design construction |
M2 | M | D04, A02, A03, A07, K04, R11 |
D15 International sound versioning and M&E |
M3 | M | D04, D07, K04 |
D16 Dramatic mix control |
M2 | M | D07, D04, K04, A03 |
D17 Walla, loop group and efforts |
M2 | S | D01, D02, D03, D04, D10, R17, W07, N08, O01 |
K — Editorial, finishing, localisation and delivery (lilith-v1-5:K)
| Task | M | Effort | Depends on |
|---|---|---|---|
K01 Edit room |
M1 | L | isis-chroma-runpod:F.17.01, isis-chroma-runpod:F.17.02, isis-chroma-runpod:F.17.05, G03 |
K02 Assembly and alternate cuts |
M1 | M | isis-chroma-runpod:F.17.04, K01 |
K03 Cut review |
M1 | S | isis-chroma-runpod:F.17.06, Q02 |
K04 Conform |
M2 | S | isis-chroma-runpod:F.17.08, G03 |
K05 Titles, credits and the AI disclosure |
M2 | S | isis-chroma-runpod:F.19.03, R01 |
K06 Grade room |
M1 | M | isis-chroma-runpod:F.19.01, isis-chroma-runpod:F.19.02, G03 |
K07 Deliverables and final QC |
M1 | M | isis-chroma-runpod:F.19.04, isis-chroma-runpod:F.19.05, R11, R16, R18 |
K08 English subtitles (translation deferred) |
M2 | S | D08 |
K09 Dubbing with re-keyed lip sync (deferred) |
M3 | L | D02, N06, K08, lilith-v1-4:V06, K13, G02, D15 · outside: native-speaker reviewers engaged (Elikem); the languages are chosen by Elikem when he opens Decision 12 again; later until then (section 19) |
K10 Audio description and accessible delivery |
M2 | M | K08, content-workbenches:CR15, G02, D07, K04, O01 |
K11 Trailers, teasers and promos |
M2 | M | K02, lilith-v1-4:W10, G02 |
K12 Posters, thumbnails and key art |
M2 | S | V04, lilith-v1-4:D06, G02 |
K13 Language qualification (deferred) |
M3 | M | D03, W07, N06, lilith-v1-4:V06, G02 · outside: native-speaker reviewers per language (Elikem engages); later until Decision 12 opens a language (section 19) |
K14 Festival exhibition copy: DCP and cinema sound |
M2 | M | K07, D07, G02 |
K15 Textless elements and forced narratives (locale variants deferred) |
M3 | M | V10, V13, K05, K08, K04, R05, A06 |
K16 Secondary grading |
M2 | M | K06, X06, K04, A12 |
K17 Vertical and square versions re-staged in 3D |
M2 | M | K02, K07, K11, P04, X05, B07, A02, A06, Y03, Y04 |
K18 Versions and editions: rating, platform and runtime edits |
M2 | M | K02, K04, K07, R06, R18, R19, A02, D02 |
H — The agent crew and autonomy (lilith-v1-5:H)
| Task | M | Effort | Depends on |
|---|---|---|---|
H01 Crew runtime |
M0 | L | G03, A04, isis-chroma-runpod:F.01.03, G07, B03, H15 |
H02 Crew design: roles and descriptors |
M0 | M | H01 |
H03 Propose-mode protocol |
M0 | M | A02, H01 |
H04 Controller bake-off per role |
M1 | M | isis-chroma-runpod:F.01.06, Q01, H11 · outside: a frontier leg runs only after Elikem enables that profile for a production with a cap (Decision 3, accepted 2026-09-29) |
H05 Codex and Astra lane |
M1 | M | isis-chroma-runpod:F.01.08, isis-chroma-runpod:F.01.09, G03 |
H06 Opus 5.5 profile |
M1 | S | G07, H01 · outside: Elikem enables the profile per production and sets its cap (Decision 3, accepted 2026-09-29: default off, never on a test path) |
H07 Overnight and long-horizon runs |
M1 | M | H01, G07, H15 |
H08 Autonomy dial engine |
M1 | M | H03, A05 |
H09 Production memory |
M1 | M | A10, Q04, A06, A07 |
H10 Self-improvement from trajectories |
M2 | M | H07, Q01, Q09, lilith-v1-4:H03, G02 |
H11 Traces, replay and cost attribution |
M0 | M | H01, lilith-v1-4:H05, G02 |
H12 Failure recovery and escalation |
M1 | S | H07 |
H13 End-to-end autonomous production |
M2 | L | H08, Q05, I09, W12, C13, T12, C06, P05, N11, X05, X08, D09, K02, K07, R07, G07, Y09, H16, H17 |
H14 Crew safety with untrusted content |
M1 | S | B05 |
H15 Durable job reconciliation |
M0 | M | A03, A07, isis-chroma-runpod:F.02.05 |
H16 Candidate search with verifiers |
M1 | M | H01, Q02, Q12, G07 |
H17 Sequence and temporal search for generated shots |
M2 | M | H16, Y03, Y05, Q12 |
Q — Quality, evaluation and taste (lilith-v1-5:Q)
| Task | M | Effort | Depends on |
|---|---|---|---|
Q01 Film benchmark ladder |
M0 | M | isis-chroma-runpod:F.20.01, isis-chroma-runpod:F.20.02, Q12 |
Q02 Critic panels |
M1 | M | Q01 |
Q03 Continuity verification after each shot |
M1 | M | T06, W03 |
Q04 Elikem's taste model |
M1 | M | H03, content-workbenches:CR16, G02 |
Q05 Autonomy evidence gate |
M1 | S | Q01, H08 |
Q06 Blind screening protocol and panel |
M2 | M | K07, lilith-v1-4:W12, R10, Q12, G02, R20 · outside: invited viewers recruited, participant consent, and screening rights for the comparison films (Elikem's acts and spend) |
Q07 Directed versus autonomous comparison |
M2 | S | H13, Q06, M01, M02 |
Q08 Capability matrix |
M2 | S | isis-chroma-runpod:F.20.06, G03 |
Q09 Quality regression suite |
M2 | S | Q01 |
Q10 Field watch |
M1 | S | isis-chroma-runpod:F.20.07, G03 |
Q11 Department autonomy acceptance matrix |
M2 | M | Q01, Q05, Q09, P05, N11, X08, D09, K02, T12, I09, W12, C13, Y09, Q12, O02 |
Q12 Frozen evaluation protocol |
M0 | M | G10, isis-chroma-runpod:F.20.01 · outside: Elikem approves the thresholds |
Q13 Formative animatic screenings |
M2 | M | P02, S04, R10, Q12, content-workbenches:CR23, lilith-v1-4:W12, R20 |
Q14 Face performance benchmark rung |
M2 | M | Q01, Q12, N05, N06, N21, Q09 |
R — Rights, authorship, provenance and standards (lilith-v1-5:R)
| Task | M | Effort | Depends on |
|---|---|---|---|
R01 Authorship ledger |
M0 | M | A03 |
R02 Chain of title |
M1 | M | R05, A06 |
R03 Originality and likeness screen |
M1 | M | content-workbenches:CR19, G03, G02 |
R04 Provenance and disclosure |
M2 | M | K07, lilith-v1-4:F06, G02, R01, H11, A06, R10 · outside: the C2PA signing certificate bought before M2 deliveries (Decision 17, accepted 2026-09-29) |
R05 Rights and licence register |
M0 | S | G02, lilith-v1-4:S03, isis-chroma-runpod:F.01.01 |
R06 Film standards and ratings |
M1 | S | G10 |
R07 Labelling autonomous work |
M2 | S | R01, H08 |
R08 Cultural consultation |
M1 | S | G10 · outside: consultants engaged (Decision 14, accepted 2026-09-29: from M1, because Decision 5 chose a West-African-rooted world for the short) |
R09 Festival, awards and copyright pack |
M2 | S | R01, T13, R14 |
R10 Production data handling |
M1 | S | R05 |
R11 Music and sample rights |
M1 | S | R05, D04 |
R12 Title and trademark clearance |
M2 | S | I06 |
R13 Asset and reference intake gate |
M0 | M | R05, A07, G01 |
R14 Legal review before release |
M2 | S | R02, R04, R07 · outside: counsel engaged (Decision 19, accepted 2026-09-29; amounts decided later) |
R15 Folklore and source clearance |
M1 | M | R05, W01, G10 · outside: permissions and fees where a source requires them (Elikem) |
R16 Use-specific clearance |
M1 | M | R02, R05, lilith-v1-4:S03, G02 |
R17 Performer permissions |
M1 | M | R05, R06 |
R18 Film-output safety integration |
M1 | M | R03, R06, lilith-v1-4:S02, lilith-v1-4:S10, lilith-v1-4:S11, G02 |
R19 Audience-age policy integration |
M1 | M | R06, lilith-v1-4:S07, lilith-v1-4:R01, lilith-v1-4:R11, G02 |
R20 Controlled previews and piracy response |
M2 | M | A07, G01, G11, R05, R10, R16, Q12 · outside: YouTube's decisions on Content ID and Copyright Match Tool access for Elikem's account |
R21 Face and motion stack licence admission |
M1 | S | R05, B13 · outside: a Meshcapade commercial licence for FLAME and the SMPL family (Isis T.12.07, Decision 21); without it both are refused |
L — Distribution, premieres and audience (lilith-v1-5:L)
| Task | M | Effort | Depends on |
|---|---|---|---|
L01 Films and series channel on Lilith TV |
M3 | M | lilith-v1-4:Z02, G11, G02 · outside: lilith-v1-4:Z02 brought forward through G02 (Decision 18, accepted 2026-09-29); if it cannot land in time, the season premieres on demand through L02, which M06 and E07 depend on instead of this task; L01 then adds the channel premiere when Z02 lands |
L02 On-demand library in the watch area |
M3 | M | lilith-v1-4:T08, G03, G02, L09 |
L03 Premiere events |
M3 | S | L01 |
L04 Syndication of films |
M3 | S | lilith-v1-4:Y02, R11, G11, R16, G02, L09, K17, K18 |
L05 Festival submissions |
M2 | S | R09, K07, R16, R20 · outside: festival fees (Elikem) |
L06 Audience analytics into development |
M3 | S | lilith-v1-4:A03, G03, G02 |
L07 Press kit and EPK |
M3 | S | K11, K12 |
L08 Campaign plan and reveal control |
M2 | S | K11, K12, I07, T11, G11, R16, content-workbenches:CR34, L09, K17 |
L09 Release windows and premiere status |
M2 | S | G01, G11, G10, R16, L05 |
L10 Sales desk and deal terms |
M3 | M | G14, L09, R16, R20, K07, L07, R14 · outside: counsel for contracts (Decision 19) and a sales agent if Elikem engages one |
M — Productions: the slate that proves the studio (lilith-v1-5:M)
| Task | M | Effort | Depends on |
|---|---|---|---|
M01 The pilot, directed by Elikem through the studio |
M1 | L | S03, S04, S17, T04, C12, P02, P03, N06, N13, X04, X05, K01, K06, K07, D02, D04, D07, R02, R11, U02, U03, U04, U05, U06, U08, U09, X10, V11, O02, O04, H16, P01, P05, P08, X03, D03, D08, K02, N01, I06, V04, Q03, N20 · outside: F.00.03's brief (40 seconds, 8 shots, 2 characters, one interior workshop set, a four-line exchange, one simulated cloth effect, full sound, graded and delivered, 1080p at 24 fps, stylised-realistic look, rated to the mature standard (R06) rather than safe for work, superseding F.00.03(c) for V1.5); Isis F.20.05 makes the same brief with agents on the F pipeline |
M02 The pilot, autonomous remake |
M2 | M | M01, H13 |
M03 The short film (3–8 minutes) |
M2 | L | M01, K07, D05, D07, K06, X06, R04, R08, R15, W06, C14, O03, K10, Q13, M10, D17 |
M04 The short, autonomous twin |
M2 | M | M03, H13, Q07 |
M05 Series pilot episode |
M3 | L | E03, E04, E05, M03 |
M06 Season one |
M3 | L | M05, E06, E07, L02, G11 |
M07 An autonomous episode, screened and compared |
M3 | M | E08, M05, Q07 |
M08 Post-mortems and pipeline learnings |
M2 | S | M01 |
M09 Screen the short and submit it |
M2 | M | M03, Q06, L05, R09, G11, K14, L09 · outside: panel recruited and festival fees paid (Elikem) |
M10 The short's pre-production pack |
M2 | S | R08, R15, O01, O05, W06, W07, G06, G10 |
Z — Horizon (lilith-v1-5:Z)
| Task | M | Effort | Depends on |
|---|---|---|---|
Z01 Feature film readiness |
M4 | L | M06 |
Z02 Opening the studio to Lilith members |
M4 | L | M06, S14 · outside: Elikem's go after M3 evidence (Decision 15, accepted 2026-09-29) |
Z03 Creator studio for outside creators |
M4 | L | Z02, lilith-v1-4:Z08, G02 |
Z04 Interactive and branching films |
M4 | M | M06 |
Z05 Live animated episodes |
M4 | M | M06, lilith-v1-4:P02, G02 |
Z06 Game tie-ins from film assets |
M4 | M | M06 · outside: Isis T.19/T.21 Unity lanes |
Z07 Immersive worlds from films |
M4 | M | M06 |
Z08 Character licensing and merchandise |
M4 | S | M06, R12 |
Z09 Franchise development plan |
M4 | S | W02, G06, G12, R16, Z06, Z08, M06, content-workbenches:CR23 |
O — Consultants and subject-matter experts (lilith-v1-5:O)
| Task | M | Effort | Depends on |
|---|---|---|---|
O01 Expert network and consultation records |
M1 | M | R08, G06, A05 |
O02 AI expert panel |
M1 | M | O01, T14, H02, H03, Q02, H14 |
O03 Period research and anachronism check |
M2 | M | T14, W05, O01, T06, H14 |
O04 Psychology and portrayal review |
M1 | M | C01, T03, T10, O01, R06, A06 |
O05 Field research and reference shoots |
M2 | S | O01, R13, R17, R10, A07, B08 |
U — The studio inside Blender (lilith-v1-5:U)
| Task | M | Effort | Depends on |
|---|---|---|---|
U01 Studio extension and app template |
M1 | M | S08, B01, B04, isis-chroma-runpod:F.03.03, B11 |
U02 A workspace for every department |
M1 | L | U01, A12, C06, C07, V05, X01 |
U03 Operation parity and native-edit capture |
M1 | M | U01, A03, A04, A12 |
U04 Agent proposals and co-driving in the viewport |
M1 | M | U01, H03, B04, S06, S10, H15, S16 |
U05 Review notes and approvals inside Blender |
M1 | S | U01, S04, isis-chroma-runpod:F.06.05 |
U06 Context panels for the shot in hand |
M1 | M | U01, A08, T06, P04, V11, C09, W10 |
U07 Craft instruments and overlays |
M2 | M | U01, isis-chroma-runpod:F.06.04, isis-chroma-runpod:F.07.05, isis-chroma-runpod:F.12.02, D03, C11, V11 |
U08 Studio asset shelves |
M1 | S | U01, B08, R05 |
U09 Blender workspace tests and budget |
M1 | S | U02, U03, U04, U05, U06, U08, B11 |
U10 Control catalogues for every department |
M2 | M | U03, N20 |
Total: 318 tasks and 24 epics (revision 14 added 31 tasks and the O and U epics; revision 20 added T15, C15, N18, X11, V12, V13 and O05; the second role audit adds V14, V15, X12, C16, N19, D14, D15, D16, K15, K16, Q13, A14, L08, Z09 and G13), section 17 adds N20, U10 and M10, and section 18 adds G14, K17, K18, L09, L10, R20, C17, D17 and S18; 170 task dependencies point into other families.
12. Field review, 2026-09-29 (revision 14)#
Elikem asked, after approving revision 6, for the plan to capture the best and latest industry techniques and the newest effective ones. Two passes were made the same day (see Reviews below), each reading this PRD, every task body and the Isis F and T titles. Anything the F, T, CR or V1.4 plans already own was left alone. Facts in section 3 were corrected where they had gone stale; the changes below are the ones that move film quality or Elikem's speed, ranked by expected effect.
| # | Change | Why it matters | Where |
|---|---|---|---|
| 1 | Protect authored timing end to end: approved cadence (ones, twos, holds, smears) is a record every stage honours; interpolation never crosses stepped animation | Spider-Verse-style timing is what makes a stylised film feel authored; spline conversion, interpolation and generative finishing silently erase it | N02, N07, Y08 |
| 2 | Search, then choose (principle 12): candidates with verifiers, a per-shot trunk, verified entity memory, backtracking from the last good prefix, N set from a measured curve; refinement versus best-of-N compared at equal cost | The strongest 2026 results in multi-shot generation come from verified search, not one-shot prompting | H16, H17, H04, principle 12 |
| 3 | Stylised identity: adapters trained on each character's own renders, and an on-model check calibrated on multi-view renders, whole-character cues, occlusion and intentional deformation | Identity drift is the documented failure of generated finishing; face embeddings misjudge stylised and non-human faces | Y11, C11, Y04 |
| 4 | Voice performance: duration-targeted takes and directing a line by acting it, converted to the character's original voice | Acting and timing are where synthetic voices fall short; a solo director can now perform the read he wants | D02, D12, K09 |
| 5 | EEVEE as an approved final engine, measured per look | Flow and Il Baracchino finished in EEVEE; for a stylised look it can multiply iteration speed and cut farm cost | V05, X05, B09 |
| 6 | Revision-aware visual memory for long shots and for production questions | Identity after a character leaves and returns, and "what happened in that shot", are where overlap-only generation and text memory fail | H09, Y05 |
| 7 | A stylised facial performance contract instead of a fixed ARKit 52 | KPop Demon Hunters-style faces need view-dependent correctives and mesh swaps, and creatures have no human face | C07 |
| 8 | Learned in-betweening from blocking as editable keys | Production-proven at Animaj (3.5x faster in-betweening) and newly open at SIGGRAPH 2026 | N14 |
| 9 | MCP 2026-07-28, Agent Skills and GEPA-style trace optimisation for the crew | The protocol changed in July; skills make a role portable across Astra, Opus and cheap bindings | A04, H02, H10, H11 |
| 10 | Provenance and rights: C2PA 2.4, CAWG 1.3, IPTC source types derived from the operations, the EU icon; licences checked for UK territory | Accurate labels make disclosure interoperable; Tencent Hunyuan licences exclude the UK | R04, R05, G02 |
| 11 | Blender's NPR nodes: the Raycast node now, Material Lighting nodes with 5.3; comic, tactile and watercolour kits | The look kits were written as if Blender had no NPR nodes | V05, B10 |
| 12 | Festival exhibition copy (DCP, stereo or 5.1) | Nothing made one, and a selected short must screen | K14 |
| 13 | Factual versus perceptual scoring, agentic video critics that report what they watched | An attractive but wrong result must not pass; a critic that skipped the frame cannot vouch for it | Q01, Q02 |
| 14 | Video-to-audio foley proposals beside 3D-event sync | New models propose synchronised layers; stems and scoring keep them editable | D04 |
| 15 | Live look preview (optional) | Real-time video-to-video can show the final look while blocking; kept only if it changes decisions | X09 |
Corrected facts (section 3): Blender has NPR nodes (the 5.1 Raycast node in the pinned LTS; Material Lighting nodes in 5.3 development), Blender 5.2.2 exists, Gemini no longer reads video only at a fixed 1 fps, Hunyuan weights are unusable by a UK company, Sora's API ended on 2026-09-24, the controller field now includes GPT-6 Sol and Luna, Claude Sonnet 5.5, Gemini 3.8 and new open models, IndexTTS's released duration control is a factor rather than an exact mode, LTX-2.5 is the current open release, C2PA is at 2.4 and CAWG identity at 1.3, the SAG-AFTRA animation agreements expired on 2026-06-30 with a successor only tentatively agreed, universal "best model" claims are softened, and unverifiable anecdotes (costs, premieres, per-minute prices) are labelled as such.
Deliberately not adopted: USD as the department-layer backbone (Blender's library overrides, F.03.03, already give per-department layering and Blender is the source of truth; USD stays an interchange format, A09); switching the pilot from AgX to ACES 2.0 (F.00.03's decision stands; HDR is a per-title option); a paid avatar product; any Tencent Hunyuan model; universal "best model" rankings (every lane is measured per shot class, Y09); anecdotal cost figures as planning inputs.
Fixed after the Eve review of the submitted amendment (review 86559333, eve-reviewer-01): Decision 18's on-demand fallback could not unblock the first season, because M06 depended on the channel task L01 and L01 on V1.4's Z02. The season (M06) and its schedule (E07) now depend on the on-demand library (L02), and L01 adds the channel premiere when Z02 lands. This flaw was already present in revision 6.
Reviews. Claude's own search of primary sources, and GPT-6 Astra at xhigh
effort through codex --search exec -s read-only on the ChatGPT subscription
(about 160 searches). Astra confirmed the architecture and most agent patterns
as current, and contributed the cadence, identity-calibration, facial-contract,
visual-memory and provenance-version findings and most corrections above. A
second read-only Astra pass reviewed the role audit and the Blender workspace
(sections 13 and 14; 32 findings, all folded in), and a separate read-only
research pass compiled the Blender 5.0–5.3 feature digest behind section 14.
What revision 14 changes on the board: the new tasks and the O and U epics
are created as later until this revision is approved; the task amendments and
new dependencies are applied only after approval. The evidence document
(b223ede2) carries the same changes at its revision 3.
13. Every role, department and perspective#
Elikem asked (2026-09-29, after approving revision 6) that every role a top studio staffs is accounted for, and that every role gets three things: the full tools of the craft, granular control for him, and an agent that can do the job. This table is that audit. It draws on how feature animation (Pixar, Walt Disney Animation, DreamWorks, Sony Pictures Animation and Imageworks, Laika, Aardman, Fortiche, Cartoon Saloon, Studio Ghibli and the anime studios' production roles), prestige television and live-action features crew a production, and it keeps the live-action roles whose knowledge an animated film still needs (fight choreography, dialect coaching, historical and technical consultants). Roles that exist only because of physical sets or cameras (grips, gaffers, catering, on-set safety) have no counterpart here.
References for the department structure, checked 2026-09-29: Walt Disney Animation Studios' process and team pages (technical direction, stereo); Sony Pictures Imageworks' production accounts (Spider-Verse; KPop Demon Hunters, 2026-01-15); Fortiche's Arcane season 2 department pages; Cartoon Saloon's production study guide; Pixar's Inside Out 2 production notes (2024-05-23); ScreenSkills job profiles (music editor and others); Blender's Flow and Il Baracchino production stories. The audit is as complete as these sources and the two reviews could make it; the field watch (Q10) adds any role a production shows is missing.
Reading the columns: Tools are the tasks that give the role its craft (tasks in other families are written in full); Control is where Elikem directs or overrides it: the web room, the Blender workspace (U pillar) or a record he locks; Agent is the crew role in H02's roster; the dial (section 5) sets how far it may act. Status: covered (revision 6 already had it), deepened (revision 14 amends a task) or new (revision 14 adds a task).
13.1 Studio leadership, producing and the production office#
| Role | Tools | Control | Agent | Status |
|---|---|---|---|---|
| Studio head, executive producer | I06 slate and greenlight, G06, G10 decision register | Greenlight records he signs | none (Elikem only) | covered |
| Producer | G06, G07 cost ledger, B07 estimates, S11 meter | Budgets and caps per title | producer | covered |
| Line producer, production manager | G12 bids, schedule, quotas, burn-down and what-if | The schedule and quotas he approves | production manager | new |
| Production coordinator, assistants | S02 "what needs me", S03 tracker, A08 events | Batch decisions on the home screen | producer | covered |
| Production accountant | G07 with incentive fields, S11, G14 | Caps; spend approvals; which incentives to pursue | producer | deepened (section 18) |
| Development executive, story analyst | I04 comparables, T07 coverage | Coverage notes accepted or rejected | story editor | covered |
| Showrunner | E01–E09, S07 director chat at series scope | Season arc and episode locks | showrunner | covered |
| Post-production supervisor | K pillar, K07, K14 | Delivery approvals | post supervisor | deepened (roster) |
| Business and legal affairs | R12 title and script clearance, R14 legal review and E&O readiness, R16 | Counsel engagement; release authorisation (G11) | none (people only) | deepened |
| Archivist, librarian | A11 archive and rebuild, B08 | Retention rules | QC | covered |
| Pipeline technical director | H12 diagnosis and repair diffs, B05, B12, B14, isis-chroma-runpod:F.02.08 | Dependency and diagnostic view; repair approvals | pipeline TD | deepened |
13.2 Development, writing and story#
| Role | Tools | Control | Agent | Status |
|---|---|---|---|---|
| Idea generation, development | I01–I09, content-workbenches:CR04–CR06 | Idea inbox, taste board (I08) | writer, showrunner | covered |
| Screenwriter, head writer | T01–T04, T13 human-authored mode | Every line authored, locked or accepted | writer | covered |
| Writers' room, staff writers | T12, lilith-v1-4:W06 | Room settings, pitch acceptance | writer (several) | covered |
| Story editor, script editor | T05, T07, T11 | Notes accepted per line | story editor | covered |
| Script coordinator | T10 with revision colours, A-pages and revision marks | Locks and revision sets | script coordinator | deepened |
| Comedy, gag and punch-up writer | T05 comedy and gag pass | Gag bank per character | writer | deepened |
| Script doctor | T07, lilith-v1-4:W06 | Accept or reject each note | story editor | covered |
| Head of story, story artists | P01 boards room, P02 animatic, content-workbenches:CR32 | Panels and timing | board artist | covered |
| Story trust ("braintrust") screenings | P06, K03, Q02, content-workbenches:CR24 | Notes sessions he chairs | critic panel | covered |
| Series composition, season arc | E02, E03 | Arc locks | showrunner | covered |
| Researcher | T14 research desk, O03 period research | Sources accepted per record | researcher | deepened |
| Lyricist, songwriter | D06 | Lyrics authored or accepted | composer | covered |
13.3 Direction#
| Role | Tools | Control | Agent | Status |
|---|---|---|---|---|
| Director | Every room; S07 director chat; U pillar in Blender | All locks and the dial | director agent (propose) | covered |
| Co-director, sequence and episodic directors | S07 at sequence and episode scope, E05 | Scope-level dial | director agent | covered |
| Animation director, supervising animator | N01, N02, N11, N13 | Performance locks, N13 layers | animator | covered |
| Action director (second unit) | N15 action coverage and readability rules | Fight phrase approvals | fight choreographer | new |
13.4 Art department and visual development#
| Role | Tools | Control | Agent | Status |
|---|---|---|---|---|
| Production designer | V01 style bible, X01 sets, W09 | Style bible locks | production designer | deepened (roster) |
| Art director | V01, V09 art department review | Art review decisions | art director | covered |
| Visual development, concept artist | V02, Y07, V06 own-art adapters | Pick, edit, lock designs | art director | covered |
| Character designer | C04, C05, content-workbenches:CR08 | Model sheet lock | character designer | covered |
| Costume designer | C14 costume bible, plot, pattern garments, fabric presets | Costume locks per scene | costume designer | new |
| Hair and makeup designer | C14 (hair and makeup states, ageing, wounds), isis-chroma-runpod:F.09.04 | Continuity states | costume designer | new |
| Colour designer, colour stylist | V11 colour models per lighting condition, palettes, symbolism, colour-vision checks; V03 colour script | Palette locks | colour designer | new |
| Graphic designer (in-world graphics) | V10 signage, documents, currency, screens, heraldry, scripts, fonts | Graphic locks | graphic designer | new |
| Prop designer, prop master, armorer | X10 prop plot, hero props, prop state, vehicles, weapons | Prop locks and plot | prop master | new |
| Mechanical and vehicle designer | X10 articulation for vehicles and mechanisms | Design locks | prop master | new |
| Set and environment designer | X01, X02, isis-chroma-runpod:F.08.01–F.08.03, W04 maps | Set locks | production designer | covered |
| Set dresser | isis-chroma-runpod:F.08.02, U08 shelves | Dressing edits in Blender | production designer | covered |
| Matte painter | X07, isis-chroma-runpod:F.08.06 | Plate approval | art director | covered |
| Location scout | X02, W09 splat worlds | Location picks | production designer | covered |
| Cartographer | W04 | Map edits | world-builder | covered |
| Colour key, lighting key artist | V03, V04 style frames | Key frame locks | colour designer | covered |
| Title designer, typographer | K05, V10 fonts under R05 | Title approval | graphic designer | covered |
| Key art and poster illustrator | K12, V04 | Key art approval | art director | covered |
| Background art director (2D and anime) | V07 | Background locks | art director | covered |
13.5 Characters, rigging and performance#
| Role | Tools | Control | Agent | Status |
|---|---|---|---|---|
| Character psychology and arc | C01, C02, C03 interviews, O04 depth and consistency | Trait locks | writer, psychology expert agent | deepened |
| Modeller, sculptor | C06, isis-chroma-runpod T track | Asset locks | character designer | covered |
| Texture, surfacing and look-development artist | V05 look kits, isis-chroma-runpod:F.12.04 | Look locks | art director | covered |
| Rigger, character TD | C06, C07, isis-chroma-runpod T.11, C18 face basis and its projections | Rig certification | character TD | covered |
| Creature designer and creature TD | N03 creature and animal gaits, N12 generated creature motion (Isis T.12.10), N28 gait grammar, W08 ecology | Creature locks | animator | deepened |
| Groom artist | isis-chroma-runpod:F.09.04 | Groom locks | costume designer | covered |
| Cloth and CFX artist | isis-chroma-runpod:F.13.02, C14 fabric presets | Simulation stage approval | effects | covered |
| CFX shot finaler | N17 shot-local correctives, sim and performance blending, protected ranges | Corrections he sculpts and locks | CFX finaler | new |
| Animator (blocking, spline, polish) | N01, N02, N13, N14 learned in-betweens, U07 instruments | Keys on his own layer; locks | animator | covered |
| Facial and lip-sync animator | N05, N06, isis-chroma-runpod:F.11.03–F.11.05, N24 dialogue fusion, N25 stylisation layer | Mouth and face keys | animator | covered |
| Stylised animation (twos, smears) | N07, N02 holds and twos kept | Timing locks | animator | covered |
| Motion-capture supervisor and performer | N04 body capture, N05 face capture, N22 and N26 solvers with their re-render and physics gates | Takes he performs | animator | covered |
| Fight choreographer, stunt coordinator, fight-style researcher | N15 research, fight bible, beat notation, stunt-vis, readability | Phrase approvals | fight choreographer | new |
| Dance choreographer | N16 beat-grid choreography, cultural forms with consultation | Phrase approvals | dance choreographer | new |
| Movement coach, physicality | C09 acting profile | Profile edits | animator | covered |
| Crowd TD | N10 paths, groups, cue reactions, per-agent overrides, hero promotion | Crowd approval and overrides | crowd specialisation | deepened |
| 2D key animator, in-betweener, 2D animation director | N09, V07, S04 drawovers | Key and correction drawings | animator | covered |
| Clean-up artist, animation checker, ink and paint, 2D effects animator | N09 production stages, correction layers, exposure IDs, fill and palette checks, hand-drawn effects | Stage approvals | clean-up, checker, paint and 2D-FX specialisations | deepened |
| Performance-mining wrangler (reference libraries) | N21 and N27 libraries mined from generated reference, reference-only under Y01; N28 gait grammar; Q14 face rung | Library approvals; a rejected clip never yields a pose or phrase | animator | new (section 20) |
13.6 Camera, staging and lighting#
| Role | Tools | Control | Agent | Status |
|---|---|---|---|---|
| Cinematographer, director of photography | P04 camera language, P05 coverage, X03 | Camera rules he writes | cinematographer | covered |
| Layout artist and supervisor | P03, P08, isis-chroma-runpod:F.06.03 | Layout locks | layout artist | covered |
| Previs supervisor | P06, isis-chroma-runpod:F.06.05 | Previs approval | layout artist | covered |
| Camera operator, virtual camera | P07 | Moves he records | cinematographer | covered |
| Stereographer (opt-in per title) | P09 depth script, stereo cameras, depth budgets, window and disparity checks | Parameter locks | stereographer | new (M4) |
| Focus puller | isis-chroma-runpod:F.07.03 | Focus keys | cinematographer | covered |
| Lighting supervisor, lighter | X03, X08, isis-chroma-runpod:F.12.01–F.12.06 | Light locks per shot | lighter | covered |
13.7 Effects, rendering, finishing and editorial#
| Role | Tools | Control | Agent | Status |
|---|---|---|---|---|
| VFX supervisor | Y01–Y11, X04, X06 | Lane policy and fidelity gates | VFX supervisor | deepened (roster) |
| Effects artist | X04, isis-chroma-runpod:F.13.01–F.13.06 | Stage approvals | effects | covered |
| Render wrangler | X05, B06, B07 | Farm caps | producer | covered |
| Compositor, anime photography director | X06 with interactive compositor playback and node gizmos, V07 2D compositing | Comp locks | compositor | deepened |
| Generative finish artist | Y03, Y04, Y10, Y11 | Fidelity gate and approvals | VFX supervisor | covered |
| Editor, assistant editor | K01, K02, K04 | Cut locks | editor (alternate cuts only) | covered |
| Story-reel editor | P02 | Animatic locks | editor | covered |
| Colourist | K06, isis-chroma-runpod:F.19.01–F.19.02 | Grade locks | colourist | covered |
| Online and finishing editor, QC operator | K04, K07, K14 | Delivery approval | QC | covered |
| Trailer editor | K11 | Trailer approval | marketing | covered |
13.8 Sound and music#
| Role | Tools | Control | Agent | Status |
|---|---|---|---|---|
| Supervising sound editor, sound designer | D04 with a video-to-audio lane, D09 | Cue approvals | sound designer | deepened |
| Foley artist | D04, isis-chroma-runpod:F.18.03 | Layer approvals | sound designer | covered |
| Dialogue editor, ADR | D02, D03, D12 | Take and comp locks | dialogue editor | covered |
| Voice director | D02 duration targets, dialect consistency; D12 directing by acting | Every read | voice director | deepened |
| Casting director | C10, lilith-v1-4:C03 | Casting decisions | casting director | covered |
| Dialect coach | D02 dialect consistency, W07 lexicon | Lexicon locks | voice director | deepened |
| Re-recording mixer | D07, K14 cinema mix | Mix approval | mixer | covered |
| Composer, score direction | D05 spotting sessions, leitmotif map, intensity curve, orchestration palette | Cue and theme locks | composer | deepened |
| Orchestrator, arranger | D11 editable score | Note-level edits | composer | covered |
| Music editor | D13 cues tied to picture revisions, hit points, click tracks, phrase-aware reconform | Anchors he locks | music editor | new |
| Music supervisor | R11 music and sample rights | Clearance decisions | music supervisor | deepened (roster) |
13.9 Consultants and subject-matter experts#
| Role | Tools | Control | Agent | Status |
|---|---|---|---|---|
| Cultural consultant | R08, W06 | Canon gate on sign-off | expert agent (advisory) | covered |
| Folklore and source clearance | R15 | Permissions recorded | none (people only) | covered |
| Historian, period consultant | O03 period bible and anachronism check | Accept per note | expert agent | new |
| Psychologist, mental-health and representation consultant | O04 portrayal checks and consultation | Hold at script lock | expert agent (advisory) | new |
| Military, weapons, martial-arts expert | O01, N15, X10 | Accept per note | expert agent | new |
| Medical, science, engineering and legal advisers | O01, O02 | Accept per note | expert agent | new |
| Linguist, conlanger | W07 | Lexicon locks | world-builder | covered |
| Ethnomusicologist | O01, D05 | Accept per note | expert agent | new |
| Religion and mythology adviser | O01 (human sign-off domain), R15 | Canon gate | expert agent (advisory) | new |
| Animal-behaviour adviser | O01, N03 | Accept per note | expert agent | new |
| Child-development adviser | O01, R19 | Accept per note | expert agent | new |
| Textile and costume historian | O01, C14 | Accept per note | expert agent | new |
| Architecture adviser | O01, X01 | Accept per note | expert agent | new |
| Sensitivity reader | O01, O04 | Hold at script lock | expert agent (advisory) | new |
13.10 Release, audience and the perspectives around the film#
| Role or perspective | Tools | Control | Agent | Status |
|---|---|---|---|---|
| Standards and practices, ratings | R06, R18, R19, lilith-v1-4:S01 | Rating decision | QC | covered |
| Provenance and disclosure | R01, R04, R07 | Disclosure wording | QC | covered |
| Localisation lead, translators, dubbing director | K08 English subtitles; K09 and K13 deferred (Decision 12) | Language sign-offs once a language opens | localisation lead | deferred (section 19) |
| Accessibility (audio description, captions) | K10, D08 | Approval | QC | covered |
| Audience research, test screenings | Q06, Q07, L06 | Screening protocol | none (people) | covered |
| Marketing, publicity, EPK | K11, K12, L07 | Campaign approval | marketing | covered |
| Distribution, festivals, sales | L01–L05, K14 exhibition copies, L09 release windows, L10 sales desk, G14 income | Release authorisation (G11), the release plan and every deal | producer | deepened (section 18) |
| Merchandise and licensing | Z08 | Horizon | none | covered (horizon) |
| The director's taste | I08, Q04 | Ratings he gives | taste model | covered |
| Craft teaching | S15 craft mentor | On request | mentor | covered |
Every agent role above ships as a skill directory (H02), is benchmarked per department (Q11, which now names costume, colour design, graphics, props, choreography and research), runs at the dial Elikem sets, and every role's controls exist in the web room, in the Blender workspace and as an MCP tool, which U03's parity table proves.
13.11 Added by the productions audit (section 15)#
| Role | Tools | Control | Agent | Status |
|---|---|---|---|---|
| Adaptation lead (showrunner adapting a novel or game) | T15 fidelity map, canon extraction, source rights | The fidelity map he locks | showrunner | new |
| Continuity keeper (a franchise's story group) | W02 cross-title canon and continuity database | Canon locks | continuity keeper | deepened |
| Character variant designer (ages, bodies, upgrades, clones) | C15 | Variant locks | variant designer | new |
| Vehicle-action animator, space-battle and chase choreographer | N18 flight models, formations, action cameras, readability | Maps and passes he approves | vehicle-action animator | new |
| Zero-gravity movement coach, environment physics | X11 gravity and medium profiles | Profile per scene | zero-gravity movement coach | new |
| 3DCG anime director and toon look-development TD | V12 normals, face-shadow maps, outlines, limited animation | Look locks | 3DCG anime director | new |
| Motion-graphics and interface designer (screens, holograms, HUDs) | V13 | Editable sources | motion-graphics designer | new |
| Gesture choreographer, conlang coach | W07 gesture vocabularies | Lexicon and gesture locks | gesture choreographer | deepened |
| Visual consultant (a live-action cinematographer) | O01 domain, P04, X03 | Accept per note | expert agent (advisory) | deepened |
| Field researcher, reference photographer and recordist | O05 | References he accepts | field researcher | new |
| Character lighter (eye lights, rim and kicker cheats) | X03 | Light locks | lighter | deepened |
| Anime layout checker, second key, in-between checker, finish checker | N09 | Stage approvals | 2D specialisations | deepened |
| Montage and music-video editor | K02 | Cut locks | editor | deepened |
| Opening-title designer | K05 | Title approval | graphic designer | deepened |
| Signature-sound designer | D04 sound lexicon per universe | Lexicon locks | sound designer | deepened |
| Vendor and freelancer coordinator | S14 work packages | Package approvals | producer | deepened |
| Streamer delivery specialist | K07 streamer profile | Delivery approval | post supervisor | deepened |
| Ratings and age-assurance lead | R06, R19 mature-audience standards | Rating per title | QC | reworked |
13.12 Added by the second role audit (section 16)#
| Role | Tools | Control | Agent | Status |
|---|---|---|---|---|
| Board-first story artist (story from drawings, no script first) | P01, P02, T09 reverse handoff | Which representation is authoritative | board artist, story editor | deepened |
| Line animator (expressive linework as performance) | V14 | Line locks and shot keys | line animator | new |
| 2D lighting and texture artist | V15 drawn lighting tracked across drawings | Every shape and exposure | drawn-lighting artist | new |
| Hand-drawn 2D in-betweener (no model required) | N09, V07 manual path | Generation off per exposure | animator | deepened |
| Effects art director | X12 guide curves, silhouette keys, timing envelopes | Shape, breakup, staging | effects art director | new |
| Feather and wing TD | C16 | Feather grouping and silhouette | feather and groom TD | new |
| Transformation TD | N19 | Pose-by-pose transitions | transformation TD | new |
| Capture supervisor (fidelity) | N04, N05 calibration, sync, source-versus-retarget review | Take choice, protected moments | animator | deepened |
| Singing-voice director | D06 sung-performance record | Syllable, note and phoneme edits | composer, voice director | deepened |
| Sound-design constructor | D14 editable recipes | One component of a signature sound | sound-design constructor | new |
| Re-recording mixer (dramatic control) | D16 automation, snapshots, perspective | Focus, silence and perspective per cue | mixer | new |
| M&E editor and international versioning | D15 | Exceptions per locale | M&E editor | new |
| Secondary colourist | K16 masks, tracking, qualifications | Region-level grades | colourist | new |
| Visual-localisation designer | K15 per-locale picture text, textless elements | Translations and typography | visual-localisation designer | new |
| Accessibility editor and consultants | K10 experience review | Descriptive interpretation | accessibility editor | deepened |
| Audience researcher (formative screenings) | Q13 | Which findings to act on | audience researcher | new |
| Archivist (preservation) | A14 fixity, copies, migrations | Retention and migration approvals | archivist | new |
| Campaign lead | L08 reveals, embargoes, asset-to-revision map | Every reveal | campaign lead | new |
| Franchise strategist | Z09 | Canon boundaries and greenlights | franchise strategist | new (M4) |
| Sustainability lead | G13 | Plan choice | sustainability lead | new |
13.13 Added by the feature-gap review (section 18)#
| Role | Tools | Control | Agent | Status |
|---|---|---|---|---|
| Release strategist (festival and distribution windows) | L09, L05 | The release plan he edits | producer | new |
| Sales agent, distribution executive | L10 | Every deal he signs | producer | new |
| Social and short-form editor | K17 | Every reframe he accepts | editor, layout artist | new |
| Versioning editor (rating, platform and runtime edits) | K18 | Every change in a version | editor | new |
| Loop group and walla coordinator | D17 | Density, language and lines | voice director, dialogue editor | new |
| Content protection (previews, anti-piracy) | G01 preview, R20 | Every preview recipient and every takedown | QC | new |
| Improvisers (character workshop) | C17 | Situations, steering and kept moments | one writer agent per character | new |
14. Blender: what to adopt, and the studio inside Blender#
Sources: the release notes for 5.0 to 5.3 on developer.blender.org, the 5.2 manual and API documentation, Blender's milestones and roadmap, the download server and Blender Studio's project pages, read 2026-09-29 (about 330 feature rows were checked; the working digest is kept with B14's evidence). Blender's own two-year strategy, published in September 2026, is "making Blender the tool that enables small independent studios to work on long-form projects, such as an animated feature film"; its test bed is Blender Studio's feature Overgrown, and its NPR short Cowboi (announced 2026-09-21) is driving the Material Lighting nodes. Blender's stated policy is that no generative AI is planned for core Blender, so the studio's AI stays outside Blender and reaches it only through the operation set.
14.1 Release calendar#
| Release | Date | What it means here |
|---|---|---|
| 5.2 LTS | 2026-07-14, supported to July 2028 | The pinned production version; 5.2.1 (2026-08-25) and 5.2.2 (2026-09-15) shipped, 5.2.3 is planned for 2026-10-20; B01 moves a title only after its golden shots re-render (Q09) |
| 5.3 | beta 2026-09-30, release 2026-11-10 | Adopted behind flags through B10; files using 5.3-only data are assumed not to round-trip into 5.2 |
| 5.4 | 2027-05-11 | Watched (Q10); its milestone lists sequencer effects and storyboarding markers |
| 5.5 LTS | 2027-11-09 | The next LTS; the likely pin for the series (M3 and later) |
| 6.0 | alpha from 2027-06-30 | Breaking changes to animation, sequencer, asset and node Python APIs; B14 plans the extension's migration |
14.2 Adoption map#
"Now" means in the pinned 5.2 LTS; "flag" means 5.3, adopted through B10 after its release; "watch" means roadmap only. B14 keeps this map current and owns a test for every adopted row.
| Feature | Version | What it gives the film | Owner | Adopt |
|---|---|---|---|---|
| Shader Raycast node (lines, rims, x-ray), per-object raycast visibility | 5.1, 5.2 | Outlines and contact lines inside the material | V05 | now |
| Shader closures, bundles, repeat zones, Menu Switch | 5.0 | One master toon shader whose style is an enum an agent can set | V05 | now |
| Material Lighting nodes (per-light toon ramps, rims, halftone; EEVEE) | 5.3 | Real multi-light toon shading, the look Cowboi is proving | V05, B10 | flag |
| Line Art Fill Strokes; Grease Pencil motion blur, corner types, textured and dot brushes | 5.0, 5.2 | 3D-to-cel ink with fills; painterly and stippled lines | V05, V07 | now |
| Grease Pencil fill redesign and Delaunay bucket fill | 5.1, 5.2 | Cel ink and paint with gap closing: the 2D lane's base | N09, V07 | now |
| Grease Pencil Draw-mode select and transform, Carver, multi-region fill, gradients | 5.3 | 2D animators stay in Draw mode | V07, B10 | flag |
| Sequencer edit scene, Sync Scene Time, scene assets, Storyboarding template | 5.0 | Board to reel to shot inside Blender; scrub the reel, land in the shot | P01, P02 | now |
| Compositor strips, GPU strip modifiers, text style presets | 5.2 | Transitions, grades, titles and subtitles finished in the edit | K01, K05, isis-chroma-runpod:F.15.04 | now |
| Compositor node-group assets, built-in looks, interactive playback, Image Editor node gizmos | 5.0, 5.2 | Per-show bloom, grain and colour stacks as versioned assets; live comp review | X06 | now |
| Compositor effects stack and playback cache | 5.3 | Simple layered compositing | X06, B10 | flag |
| EEVEE throughput (shader compile 32–53% faster, instancing up to 2x, sturdier SSR and fast GI) | 5.1, 5.2 | EEVEE as a final engine for stylised titles | V05, X05, B09 | now |
| XPBD cloth and hair dynamics with custom forces (experimental) | 5.2 | Art-directable secondary motion, including cartoon forces | X04, N17 | now, pinned per title |
| Geometry Nodes lists, bundles, nodes on Empties, Sample Sound Frequencies | 5.2 | Procedural effects and crowds with real data; motion driven by music or dialogue | X04, N10, N16 | now |
| Node tools registered as operators, inputs settable from Python | 5.1, 5.2 | Procedural tools callable as typed operations | B03 | now |
| Animation: object-mode in-betweens, playback loop modes and preroll, Graph Editor local view, Copy Constraints | 5.2 | Faster blocking and polish | N13 | now |
| Action evaluation +125–146% and shape-key evaluation +130–304% | 5.1 | Real-time playback of face-heavy toon rigs | C07, N06 | now |
| Shape-key workflow (multi-select, Make Basis, flipped join, Apply to Basis) | 5.0, 5.1 | Quicker toon faces and correctives | C07 | now |
| World-space animation copy and paste, rotation-mode conversion | 5.3 | Animation transfer between rigs and props | B10 | flag |
| Working colour space, ACES 2.0 and AgX HDR views, HDR export and scopes | 5.0, 5.2 | One pipeline for SDR and HDR masters (AgX stays the pilot's choice) | isis-chroma-runpod:F.14.03, K06, K07 | now |
| Cycles texture cache, Principled BSDF Thin Wall | 5.2 | Painterly Cycles shots that fit in memory; paper, leaves and cloth | V05, C14 | now |
| Light and shadow linking (copying improved) | 5.2 | Per-receiver lighting control | X03 | now |
| VR location scouting | 5.2 | Block vantage points and lenses inside the set | P07 | now |
| Remote asset libraries on static HTTP; access tokens | 5.2, 5.3 | One studio library for the director, the farm and the agents | U08, B08 | now, tokens flag |
| Projects system (.blender_project, {project_root} paths, project libraries and colour configuration) | 5.3 | The native form of per-title pinning | B10, B01 | flag |
| Packed linked data and archive libraries; Flamenco 3.9 with Blender Asset Tracer 2 | 5.0 | Self-contained shot copies for the farm and agents | B06, A11 | now |
| Headless automation (CLI commands, headless GPU, image API, reports list, path enumeration) | 5.0–5.2 | Agents run QC, thumbnails, audits and publishing without the UI | B02, H01 | now |
| Gaussian splats (import, render, USD splat schema), with documented limitations | 5.3 | Scanned and generated sets | X07, W09, B10 | flag |
| Bundled OpenTimelineIO Python module | 5.3 | A native base for the OTIO bridge | isis-chroma-runpod:F.17.02, B10 | flag |
| Blender Lab MCP server (runs generated Python "without any guards") | Lab, 5.1+ | An official bridge for agents, only behind the studio's own command layer and the contained lane | B05, A04 | now, isolated |
| Layered actions (animation layers) | roadmap; NLA-style strips 2027 or later | Real animation layers; until then N13 uses NLA | N13 | watch |
| Dynamic overrides, Story Tools Next, meta-rigs and on-model gizmos, texture layers, FLIP fluids in Geometry Nodes | roadmap | Per-shot tweaks on linked assets, board to previs, rigging at scale | B14, Q10 | watch |
Script-breaking changes the studio's code must absorb (B14 owns a fixture for each): the legacy Action API and BGL were removed in 5.0, along with dictionary access to add-on properties; 5.1 renamed the sequencer strip-time properties; 5.2 moved Geometry Nodes modifier inputs from custom properties to RNA properties. Blender Studio's GPL pipeline tools (CloudRig, Asset Pipeline, Contact Sheet, Lighting Overrider, Pose Shape Keys, Anim Cupboard, Brushstroke Tools, Blender Kitsu) are evaluated for reuse under G02 rather than rebuilt.
14.3 The studio inside Blender: how it is built#
| Need | Blender mechanism (5.2 LTS) | Task |
|---|---|---|
| Install and pin | An add-on extension (blender_manifest.toml; version range pinned per title) served from a private static extension repository with a per-title index | U01 |
| A full studio edition of Blender | An app template (startup.blend with the workspaces, userpref.blend, splash, register hooks), installed separately because it is not an extension type | U01, U02 |
| Department workspaces | Workspaces authored in the template's startup file and appended per department; owner filtering shows only the studio's panels and tools where that helps | U02 |
| Studio tools in the toolbar | WorkSpaceTool classes per space and mode, with gizmo groups for camera framing and rig handles | U02, U07 |
| Ghosts, drawovers and instruments | View-3D draw handlers with the gpu module, reading cached display data and restoring GPU state | U04, U05, U07 |
| Studio library in Blender | Asset shelves over a remote asset library | U08 |
| Talking to the studio without freezing Blender | No async operators exist and Python threads are not supported: the broker runs in a separate process and results come back through timers and modal operators on the main thread | U01 |
| Knowing when something changed | The message bus and handlers only as hints (they miss viewport transforms and animation), with explicit checkpointed transactions as the record | U03 |
| Studio data on Blender objects | Property groups with library-override flags | U03, U06 |
| Agents and the farm | Registered CLI commands for headless jobs; the Lab MCP server only behind the operation layer | B02, B05, A04 |
| Permissions | Manifest declarations for network and files, respecting online access; enforcement in the broker, not in Blender | U01 |
Everything Elikem can do in the web studio, he can do here, and every control here is an operation an agent can also call: U03's parity table is the proof, and a capability without all three is not admitted.
15. What the reference productions add, and the audience correction (2026-09-29)#
Elikem asked whether making Arcane, Star Wars: The Clone Wars, The Expanse, Altered Carbon, Altered Carbon: Resleeved, Cyberpunk: Edgerunners or a Pixar or DreamWorks feature would need roles or tasks this plan lacks, in granular technical detail. Each production's documented methods were compared with every task body after revision 14. Most of their craft was already covered (Arcane's hand-painted 3D and hand-drawn 2D effects: V05, N09; The Clone Wars' motion-captured duels: N04, N15; Coco's cultural consultants: R08, O01; the anime studio roles of colour design, compositing and checking: V11, V07, N09). What was missing:
| Production | Technique or role it depends on | Where it now lives |
|---|---|---|
| Arcane (League of Legends), Edgerunners (Cyberpunk 2077), The Expanse and both Altered Carbons (novels), The Clone Wars (the films) | Adapting a source work: canon extraction, a source-to-episode fidelity map, the source's rights | T15 (new) |
| Arcane (child and adult Vi and Jinx), Altered Carbon (re-sleeved bodies), Edgerunners (cyberware), The Clone Wars (individualised clones) | One character across ages, bodies, upgrades and copies | C15 (new); N10 draws crowds from it |
| The Clone Wars (dogfights, fleet battles), The Expanse (Newtonian flight), Edgerunners (car chases) | Vehicle and ship action with flight models, formations, action cameras and battle geography | N18 (new) |
| The Expanse | Gravity and medium as scene physics (zero, low, spin and thrust gravity; vacuum; underwater) | X11 (new) |
| Altered Carbon: Resleeved, Guilty Gear Xrd (GDC 2015), Edgerunners' CG vehicles | Cel-look 3D anime: edited normals, face-shadow maps, outline shells, limited animation per part, impact frames | V12 (new) |
| The Expanse (screens), Altered Carbon (holograms), Edgerunners (HUDs, glitch effects) | In-world motion graphics as editable animated assets | V13 (new) |
| Pixar (Coco's research trips) | Field research and reference shoots with releases | O05 (new) |
| The Expanse (Belter Creole) | A gesture vocabulary built with a linguist, a choreographer and a dialect coach | W07 |
| The Clone Wars (Lucasfilm's Story Group) | Canon across titles and media with a continuity keeper | W02 |
| How to Train Your Dragon (Roger Deakins) | A live-action cinematographer as visual consultant on layout and lighting | O01 |
| Pixar and DreamWorks lighting | Eye lights, rim and kicker cheats linked to one character | X03 |
| Edgerunners and the anime pipeline | Layout check, second key, in-between check and finish check; impact frames and speed lines; fisheye layouts | N09, P04 |
| Arcane | Montage and music-video sequences cut to a song | K02 |
| Arcane and Edgerunners | A designed opening title sequence per series | K05 |
| The Clone Wars | A signature sound lexicon kept across a universe's titles | D04 |
| Every series outsourcing sequences | Work packages for outside freelancers and vendors | S14 |
| Streamer sales | IMF, HDR mastering and immersive audio when a sale requires them | K07 |
The audience correction. Decision 10 recommended a 12+ audience and Elikem accepted all recommendations together; on 2026-09-29 he corrected it: Oshun Synthetics makes work for a Rockstar-level mature audience. Decision 10 now says so (18 by default, 15 where a title fits; violence, language, drug use and dark themes in scope; sexual content within his recorded boundary; hard lines kept), and the tasks built on the mistake were reworked: R06 (standards per rating instead of "safe for work by default"), R18 (screening against the title's rating boundary and the hard lines), R19 (age assurance and age-restriction flags instead of child-safe viewing), Y01 and Y02 (mature shots routed to self-hosted lanes where a hosted provider's terms refuse them), Q06 (adult screening panels) and K07 (ratings and descriptors on every deliverable). O04's portrayal checks stay: they shape how mature themes are shown, not whether they appear. The pilot brief's "safe for work" came from Elikem's own film-track decision (F.00.03, 2026-09-24), not from Decision 10; he decided on 2026-09-29 that the pilot follows the mature standard too, so M01 changes, G10 records it, and G02 asks the owner of the agent-made pilot (F.20.05) to match it so the two pilots stay comparable. Sexual content stays within his recorded boundary (confirmed the same day).
Board changes. Seven new tasks (later until this amendment is approved), 21
amended task bodies, updated epic counts, and every board-map row's dependencies
regenerated from the evidence document so the table matches the board.
Sources, checked 2026-09-29: RedShark News on Arcane's pipeline; Flickering Myth on Ray Park's motion capture for The Clone Wars season 7; ClubJade on Lucasfilm's Story Group; Wikipedia and Space.com on Belter Creole and its gestures; Jim Hill Media on Roger Deakins at DreamWorks; /Film on Coco's research trips; BlenderNation on Arc System Works' GDC 2015 talk; Wikipedia on Altered Carbon: Resleeved; QooApp on Edgerunners' staff credits.
16. Second role audit (2026-09-29)#
Elikem sent the every-role request again and chose a second, independent hunt for
missing roles. GPT-6 Astra at xhigh effort (read-only, through codex --search exec
on the ChatGPT subscription, about 150 searches) read revision 20 and all 291 task
bodies and compared them with productions not yet audited. It found 20 gaps, almost
all about the depth of control or the acceptance test rather than a missing
department name. Claude checked each quoted task contract against the task bodies
before accepting it. Two changes the review proposed as amendments became new
second-milestone tasks (D16, K16) so that no first-milestone task waits on a later
one.
| Production that shows it | What was missing or hollow | Where it now lives |
|---|---|---|
| Primal (Genndy Tartakovsky) | Story that starts in drawings: boards could not create or revise scenes and shots | P01, P02, T09 |
| Studio Ghibli's key and in-between workflow | The 2D lane required an admitted generative model to finish human animation | N09, V07 (manual path first) |
| Into the Spider-Verse | Expressive lines as animated performance controls, not only generated outlines | V14 (new) |
| Klaus | Lighting and texture painted on changing drawings and tracked between them | V15 (new) |
| The Wild Robot, The Bad Guys 2 | Effects shaped and timed like animation over simulations | X12 (new) |
| The Wild Robot | Feather groups and wing folds | C16 (new) |
| Nimona | Transformations inside a shot | N19 (new) |
| Avatar: The Way of Water | Capture judged on fidelity to the performance, not importability | N04, N05 |
| KPop Demon Hunters | A sung performance editable per syllable, note and phoneme | D06 |
| Dune (sound design) | Designed sounds as editable, reusable recipes | D14 (new) |
| Dune (re-recording mix) | Mix automation and dramatic control beyond loudness | D16 (new) |
| Dune (colour) | Secondary grading with masks, tracking and keyframes | K16 (new) |
| Streamer delivery (Netflix specifications) | A fully filled M&E under every dub | D15 (new) |
| Streamer delivery | Picture text per locale and textless elements | K15 (new) |
| All the Light We Cannot See | Accessibility checked with blind, low-vision, deaf and hard-of-hearing viewers | K10, K07, M03 |
| The Mitchells vs. the Machines | Audience screenings of the animatic before production commits | Q13 (new); M03 depends on it |
| LAIKA's archive; the Academy's Digital Dilemma | Preservation after delivery: fixity, copies, migrations | A14 (new) |
| The Boy and the Heron | The campaign and its reveals as a decision | L08 (new) |
| KPop Demon Hunters' licensing | Which extensions of a property should exist (for an adult audience) | Z09 (new, M4) |
| Aardman | Environmental impact beside cost | G13 (new) |
Sources, as cited by the review and dated there: AWN interviews (Primal 2020-10-22, The Mitchells vs. the Machines 2021-05-05); Studio Ghibli's Princess Mononoke production diary; Sony Pictures Imageworks' Spider-Verse account; Cartoon Brew on Klaus (2019-11-14); SIGGRAPH 2025 Talks (DreamWorks); DNEG at SIGGRAPH 2023 (Nimona); Wētā FX on Avatar: The Way of Water (2023-03-30); Netflix Tudum (KPop Demon Hunters 2025-12-22; All the Light We Cannot See 2023-11-03); Motion Picture Association and A Sound Effect interviews on Dune's sound (2022-03); Filmmakers Academy on Dune's grade (2022-01-05); Netflix partner-help delivery specifications; LAIKA's archive pages and the Academy's Digital Dilemma 2; Studio Ghibli's note on The Boy and the Heron (2023-07-10); Netflix on the KPop Demon Hunters toy partnership (2025-10-21); Aardman's sustainability update (2025-06-27).
Board changes. Fifteen new tasks (later until approval), twelve amended task
bodies, three new dependencies (M03 on K10 and Q13, K09 on D15), updated epic counts
and the crew roster, and the board map's dependency cells regenerated from the
evidence document.
17. Four targeted passes (2026-09-29)#
After the second role audit, Elikem chose four narrower passes. Each looks for a different kind of gap than a missing role.
1. The pilot walk-through. Tracing the pilot (M01) through the task graph showed its dependency chain skipped rooms a 40-second, two-character film uses and the M1 exit promises: the boards room (P01), the coverage agent (P05), the boards-to-layout handoff (P08), the lighting room (X03), dialogue alignment (D03), audio QA and captions (D08), assembly (K02), the performance plan (N01), style frames (V04), continuity checks (Q03) and the greenlight its "from idea" journey starts with (I06). M01 now depends on all of them and on N20. Three handoffs were wired: P08 waits on P01 and P03, P01 on the shot list (T09), and render management (X05) on estimates (B07), so no render spends before its estimate (principle 11). The pilot's music is one cleared cue from the audited music lane, since the score department (D05) is M2 and D05's acceptance now names the short instead of the pilot; the pilot stays in the studio until provenance and disclosure exist at M2. The pilot chain grew from 112 to 132 V1.5 tasks.
2. The animation control catalogue (Eve document "V1.5 animation control catalogue (2026-09-29)"). About 45 controls a professional animator uses, from IK/FK matching and space switching to holds, stepped parts, cycles, contact locks, layers and face work, each classed as an operation (a web control, a Blender control and an MCP tool), a native Blender edit captured as a transaction, or an agent assist on its own layer. It also states what the web room offers for animation: decisions such as retiming, holds, exposure, mirroring, pose choice, contacts, range locks and proposals, while posing and curve shaping stay in Blender. N20 builds it into the studio at M1; U10 writes the catalogues for every other department at M2, and U03's parity table reads them.
3. The autonomy readiness matrix (Eve document "V1.5 autonomy readiness matrix (2026-09-29)"). Every department checked for inputs, an automated verifier that is at least partly rule-based, a Q11 scenario and a safe failure path. Seven departments had no scenario (face and lip sync, voice, mix, localisation, campaign, production management, pipeline) and three relied on a model critic alone (art direction, boards, music). Q05 now caps any department at propose until it has a rule-based verifier, a passing scenario and H12's failure path, and names the decisions no setting delegates; H02's descriptors carry each role's verifiers and ceiling; Q11 adds the seven scenarios; Q02 adds music and vocal performance; V08, P02 and D05 gain the rule-based checks.
4. The short's pre-production pack (Eve document "V1.5 short film: pre-production pack (2026-09-29)"). What the West-African-rooted short needs before its canon locks: Ghana's Act 690 names kente and adinkra designs as folklore; the National Folklore Board's notice of 2026-04-01 demands prior authorisation and royalties for films and online video using them, dance or indigenous knowledge, with fees unpublished; kente gained geographical-indication protection in September 2025, so no invented cloth is called "kente"; video on location may need a National Film Authority permit (Act 935). The pack lists Elikem's decisions (source cultures, adaptation, sacred items, languages, music, budget), the consultants, a clearance map classing every element as original, cleared or avoided, a research trip and the order of work. M10 keeps it current and the short (M03) waits on it; R15 and O05 carry the legal specifics.
Board changes. Three new tasks (N20, U10, M10; later until approval), fifteen
amended tasks and thirteen new dependencies; the board map's dependency cells are
regenerated from the evidence document.
18. Feature-gap review (2026-09-29)#
Elikem asked whether any feature was missing. This pass read revision 28, all 309 task bodies in the evidence document, the Isis F-track items the tasks touch and the V1.4 PRD, and searched the whole board for each candidate before calling it a gap. The three role audits had left few crew roles out; what was missing sits around the films: how they reach audiences without losing festival eligibility, the versions each destination needs, the money coming in, control of unreleased copies, the voices under the dialogue, a tool for discovering characters, and running the studio from the personal agent.
| Gap | What was missing | Where it now lives |
|---|---|---|
| Release windows and premiere status | Nothing ordered a title's releases or remembered what had been shown. The short goes to festivals (M09), while on-demand release (L02), syndication (L04), trailers (K11) and the campaign (L08) could put it online first. The Academy's 99th Awards animated short rules bar a short that qualifies by a theatrical run from any non-theatrical exhibition, including the internet, before that run, except up to 15% of its running time (shorts qualifying by an award at a listed festival are exempt), and many festivals require premiere status. Exposure is cumulative, so every excerpt, version and edition counts. | L09 (new: release plan plus an append-only exhibition journal); M09, L02, L04 and L08 depend on it; L05 records each target's rules |
| Short-form versions | One trailer and a 15-second teaser; nothing made the 9:16, 4:5 or 1:1 versions where many viewers first meet animation. Every shot is a 3D scene, so a version is re-framed and re-rendered rather than cropped. | K17 (new); the campaign (L08) and syndication (L04) use it |
| Editions | With 18 as the default rating, a title needs 15-rated, platform and runtime edits; R18 and R19 flag and block, and K02's alternate cuts are creative options. | K18 (new); syndication (L04) uses it |
| Money coming in | Costs were thorough (G07, G12, B09, G13), but nothing recorded incentives, funds, sales or income, and section 13.10 listed sales as covered although no task sold anything. The UK Audio-Visual Expenditure Credit pays 39% of qualifying expenditure for animation (the lower of UK core expenditure and 80% of total core expenditure), subject to British certification and spend rules, and UK expenditure is judged by where goods and services are used or consumed, so the record has to be kept that way from the start. | G07 amended (incentive fields on every cost row, pending allowed, blank refused); G14 (new, M1); L10 sales desk (new, M3); Decision 20 |
| Controlled previews and piracy | The worker rule in every task said nothing leaves the studio before publication, which forbade the screenings Q06 and Q13 already planned and every festival screener. Nothing marked a copy per recipient, expired it or logged access; R04's marking is disclosure, not tracing, and V1.4's PRD records that the repository's in-house video mark does not survive lossy encoding. Nothing registered released titles with YouTube's rights tools or handled takedowns. | G01 amended (a preview permission beside publish); the worker rule reworded in every task; R20 (new: security profiles for studio streams and required uploads, blind-tested forensic marks, piracy response); L05, Q06, Q13 and S14 depend on it |
| Walla, loop group and efforts | No task made background voices: crowds were silent and D01 and D02 voice named characters only. The short's background voices must be in its world's languages, which voice models may not cover, so a hired loop group is the fallback. Efforts (breaths, grunts, exertion) had no timing to contact frames. | D17 (new); the short (M03) depends on it |
| Ensemble improvisation | C03 interviews one character at a time; nothing let several characters play a situation out in character, each limited to what it knows at that story time. | C17 (new) |
| The studio from the personal agent | Director chat was typed only and notifications went to web push and email, while V1.3 already builds an MCP client and approvals bound to the exact action. | S18 (new) |
Checked and already covered (not added): destruction effects (isis-chroma-runpod:F.13.03), grain, halation and letterbox (F.15.04), recaps and cold opens (E06), cut interchange for finishing elsewhere (A09, F.17.07), music cue sheets (D05), credits from the authorship ledger (K05, F.19.03), ideas captured as voice notes (I01), side-by-side review (S04) and E&O readiness (R14).
Open point for M10. If the short's dialogue, not only its walla, uses a language other than English, the language qualification K13 performs at M3 is needed at M2; the pre-production pack (M10) records the languages and G02 reschedules K13 if so.
Decision 20 (section 9) records the revenue recommendation. It takes effect when Elikem approves this revision.
Review. A read-only GPT-6.1 Sol review through codex --search exec on the
ChatGPT subscription returned 13 findings, all folded in: cumulative exposure and an
exhibition journal instead of per-item checks (L09); the credit's 80% cap, the
used-or-consumed test and pending classifications (G07, G14); previews separated from
publication (G01, the worker rule); acceptance tests that prove each new integration
in the amended tasks; destination security profiles for required uploads and a
corrected account of YouTube's Copyright Match Tool (R20); blind forensic extraction
with controls (R20); alignment, the loop-group fallback and constructed-language
review (D17); the 4:5 format and the restyle path (K17); edition producers for
syndication (L04); paired permit-and-refuse tests for deal scope (L10); knowledge
states rather than witnessing (C17); and exact-diff, script-lock and transcript tests
(S18). Checking the Academy's 99th Awards rules directly also removed a claim, taken
from a secondary source, that festival online screenings are exempt; those rules
contain no such exemption.
Board changes. Nine new tasks (later until this amendment is approved); G01 and
G07 amended; seven more task bodies amended with the integrations they must prove (L02,
L04, L05, L08, Q06, Q13, S14); twelve new dependencies from existing tasks (M09, L02,
L04 and L08 on L09; L04 and L08 on K17; L04 on K18; L05, Q06, Q13 and S14 on R20; M03
on D17); seven epics updated; the worker rule reworded in all 333 existing items; and
the board map's dependency cells regenerated from the evidence document. The pilot's
dependency chain is unchanged; the short's grows by six tasks.
Sources, checked 2026-09-29: the Academy's 99th Awards animated short film rules (oscars.org, May 2026); HMRC's Creative Industries Expenditure Credit manual, CREC021000 (qualifying films) and CREC054100 (UK expenditure); GOV.UK guidance on claiming Audio-Visual Expenditure Credits; the BFI's page on UK creative industry expenditure credits; YouTube Help on Content ID eligibility and the Copyright Match Tool; Fable Studio's SHOW-1 paper (2023); V1.4's PRD section 2 on the in-house video watermark.
19. Language decision (2026-09-29)#
Elikem decided that the films stay in English "for a good while before we explore others", and chose to defer translated subtitles and dubbing as well. This closes the open point in section 18: the short's dialogue is English, so no language qualification is needed at M2.
| Task | Change |
|---|---|
| K08 | Now English subtitles for deaf and hard-of-hearing viewers, beside D08's captions; translated subtitles wait. K10 (audio description, needed by the short) still builds on it. |
| K09, K13 | Dubbing and language qualification are deferred: later until Elikem chooses a language, which he does by revising Decision 12 again. |
| D15, K15 | Kept, because every buyer needs a filled M&E and textless elements to dub a title themselves; their per-locale parts wait. D15 no longer waits on K13. |
| M06 | The first season no longer waits on dubbing (K09). |
| D17 | Walla is English by default, or a language of the world Elikem chooses. |
| Q11 | Its localisation scenario runs once a language opens. |
Nothing was removed: the deferred tasks keep their contracts, so opening a language later means revising Decision 12 and moving K09 and K13 back to ready. The pipeline must not assume English beyond its test data (K08), so a language can be added without rework.
20. Performance mining and the open face and body stack (2026-10-05)#
Elikem asked whether a studio can mine character expressions from generated video and apply them to 3D models, whether the same holds for action, sports, dance, grappling, animals and machines, and whether the stack has to rest on Apple's 52-shape set or can be built open and in-house. The answer is one framework, performance mining, with a plug-in per domain. Twelve tasks were filed on the board at his instruction on 2026-10-05, before this revision; this section records what they commit to, what was deliberately not filed, and the one decision they left him, which he made the same day (Decision 21: build own).
20.1 The principle: generate, solve, gate, keep as editable keys#
Generated video is reference, never capture. A monocular solver always emits parameters and never says "this is not a face" or "no body moves like this", so a lane that goes video → solver → rig without a gate produces confident garbage: action-unit combinations no face makes, foot slide, limb swaps and momentum-free jumps that look fine in two dimensions because the video model painted over them. The lane is therefore: generate a short clip conditioned on the character's own render and identity adapter (Y11), never on a prompt that describes a human; solve into the character's own parameter space; gate by re-synthesis against ground truth; store what passes as editable keys on a layer (N13) with provenance; count what fails by failure class. A rejected clip never yields a pose. Dialogue is never mined from a generated mouth: phoneme timing comes from the audio (N06), and the video supplies only the emotional layer (N24).
| Domain | Parameter space | Gate | Generated video is worth it for | Tasks |
|---|---|---|---|---|
| Faces | the 52-shape set now, the studio's own basis later | render residual: re-render the rig with the solved weights and compare, region by region | expression range, stylised faces | N21, N22, N23, N24, N25, Q14, C18, C19 |
| Bodies (action, sports, dance) | the studio's own skeletons (T.11.03), never SMPL until licensed | physics tracking: a simulated character reproduces the motion in MuJoCo under the shot's gravity profile (X11); contact, balance, bone length, joint limits | stunts nobody can perform, superhuman action, crowd variety, choreography previs | N26, N27 |
| Grappling, lifts | two skeletons plus a contact graph | contact consistency plus physics | beats and camera only | refused by N26 until R21 admits a contact solver |
| Animals and creatures | species skeleton | gait grammar: footfall order, duty factor, limb phase per gait; flight and perching for birds | creatures that do not exist | N28 |
| Machines and vehicles | mechanism state and rigid-body dynamics | the simulation is the truth (N18, X11) | look, timing beats, destruction, camera: reference-only under Y01 | no task |
20.2 The 52-shape set and the stack underneath it#
Apple's tracker is closed; the 52 names and their meanings are a published convention that MediaPipe emits under Apache 2 and that most rigs, phones and vendors already speak. It binds nothing and is kept as the import and export contract; product copy never calls it "ARKit". Its limit is coarseness, not ownership. The in-house path (C18) is a FACS-style basis of roughly 100 to 150 shapes plus correctives, sculpted once on a template head and carried to each character by deformation transfer, with the 52 and the 15 visemes kept as projections so every producer and consumer keeps working. The inverse-rendering solve (N23) then fits video straight into that basis with a differentiable renderer, so there is no retarget step and the render residual is the gate itself. The stylisation layer (N25) is where a studio can exceed the research stack, which chases realism: a per-character map from realistic trajectories to the character's acting language, rule-based first and learned only from Elikem's corrections.
Every component is admitted or refused in the rights register before it is wired (R21). Clean: MediaPipe Face Landmarker and Pose, RTMPose, MuJoCo (Apache 2), DeepLabCut (LGPL), PyTorch3D and Mitsuba 3 (BSD), Wan 2.x (Apache 2), NVIDIA Audio2Face-3D (MIT code, NVIDIA Open Model License weights). To read before admission: ICT FaceKit, nvdiffrast, PHC-class physics trackers, the world-grounded monocular solvers (WHAM, TRAM, GVHMR). Refused as they stand: the Max Planck family (FLAME, SMPL, SMPL-X, SMAL, AMASS) without the Meshcapade licence that Isis T.12.07 already parks on, InsightFace-derived models and BUDDI-class contact solvers (non-commercial), AIST++ and most music-to-dance models (research-only), Hunyuan (territory, R05).
20.3 The tasks#
| Task | Title | M | Effort | What it commits to |
|---|---|---|---|---|
| R21 | Face and motion stack licence admission | M1 | S | a register row with an admit or refuse decision for every component above; FLAME and the SMPL family refused under Decision 21 |
| N22 | Face solver on the open stack, with a re-render gate | M2 | M | MediaPipe on CPU, one whole-clip temporal pass, re-render through headless Blender and reject above the Q12 threshold; also N05's non-iPhone path |
| N21 | Expression library mined from generated reference | M2 | M | at least 30 accepted poses per lead with provenance and the rejection rate; reference-only under Y01 |
| N26 | Body solver with a physics-tracking gate | M2 | L | keypoints, lifting, whole-clip pass, MuJoCo tracking under X11; unsolved intervals reported, never filled; defines the body rung on Q01 |
| N27 | Motion-phrase library mined from generated reference | M2 | M | at least 20 accepted phrases for one lead and one non-human character; feeds N15, N16, N10 and N02 without their waiting on it |
| N28 | Gait-grammar gate per species | M2 | S | per-species tables with sources and O01 review; a species without a table is "unchecked", never a pass |
| Q14 | Face performance benchmark rung | M2 | M | rule scores first, then critic, then blind preference; first baseline for N05, N06 and N21 on the cheap binding |
| C18 | Oshun face basis: the studio's own expression set | M3 | L | two leads carry the basis by transfer and pass C07's conformance check; round-trip through the 52 within tolerance |
| N23 | Inverse-rendering solve in the character's own basis | M3 | L | compared with N22's two-step on the Q14 rung and recorded either way; refused without a GPU cap |
| N24 | Audio and video fusion for dialogue takes | M3 | M | closures from audio within the F.11.03 tolerance, brows and lids from video; a conflicting cue is flagged, not averaged |
| N25 | Acting-language stylisation layer | M3 | M | the same take on a realistic and a stylised lead, passing the F.10.05 acting checks and N07's cadence rules |
| C19 | Per-character learned basis from 4D capture | M4 | L | a basis extracted through video-to-mesh's 4DGS conversion, registered to C18, detail kept as a displacement layer |
No existing task gained a dependency, so nothing already approved was re-sequenced; journal notes on N05, C07, N04, N03 and N15 point at the new work. The evidence document (b223ede2, revision 12) carries each task's full body. The board family went from 342 to 354 open tasks.
20.4 What this revision changes elsewhere#
Section 0 (status), 6.7, 6.12, 6.18 and 6.19 (prose and task lists), 7 (the M2, M3 and M4 exits), 9 (Decision 21), 11 (the C, N, Q and R tables) and 13.5 (three roles amended, one added). Decision 21 was decided by Elikem on 2026-10-05: build the studio's own basis and skeletons; the Meshcapade licence is reconsidered only against a named failure, by revising the decision.
21. Text-to-motion on cleared data (2026-10-06)#
Section 6.12 and N12 were written on the assumption that text-to-motion needs the Max Planck body model and motion data (SMPL and AMASS), which are research-only and licensed commercially through Meshcapade (Isis T.12.07). Decision 21 declined that purchase, which left N12 with no route. UniMate (SIGGRAPH Asia 2026, arXiv 2609.05415), read at its sources on 2026-10-06 after Elikem shared a video roundup about it, shows one: a text-conditioned flow-matching model that animates a skeleton of any shape from its rest pose, hierarchy and joint labels, with no body model at all. Its released weights cannot be used; its code and method can.
21.1 What was read#
| Item | Terms | Verdict |
|---|---|---|
Code, github.com/Friedrich-M/UniMate |
MIT, copyright 2026 Linzhan Mou | admit as code (Isis T.01.21) |
Six checkpoints, huggingface.co/Linzhan/UniMate |
CC BY-NC 4.0, every one | refuse; not run even as an internal baseline. The internal_only outcome of Isis T.01.18 covers training-data refusals only, not a weight licence that forbids commercial use |
| UniML3D: Truebones ZOO, 1,097 clips | a commercial pack that may not be redistributed; the model card reports unverified third-party allegations that some of its animal assets come from commercial games | refuse for training and delivery while the dispute stands |
| UniML3D: Mixamo, 2,317 clips | Adobe's Mixamo terms | refuse, as Isis T.01.05 already does |
| UniML3D: Objaverse-XL, 10,355 clips on 7,355 skeletons | each clip carries the licence of its source object, found through the object id | keep CC0 and CC BY after a provenance screen (Isis T.12.08) |
| UniML3D's own captions, joint labels, categories and indexes | ODC-BY 1.0 | admit with attribution |
The model card puts the full model at 74M parameters and about 280 H100-hours of training (8 GPUs for about 35 hours). The cost is compute the studio can buy; the scarce input is cleared data. The card's own limits are the ones this PRD's gates were built for: ground contact slides or hovers, clips are a fixed 2 seconds (longer motion is chained), and quality falls on rare skeletons and unusual motions.
21.2 The route#
| Task | Family | What it commits to |
|---|---|---|
| T.01.21 | Isis | the register rows above, each with the revision read and a re-read date, mirrored in R21 |
| T.12.08 | Isis | a cleared motion corpus for skeletons of any shape: a licence join per clip, CC0 and CC BY only, a person reviewing every clip flagged as a possible game rip, an attribution bundle, and honest counts per body plan, however small |
| T.12.09 | Isis | studio rigs converted to the model's skeleton input and its output brought back as keys, with a person checking joint labels and facing |
| T.12.10 | Isis | the studio's own model trained from the MIT code on the cleared corpus, the T.12.03 presets and, later, motion solved from Elikem's own capture (N26); admitted through the training-data rule; judged on the T.17.02 thresholds per mode (text-to-motion, in-betweening, editing, extension); GPU spend only on Elikem's go with an estimate |
N12 accepts either route: Isis T.12.10, or T.12.01/T.12.02 with the T.12.07 licence if a named failure ever reopens Decision 21. It now covers human and non-human cast, and its clips must pass the T.17.02 foot-slide threshold plus N26's physics tracking for a humanoid or N28's gait grammar for an animal. N14 lists the in-betweening mode of the T.12.10 model as a candidate beside AIS and ARDY. N27 and N28 carry journal notes: the studio gains its first generator for animals and creatures, and N28 becomes its gate. Generated motion stays what section 20 makes all mined and generated performance: an editable proposal on a layer, never final.
21.3 What it teaches the register#
R05 now records the terms revision it read (a model card's commit or a dated capture) and a re-read date, because UniMate's card went from v1 to v3 in nine days and gained its provenance disclaimer only after release. It checks four things separately: the code, the weights, the training data, and the training data's own provenance. A disputed provenance refuses a source until the dispute is resolved, even a pack bought commercially. The video that prompted this section described UniMate as MIT and therefore open, true of the code and false of the weights; licence facts come from the primary source, never from a summary.
21.4 What waits for Elikem#
No new decision row. Three amounts and choices wait in the tasks that need them, as section 9 says they should: the GPU spend for T.12.10 (an estimate first); whether CC BY-SA clips may train a model whose weights the studio keeps (T.12.08 leaves them out until he answers); and, outside this family, a $15 add-on purchase for the Isis mesh bake-off (T.09.08).
21.5 What this revision changes elsewhere#
Section 0 (status), 3.3 (the motion row), 6.12 and 6.19 (prose), 9 (the preamble and Decision 21's tasks), 11 (N12's title and outside dependency) and 13.5 (the creature row). No Lilith task was added and none gained a dependency; the four new tasks are in the Isis family. On the board, N12, N14, R05 and R21 were amended and N12 retitled "Text-to-motion on a cleared model"; the evidence document (b223ede2, revision 13) carries their bodies.