Discovery 00 — What earlier work already found (and what it assumed)

Note

Generated from docs/discovery/00-prior-work.md.

• Date: 2026-10-07 · Feeds: RFC 0001 R2 (“build on prior discussions, don’t restart”)
• Method: a read-only agent reviewed agent-native-lab (ANL) issues and files, mytodo issues and sessions, and the ChatGPT transcript ~/mytodo-data/chatgpt/2026-10-04-blender-mcp-eval/transcript.md. Paths are relative to ~/agent-native-lab unless noted.

A. Human lens: how professionals work

• Studio pipeline (transcript ~2618–2720): Editorial/Previz → Concept → Asset Creation (Model/Shade/Rig) → Shot Assembly → Animation → Lighting → Render → Review.
• Studio asset pipeline: task layers with an owner per data element, ending in a “Published Asset”. A production artifact = Intent + Source + Task layers + Dependencies + State + Validation + Review + Version + Publish, not a single .blend.
• Studio modelling flow: Reference → Setup → Blocking → Modifiers → Topology → UV → Handover. Each stage serves the next, so an agent needs “downstream awareness”.
• Human-to-Artifact decomposition (ANL#7): per stage, what the human sees / knows / decides / manipulates / verifies, and what Blender changes underneath. “Human UI is a baseline, not a specification.”
• Sources (ANL#18, knowledge/blender/sources/SOURCES.md, 139 tiered sources):
  • Studio Tools/Pipeline docs, rated the richest agent-readable account of pro work.
  • blender-studio-tools (asset_pipeline, blender_kitsu, CloudRig).
  • About 25 Studio trainings (Scripting for Artists, Fundamentals, Geometry Nodes from Scratch).
  • Production files: Sprite Fright, Wing It!, the Vault. Kitsu and Watchtower for production state.
• Human UI summary: knowledge/blender/MASTERY.md §1 covers the HIG paradigms (select → operate, non-modal), modes, active vs selected, keymaps and redo.

B. Source lens: modules and architecture

• Chain (transcript ~2347): DNA → RNA → Data → Operators → Depsgraph → Nodes → Evaluation → Render.
• Four interfaces: Human, Python, Internal, AI. The depsgraph is a ready-made “causal dependency model”; Geometry Nodes are Blender’s own move toward declarative authoring.
• Proposed stack (~2960–3020): semantic 3D model → Blender IR/DSL → RNA/BMesh/Nodes → core.
• knowledge/blender/MASTERY.md: 4 layers, 143 path:line citations against v5.2.2. Covers operators (~2000 types; redo = undo + re-exec), ID data-blocks, RNA/DNA, original vs evaluated depsgraph, GeometrySet/fields, BMesh vs Mesh, undo.
  • Its conclusion: prefer declarative layers; operators are a UI contract, not a function API; verify through evaluated state.
• No full module inventory existed before Discovery 01.
• Pinned source: v5.2.2 d13f752, searchable with bun _tools/search-sources.ts.

C. Methods and status

Method	Where	Status
Phase ladder: observe → map → expose → compose → benchmark → missing primitives → extend → modify core	ANL#4, #23 (W0–W7), #14	Only W0 (sources) done
End-goal hypotheses	ANL#22 → vision/blender/END-GOAL.md	Not written
Knowledge map: 4 tiers, executable knowledge, failures → knowledge	ANL#5	Tiering applied; map not written
Four-layer trace + golden table	ANL#7; lab/blender/golden-table.md	Done for 5 tasks; sculpt/UV/rig and real Info-log sessions not done
Operability baseline, T6 animation	ANL#32, #35	Closed. Data/RNA won as control surface; official MCP saw 2/6 and 0/5 of the needed state
Primitives v0: apply / inspect / render / checkpoint / verify, JSON-only	ANL#37; lab/blender/operability/2026-10-06-primitives-decision.md	Closed
Agent experiment A1	ANL#43	Closed
Benchmarks A–F, MCP bake-off, community atlas	ANL#3, #20, #21, #19	Not started

D. Parallel efforts already touched

• storytold / ArtCraft spike (mytodo#358; ~/mytodo/docs/sessions/2026-10-06-storytold-spike/).
  • Result: FilmCraft produced byte-identical MP4s, but the project can’t be reopened after an OTIO import (mytodo#365).
  • Decisions: FilmCraft sits next to Resolve; EffectCraft is a candidate; PhotoCraft gets one batch spike; ArtCraft no.
• “Model supplies data, reviewed code executes” pattern across engines: fusion_kit v0, Chrome engines v0, mytodo Blender loop.
• Blender agent projects inventoried (ANL#6, #16): mcp-blender, BlendRelay, LL3M, blender-open-mcp, Blender Agent Studio. Patterns to borrow: BlenderProc, AYON, Sverchok.

E. Prior claims resting on the dead facts (RFC R1): re-examine, don’t inherit

Dead fact	Prior claims built on it
1. Human is the operator / viewport is the core loop	MASTERY’s “F9 handle” for the human; golden table measured against human steps; transcript architecture starts at “HUMAN → Intent”; vision/PRINCIPLES.md #21 (“human creative judgment remains…”); ANL#45 treats human review hours as the main cost
2. Humans write code	“No model-generated runtime Python”: templates and dispatcher are human-reviewed trusted code (ANL#37, fusion_kit, loop.py); W1 = lessons for a human learner
3. Tiny imperative steps	Transcript’s move_to() / align() one-call ops; ANL#2’s many small per-task agents. The trace itself argues the opposite: declare outcomes.
4. The human eye verifies	“Validate = visual inspection”; ANL#14 Phase 7. The operability work already moved the gate to evaluated state, with render comparison as a warning only.

Keep vs drop:

• Keep:
  • The evidence (traces, operability results, source corpus).
  • The methods (four-layer trace, tiered sources).
  • The safety lesson behind “no model-generated code”: the retrofit MCPs run LLM code “without any guards”.
• Drop: using the human path as the yardstick for what newblender should be. Human steps stay a baseline for measurement only.