Research log

Note

Generated from docs/research/log.md.

A dated record of what was done, in order, with links to the evidence. It is append-only: new entries go at the bottom, and old entries are corrected only by a later entry that says so.

2026-10-07: the RFC and the framing

• Opened RFC 0001, starting from the owner’s framing:
  • Good: Blender is open source.
  • Bad: its whole architecture assumes a human operator.
  • Neither: today’s retrofits (MCP servers, bpy wrappers) sit on facts that no longer hold.
• Resolved R1–R5 in one sitting:
  • the four dead facts (DF1–DF4) and two parked gray zones
  • discovery first
  • two tracks meeting on capability cards
  • a virtual production from free sources
  • the four-pass study method
• Prior evidence carried in from agent-native-lab: a golden-table trace of 5 tasks. Humans needed 26 steps where 5 declarations would do, and 9 of the 26 steps could not run without GUI context.

2026-10-07: discovery corpus

• Discovery 00: which claims in our own earlier work rested on dead facts.
• Discovery 01: a map of the source modules.
• Discovery 02: parallel efforts. No confirmed agent-native fork of Blender exists. ArtCraft chose a Rust rewrite. Upstream’s “made by humans for humans” stance makes upstreaming unlikely.
• Discovery 03: an inventory of free sources per pipeline stage. Real production files were downloaded to ~/newblender-data/ (outside the repo).
• Tools built:
  • tools/registry/ dumps all 2,498 operators and scans the C/C++ source
  • tools/census/ dissects .blend files headless

2026-10-07: the S01 Modeling pilot

All four passes ran on the first stage, plus the distillation:

• Pass 1, workflow script: 73 cited professional steps.
• Pass 2, census: 78 free production files.
  • 91% of mesh objects carry modifiers.
  • 94.5% of faces are quads.
  • The median mesh has 111 vertices.
• Pass 3, trace: 320 modelling operators.
  • 70 poll bodies were read.
  • 178 operators are gated by mode alone.
  • Only 50 (16%) truly need a GUI.
• Pass 4, why: 38 design sources.
• 320 capability cards: 124 declare, 73 free, 67 keep, 56 remove.
• Transformation ledger: L-001 … L-011, all at evidence level L1.
• Stage brief, with a retrospective. The biggest lesson: a name-based heuristic for “is this operator gated?” was wrong half the time. Pass 3 must read poll bodies.

2026-10-07: the showcase

• Picked 20 real Blender works, all verified against sources, covering the whole range of the tool: feature films, open movies, VFX, NPR, a TV series, data visualisation, product viz at scale, a game and fluid simulation.
• Wrote an agent build plan for each one: readiness (3 Today, 11 Partial, 6 Taste-bound), difficulty, a first slice and a six-layer verification plan.
• Added R6, the six-layer verification model.

2026-10-07 → 08: live demos

The agent built three first slices live in Blender 5.2.2 through the Blender Lab MCP bridge. Each checks script was written before its build script.

• 001 Chair: a warm-up. 12/12 checks pass.
• 002 Flow slice: a flooded-forest boat shot in the style of Flow, rendered in EEVEE at 4K.
  • Checks: 11/11, the reproducible fingerprint is identical across a fresh headless rebuild, and the render budget passes.
  • The agent hit and fixed eight problems along the way, and never loosened a threshold. Examples: a camera inside a tree, and haze that turned the frame black.
• 003 Charge slice: a battery-factory bay in the style of Charge.
  • Checks: 11/11 live, and the fingerprint reproduces.
  • Two layer-4 checks fail and are waiting on owner rulings:
    • UV stretch: 12.4% of faces are in tolerance by face count, but 94–99% by area.
    • Fireflies: 24 against a limit of 20, all of them spark pixels.
  • One checker change was flagged in the report.

2026-10-08: capability scoring and the first blind session

• Added R7: five measures, written-before-build checks, and blind quality.
• Built tools/blind/ (4/4 tests). The answer key cannot be read before the owner locks.
• Session s01-flow-charge: 12 images, 6 agent and 6 studio. The owner rated them blind.
  • Mean quality: agent 1.67, studio 4.00.
  • Source guessing: 83% correct.
  • Flow slice: 1.00 blind against a self score of 2.5. Charge slice: 2.33 blind against a self score of 2.5.
  • One Charge frame passed as studio work; one real Flow frame was taken for agent work.

2026-10-08: going public

• Set up branch rails: dev is the integration branch, and main moves by tag only.
• Moved the demo reports into the repo and added a README.
• Made the repo public at mbaneshi/newblender and launched the site (English and Persian), generated from docs/.
• Added a newblender card to mbaneshi.ir and released it as v2.3.1. The deploy sat queued until the self-hosted runner came back, then succeeded; the link is live in all three locales. The Docker image-removal failure from the previous deploy is tracked in distributed-computation/mbaneshi-ir-site#42.
• Started this research section: approach, this log, findings and the roadmap.

2026-10-08: probe P01, region-gated operators headless

The first runtime probe from the roadmap. Full write-up: P01.

• Polls: 0 of the 23 region-gated modelling operators pass their poll headless, and 21 pass once any region is borrowed, even one that has never been drawn.
• Exec outcomes:
  • 5 did their job, verified by a before/after fingerprint.
  • 4 returned FINISHED with real arguments and changed nothing.
  • 5 sculpt gesture operators crashed Blender in their undo push.
• Our own mistake: on the first run, loop_select returned CANCELLED. That looked like evidence against the S01 trace, until reading the exec body showed the probe had not passed the hidden edge_index. With the index it works. The claim stands, and the episode is logged because it is the same anti-pattern seen from the agent’s side.
• Ledger: L-003 and L-004 moved from L1 to L2. New runtime evidence was added to L-008 and L-011.