Roadmap
What comes next, in order of when we expect to do it. Nothing here is a commitment to write engine code: the process stays RFC → discovery → spec → plan, and the spec needs the owner’s approval.
Now: turn L1 claims into L2 evidence
Section titled “Now: turn L1 claims into L2 evidence”Most ledger entries are documented and traced (L1) but not yet reproduced at runtime (L2). These are the cheap probes from the S01 brief §8. Results go into the log.
| Probe | Question | Ledger |
|---|---|---|
| Undo | Do scripted operator calls really push no undo step? | L-008 |
| Done, 2026-10-08: P01. The poll passes 21/23 with a borrowed region; 5 operators did the job, 4 failed silently and 5 crashed. Follow-ups: a drawn-region baseline, and gesture point lists. | L-004, L-011 | |
| Twin parity | Do “exact” declarative twins give identical meshes? Tested with an isomorphism compare. | L-006, L-010 |
| Stroke replay | Can a sculpt brush stroke be replayed headless? | L-009 |
Next: owner rulings and a second blind session
Section titled “Next: owner rulings and a second blind session”- Owner rulings on the two open Charge checks:
- Should UV stretch be weighted by area?
- Should known emitters be masked out of the firefly count?
- Quality iteration. Rebuild the Charge slice aiming at the levers the blind session pointed to (lighting contrast, emissive cells, smoke in EEVEE). Then run a second blind session with more items, to test P3 and P4 in findings.
- “Today” slices. Build one of the three works predicted Today (procedural still life, Molecular Nodes or product viz at scale). This calibrates the readiness predictions against outcomes, as R7 asks.
Then: the second pipeline stage
Section titled “Then: the second pipeline stage”- Apply the S01 retrospective first:
- read poll bodies (not names)
- record the operator id for each workflow step
- name a counter-corpus up front (hard-surface / archviz)
- split the trace pass across parallel agents
- Candidate stages:
- S04 Geometry Nodes, which is closest to the declarative core
- S05 Rigging, which is likely the most human-shaped, and so the best test of P1
Later: towards a spec
Section titled “Later: towards a spec”- Once enough stages have data, decide RFC 0001 Q4 (reuse vs. rewrite) and Q5 (relationship to upstream).
- Write the newblender spec: the agent-native shape across stages, starting from the S01 proposal.
- RFC 0002: decide whether and how to publish the discovery work as a research series.
Parked (gray zones)
Section titled “Parked (gray zones)”- G1: fine-tuning models to act like a human in 3D, or to judge art the way a human would.
- G2: a self-learning lab that improves from its own runs.