Official Codebase of 3DCodeVerse
Python
Commits on main
flywheel: the refine-pairs review sweep — textconv RCE, the real brief, the physical battery
c0b84f1 ziyao committed 2 months ago
prompts: the CV3D_LEAN_PROMPT review — narrow the parts table only on all-part rounds
08e9d40 ziyao committed 2 months ago
tools: close the review of the verdict/failed split — and bound the payload
933b8ec ziyao committed 2 months ago
prompts: CV3D_LEAN_PROMPT drops the material the articulated agent already has (-45.7 % baseline)
3272f18 ziyao committed 2 months ago
flywheel: export refine rounds as (brief -> diff -> delta) training pairs
92ed4a6 Ziyao Zeng committed 2 months ago
tools: a FAIL verdict is no longer an MCP error — 63% of joint_sweep calls were
bebe0f8 ziyao committed 2 months ago
planner: a degenerate plan is re-sampled from the original prompt, not edited in context
8f302d2 ziyao committed 2 months ago
docs: D51 — the effect library is harness-owned starter code; post chain; ablation probe
f2926e8 gaoypeng committed 2 months ago
scene host: a post chain — GTAO, selective glow bloom, grade (default ON for scenes)
f73c76e gaoypeng committed 2 months ago
scene: the effect library — 52 harness-owned modules ship in every scene workspace (D51)
5f05441 gaoypeng committed 2 months ago
docs: the sceneloop ledger — efficiency ranked, one A/B inconclusive, two hypotheses refuted
d1ae6b0 gaoypeng committed 2 months ago
skills: the scene-graphics recipe port — atmosphere, water, night, materials (R25-R28)
7797b37 gaoypeng committed 2 months ago
rubric: untextured_flat states its operable test — measured wording A/B, adopted
f010f02 gaoypeng committed 2 months ago
judge/texture/zone fixes from the PR #3 review sweep — every finding reproduced or refuted first
83408ba gaoypeng committed 2 months ago
articulated/scene: three review-of-PR#3 fixes, each reproduced first
69bf30a gaoypeng committed 2 months ago
Merge PR #3: articulated wave — measured levers (OFF), sweep aggregation, buried-link gate, planner re-asks, provider-failure markers (D49, D50)
07ce9bb Yipeng Gao committed 2 months ago
eval: after the runaway is trained out, greedy beats sampling again
6ba733e gaoypeng committed 2 months ago
stop-DPO: the termination pairs transfer, and generating per-model ones is not worth it
01a820f gaoypeng committed 2 months ago
Merge origin/main into the articulated wave — upstream owns the axis repair and the loose root_link match; this branch keeps what it lacks
26a4a13 ziyao committed 2 months ago
report: condense to what changed how we work; archive the blow-by-blow
c344991 gaoypeng committed 2 months ago
toolkits: a full-scan audit for the pair datasets, and the rule it exists to enforce
f7c702c gaoypeng committed 2 months ago
d48 tuning: the slice drawing stops describing itself as a defect
c2ee54a gaoypeng committed 2 months ago
toolkits: the pair pipeline, image shrinking, and a rule about deleting apparent duplicates
ae72b27 gaoypeng committed 2 months ago
docs: CLAUDE.md — 19 registry tools, filtered per track; the test counts stop being a number to maintain
754de1d gaoypeng committed 2 months ago