Skip to content

Official Codebase of 3DCodeVerse

Python

docs: D82 records the verdict A/B — nothing broke, and five prompts cannot resolve a gain

gaoypeng18e8861cdeb5
diff --git a/harness/docs/DECISIONS.md b/harness/docs/DECISIONS.md
index 310f0978..832d7095 100644
--- a/harness/docs/DECISIONS.md
+++ b/harness/docs/DECISIONS.md
@@ -883,7 +883,10 @@ Pointers: EVAL, PAPER_WRITING = `eval/docs/*.md`; COST, RUNBOOK, ARCHITECTURE, I
empty) went; the URDF build collides the rest pose only (D17) and writes no `articulation.json`;
`GateReport.of` builds every report whose verdict is its findings. Deleted as unused (owner):
`effect_ablation`, `compare_silhouette` / `texture_pass` / `texture_preview` (0/1/0 calls in 616
- sessions; 16 tools remain), `glsl_cookbook.md`.
+ sessions; 16 tools remain), `glsl_cookbook.md`. Small A/B against its base (2026-09-22; 3 urdf + 2
+ scene prompts, `--rounds 1`, one run each, inside a 503 storm): articulated picked 0.75 / 0.61 / 0.50
+ → 0.63 / 0.70 / 0.59 (mean 0.621 → 0.642), scenes 0.03 / unjudged → 0.09 / 0.47 — nothing broke, and
+ n this small cannot resolve a gain (runs in `~/3dcodeverse_runs/ab_verdict_2026-09-22`).
* **D83 A vendor session gets its prompt on stdin, sees only routed skills, and is read back from its
own record (2026-09-22).** Every CLI reads the prompt from stdin, byte-identical (a 5-zone scene prompt
passed gemini's 2 000-line `read_file` cut when it went through `task_prompt.md`; argv caps an argument