Official Codebase of 3DCodeVerse
Python
rubric: untextured_flat states its operable test — measured wording A/B, adopted
gaoypengf010f02ed54b
diff --git a/harness/codeverse/judges/rubrics/static_object_v1.yaml b/harness/codeverse/judges/rubrics/static_object_v1.yamlindex 846caaa2..d1c0c9c2 100644--- a/harness/codeverse/judges/rubrics/static_object_v1.yaml+++ b/harness/codeverse/judges/rubrics/static_object_v1.yaml@@ -217,7 +217,11 @@ defects:text: The object has TEN or more nameable parts but fewer than THREE distinct refinement kinds (the numbered list in the rubric notes) are visible across all of them together — many primitives, little shape. Count kinds, not instances; do not count colour.penalty: 0.05- id: untextured_flat- text: Default grey / magenta placeholder materials are visible, or surfaces are unlit flat fills with no shading response. A uniform but sensible colour (an all-walnut chair, an all-oak cabinet) is NOT this defect.+ text: Default grey / magenta placeholder materials are visible (the untinted "nothing was+ assigned" look — a defect even when lit), or surfaces are UNLIT flat fills, i.e.+ differently-angled faces render the SAME brightness. "No texture map" is NOT the test —+ a lit uniform sensible colour (an all-walnut chair, an all-oak cabinet) shades normally+ across faces and is NOT this defect.penalty: 0.05cap: 0.75- id: holes_or_inverted_facesdiff --git a/harness/docs/DECISIONS.md b/harness/docs/DECISIONS.mdindex 056f0a19..fa847637 100644--- a/harness/docs/DECISIONS.md+++ b/harness/docs/DECISIONS.md@@ -547,9 +547,14 @@ written) that were accepted because the code works that way and the tests pin itClean-row invariant now MEASURED, not inherited: all 6 controls rebuildbyte-identical under both dials. Forcing the sentence onto clean rows costs −0.092,so it must never be promoted to unconditional — the on-error gate is load-bearing.- Separately fundable, NOT D48's: `untextured_flat` is over-applied by BOTH arms on+ Separately fundable, NOT D48's: `untextured_flat` was over-applied by BOTH arms onshaded models with a uniform sensible colour, which the rubric item's own text- exempts.+ exempts — funded and FIXED 2026-09-01 (wording A/B on the same 22 cases, see+ docs/EVAL.md): the item now states the operable test ("differently-angled faces+ render the SAME brightness"; "no texture map is NOT the test") and keeps default+ grey/magenta a defect even when lit. Both adjudicated-false marks of the+ shaded-uniform class dropped to 0/3, the true default-grey mark held 3/3, and the+ arm's drift is SMALLER than the near-A/A band (elicit−off) — noise, not wording.* **D49 An articulated plan is checked for geometric self-consistency before any code iswritten — measured, ships OFF (2026-08-28/29).** Context: compare_art_v3's low scorersdiff --git a/harness/docs/EVAL.md b/harness/docs/EVAL.mdindex 79786180..4ca97b2e 100644--- a/harness/docs/EVAL.md+++ b/harness/docs/EVAL.md@@ -864,3 +864,26 @@ identity is now measured rather than inherited (all 6 controls rebuild byte-idendials); forcing the sentence onto clean rows costs −0.092, which the on-error gate prevents.Open, not D48's: `untextured_flat` is over-applied by BOTH arms on shaded models with a uniformsensible colour — exactly what the rubric item's own text exempts.++### 2026-09-01 — `untextured_flat` wording: the exemption becomes an operable test (adopted)++Both D48-battery arms over-applied `untextured_flat` to shaded models with a uniform sensible+colour — the case the item's own text exempts; the judges were reading it as "no texture map".+Wording A/B on the same 22-case corpus, OLD = the stored `off` arm ($0), NEW = one re-judge with+the sharpened text (n=3 pro, $3.99), drift calibrated against the near-A/A band (`elicit−off`:+same payload plus one inert sentence).++Pre-registered gates: (1) adjudicated-false marks drop — sv2 violin 2/3 → **0/3**, clean-control+minority 1/3 → **0/3**, but the lapstrake boat holds 2/3 → 2/3 (its per-plank tone variation is+invisible at montage scale — a payload-resolution limit, not wording); (2) the true default-grey+mark holds — office chair 3/3 → 3/3, protected by the new "a defect even when lit" clause;+(3) collateral drift is NOISE: newword−off meanΔ −0.0212 with 4 movers > |0.1| vs the near-A/A+band's −0.0225 with 6 movers, and the big movers are the SAME cases moving the SAME way in both+re-rolls (c02 −0.254 in both; d10 −0.337/−0.338 — the re-rolls both catch the armrest floating+the original `off` roll missed, so the "drift" is the baseline's own vote variance).+`untextured_flat` votes overall: 8 → 5, the drop landing exactly on the adjudicated-false cases.+New text: default grey / magenta placeholder (the untinted "nothing was assigned" look — a defect+even when lit), or UNLIT fills where differently-angled faces render the SAME brightness; "no+texture map" is never the test. `judge_prompt_hash` moves for `static_object_v1`, as any wording+change does; scoring is untouched.+