Official Codebase of 3DCodeVerse
Python
d48 tuning: the slice drawing stops describing itself as a defect
gaoypengc2ee54a8ebc4
diff --git a/harness/codeverse/judges/prompt_builder.py b/harness/codeverse/judges/prompt_builder.pyindex 39401927..d89cc8ca 100644--- a/harness/codeverse/judges/prompt_builder.py+++ b/harness/codeverse/judges/prompt_builder.py@@ -865,7 +865,9 @@ def slice_rig_section(labels: list[str]) -> str:"overlap are hatched red. ""A gap between parts IN THE CUT PLANE is not evidence of disconnection — parts ""may join outside this plane; the measured structure block is authoritative for "- "connectivity.\n" + lines)+ "connectivity. Judge floating_part and holes_or_inverted_faces from the shaded "+ "and geometry views, which see the whole surface; a slice shows one cut only.\n"+ + lines)def judge_prompt_hash(rubric: Rubric) -> str:diff --git a/harness/codeverse/spatial/sections.py b/harness/codeverse/spatial/sections.pyindex 476be329..28f94492 100644--- a/harness/codeverse/spatial/sections.py+++ b/harness/codeverse/spatial/sections.py@@ -626,7 +626,7 @@ def judge_slices(label = g if len(groups[g]) == 1 else f"{g} (x{len(groups[g])})"if drawn_groups[g]:handles.append(Line2D([], [], color=color_of_group[g], linewidth=2,- label=label + " [outline: open section]"))+ label=label + " [outline only — not filled; NOT a hole]"))else:handles.append(Patch(facecolor=color_of_group[g], edgecolor="#333333", label=label))if len(drawn_groups) > JUDGE_SLICE_LEGEND_MAX:diff --git a/harness/docs/DECISIONS.md b/harness/docs/DECISIONS.mdindex 9518776e..c8675d64 100644--- a/harness/docs/DECISIONS.md+++ b/harness/docs/DECISIONS.md@@ -502,12 +502,51 @@ written) that were accepted because the code works that way and the tests pin it("on-error" default | "off", `CV3D_JUDGE__SLICES`), slices appended AFTER themontages/crops with the F2/F3 rig text and the elicitation sentence in the DEFECTCHECKLIST bullet; clean rounds byte-identical (test-pinned), `judge_prompt_hash`- untouched. **Watch item before D48 is called done** (b36_v0_02): the elicitation- sentence changes defect-marking *behaviour*, not just narration — its per-defect- sweep made three samples consistent on render-true defects and the multiplicative- cap stack took the row 0.54 → 0.008 (every added defect was already a minority vote- in the other arms; direction correct, magnitude brutal). The first production- battery must re-check dirty-round defect rates and cap stacking.+ untouched.+ **Watch item ANSWERED 2026-08-31 — the channel stays, two sentences change.** The+ battery ran on a REBUILT corpus (the original scratchpad was destroyed by a /tmp+ cleanup, so this is a fresh selection, not a replay): all 223 `static_object` runs+ under `bench/out` with a GLB + plan re-gated live (92 dirty / 131 clean), then 16+ dirty (12 blender / 4 threejs, 12 batteries, stored 0.052–0.980, 1–37 ERRORs, five+ ≥ 0.88) + 6 clean controls re-rendered on the D47 rig and judged UNPATCHED in three+ arms × n=3 ($11.45, 0 errors). Payload drift vs the measured shim was not re-run —+ it was closed line by line at the implementation review, and the shim is gone.+ (a) **It is the slice IMAGES, not the sentence.** Dirty rows 0.468 (off) → 0.429+ (shipped), marked defects 2.50 → 2.75; the sentence-only arm (`slice_payload →+ ([], True)`, the state a crashed drawing already produces) is FLAT at 0.471 / 2.38 —+ it carries +0.0035 of the −0.039. The mean move is inside noise (paired bootstrap+ 95 % CI [−0.114, +0.035]; 9 down / 5 up / 2 tie), i.e. not a uniform penalty.+ (b) **The cap stack was the wrong thing to fear.** Hard-cap loss FALLS 0.0204 →+ 0.0149 (the additive defect penalty is what rises, 0.189 → 0.204), is exactly 0.000+ on 12 of 16 rows in BOTH arms, no row loses > 0.1 to a cap `off` did not apply+ (max +0.032), and no cap brought by a new mark binds anywhere. b36_v0_02 reproduces+ only weakly — 0.579 → 0.202, of which 0.000 is caps — not the 0.54 → 0.008 that+ wrote this watch item.+ (c) **The shipped arm is the more ACCURATE one, not merely the harsher one.** All 16+ majority-marked disagreements were adjudicated against the rig renders and the exact+ slice PNGs: +6 true marks gained, 0 true marks lost, 6 false `off` marks deleted, 4+ false gained (net true +6 / net false −2). Real wins: `h2h_office_chair`'s armrests+ standing clear of the seat (off missed it 0/3 → 3/3), b36_v0_02's stacked-cylinder+ scroll and painted-on f-holes. Real deletions: the `sv2_violin` volute/C-bouts/+ cut-out f-holes and the `machinist_vise` lofted casting that `off` had called+ primitive.+ (d) **The debt is two sentences, both inside slices-only text** (so clean-round+ identity and `judge_prompt_hash` are untouched): the legend suffix+ `[outline: open section]` named a MESH property and read as a hole report — it alone+ moved `holes_or_inverted_faces` 0 → 2 cases, now `[outline only — not filled; NOT a+ hole]`; and the in-plane caveat was too generic to stop a 2.4 mm open join and+ section-cut islands being marked floating — it now POINTS AT the views that can+ decide ("judge floating_part and holes_or_inverted_faces from the shaded and+ geometry views") rather than forbidding the mark, because a bare prohibition would+ suppress the true marks the channel exists to win. Rejected: a revert (it nets +6+ true marks), dropping the elicitation (inert on dirty rounds, and it produces the+ citations this audit was read from), a cap-interaction guard (no such cap fired).+ Clean-row invariant now MEASURED, not inherited: all 6 controls rebuild+ byte-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+ shaded models with a uniform sensible colour, which the rubric item's own text+ exempts.## Rejected / deferreddiff --git a/harness/docs/EVAL.md b/harness/docs/EVAL.mdindex 459a4f10..79786180 100644--- a/harness/docs/EVAL.md+++ b/harness/docs/EVAL.md@@ -827,3 +827,40 @@ Confounds recorded honestly: this arm ran on 3.6-flash (3.7 outage, 000×3probes), and `21c34c1` transport-retry landed mid-battery (these runs predateit). Validation arm `art_axr_v1` (same briefs, 3.7, both levers live) is inflight; its row decides the lever.++### 2026-08-31 — D48 watch item: it is the images, the caps never bind, and the channel is the more accurate arm++Rebuilt corpus (the original scratchpad was destroyed by a `/tmp` cleanup — a fresh selection,+not a replay): all 223 `static_object` runs under `bench/out` carrying a GLB + plan re-gated live+(0 failures → 92 dirty / 131 clean); 16 dirty + 6 clean controls re-rendered on the D47 14+4 rig+and judged through the UNPATCHED shipped code, three arms × n=3 `gemini-3.1-pro-preview`+($11.45, 0 errors, every row `n_used=3`). Payload drift vs the measured shim was closed at the+implementation review, not re-run.++| dirty n=16 | mean | median | marked/case | hard-cap loss | defect penalty |+|---|---|---|---|---|---|+| `off` (pre-D48 payload) | 0.468 | 0.546 | 2.50 | 0.0204 | 0.1888 |+| `slices` (shipped) | 0.429 | 0.447 | 2.75 | **0.0149** | **0.2040** |+| `elicit` (sentence only) | 0.471 | 0.506 | 2.38 | 0.0148 | 0.1819 |++`elicit` is the shipped `slice_payload` returning `([], True)` — the state a crashed drawing+already produces — so `elicit − off` isolates the SENTENCE and `slices − elicit` the IMAGES. The+sentence is inert on dirty rounds (+0.0035, and it *lowers* marked defects); the images carry the+move. The cap stack is not the mechanism the watch item feared: hard-cap loss FALLS, is exactly+0.000 on 12 of 16 rows in both arms, no row loses > 0.1 to a cap `off` did not apply (max +0.032),+and no cap brought by a new mark binds. b36_v0_02 moves 0.579 → 0.202 (not 0.54 → 0.008) with a+0.000 cap contribution. The mean move is inside noise (bootstrap 95 % CI [−0.114, +0.035];+9 down / 5 up / 2 tie).++Every one of the 16 majority-marked disagreements was adjudicated against the rig renders and the+exact slice PNGs the judge saw: **+6 true marks, −0 true marks, −6 false `off` marks, +4 false+marks** — net true +6 / net false −2. Two of the four false gains are D48's own drawing+describing itself, and both are now fixed in slices-only text: the legend suffix+`[outline: open section]` (read as a hole report; it alone moved `holes_or_inverted_faces` 0 → 2+cases) is now `[outline only — not filled; NOT a hole]`, and the in-plane caveat now points at the+shaded and geometry views for `floating_part` / `holes_or_inverted_faces` instead of denying the+slice — a bare prohibition would suppress the true marks the channel exists to win. Clean-row+identity is now measured rather than inherited (all 6 controls rebuild byte-identical under both+dials); 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 uniform+sensible colour — exactly what the rubric item's own text exempts.diff --git a/harness/tests/judges/test_judge_slices.py b/harness/tests/judges/test_judge_slices.pyindex 25324ea1..9b2152fe 100644--- a/harness/tests/judges/test_judge_slices.py+++ b/harness/tests/judges/test_judge_slices.py@@ -135,6 +135,11 @@ def test_dirty_gates_append_slices_after_crops_with_rig_text_and_one_elicitationtext = text_of(req)assert "After the crops, 2 cross-section slice(s) show the interior" in textassert "A gap between parts IN THE CUT PLANE is not evidence of disconnection" in text+ # the 2026-08-31 watch-item battery: the generic caveat did not stop a 2.4 mm open join+ # and section-cut islands being marked floating_part, so the text names the views that+ # CAN decide instead of merely denying the slice (a bare prohibition would suppress the+ # true marks the channel is there to win)+ assert "Judge floating_part and holes_or_inverted_faces from the shaded" in textassert f"- slice 1: {SLICE_LABELS['front_back']}" in textassert req.system.count(PROVENANCE_ELICITATION) == 1assert 'do not hide one to be kind. ' + PROVENANCE_ELICITATION in req.systemdiff --git a/harness/tests/orchestrator_tracks/test_prompts_assets.py b/harness/tests/orchestrator_tracks/test_prompts_assets.pyindex 16c8a774..ef8909ec 100644--- a/harness/tests/orchestrator_tracks/test_prompts_assets.py+++ b/harness/tests/orchestrator_tracks/test_prompts_assets.py@@ -126,6 +126,7 @@ def test_scene_templates_render_and_asset_stage_with_blender(tmp_ws, settings):assert "buildBollard" in sa and "FIX PASS" in sa[email protected] # the threejs asset check imports three under nodedef test_a_committed_asset_is_reused_on_a_second_stage_run(tmp_ws, settings):"""Review-3 S2 (V3-claim4): re-entering the stage (budget stop after commit, or afailed sibling) reuses a committed, import-clean module without a model call;@@ -150,6 +151,7 @@ def test_a_committed_asset_is_reused_on_a_second_stage_run(tmp_ws, settings):assert len(agent.jobs) > n_jobs and results3["FishingBoat"].strategy != "reused"[email protected] # the threejs asset check imports three under nodedef test_a_replanned_asset_with_the_same_name_is_not_reused(tmp_ws, settings):"""Review-3 S2 sharp edge: after a --force re-plan, a committed module that keptits NAME but changed its plan slice (description/dims) must regenerate — the reusediff --git a/harness/tests/threejs_render/test_studio_rig.py b/harness/tests/threejs_render/test_studio_rig.pyindex 5e862692..e17788a6 100644--- a/harness/tests/threejs_render/test_studio_rig.py+++ b/harness/tests/threejs_render/test_studio_rig.py@@ -103,6 +103,7 @@ def test_studio_render_is_reproducible_and_stamps_the_rig_version(stool_glb: Patassert Path(va.path).read_bytes() == Path(vb.path).read_bytes(), f"{va.name} is not reproducible"[email protected] # renders through the node studio rigdef test_orbit_views_share_one_camera_distance(stool_glb: Path, tmp_path: Path):"""Consistent framing: views in the orbit band (|elevation| <= 60) are pulled towardsone shared camera distance so the object keeps its apparent size across the montage —