Official Codebase of 3DCodeVerse
Python
docs: D47 finals — the shipped grouping re-measured, Cplus and t=0 rejected with numbers
gaoypeng2eae1e75da82
diff --git a/harness/docs/DECISIONS.md b/harness/docs/DECISIONS.mdindex 5b393f6e..e3f29aa0 100644--- a/harness/docs/DECISIONS.md+++ b/harness/docs/DECISIONS.md@@ -448,9 +448,9 @@ written) that were accepted because the code works that way and the tests pin itbaseline); A2 = production-faithful 8 views + clay; B = 14 labelled 640 px singleviews; C = 14 views + clay through our montage machinery (this decision). Theonly comparison that survives multiplicity is same-cap C−A **+0.038** (n=29,- p=.009); the raw deltas C−B p=.041 and C−A p=.024 do NOT survive Holm — cap flips+ SE 0.013, t≈2.98); the raw deltas C−B p=.041 and C−A p=.024 do NOT survive Holm — cap flipsmake the raw delta distribution heavy-tailed (sd(d) ~0.10–0.14). The cap story- is the visible one: clean-item `untextured_flat` caps 11(A) / 6(A2) / 8(B) /+ is the visible one: `untextured_flat` cap-rule fires 11(A) / 6(A2) / 8(B) /**3(C)** — more sky and underside pixels stop the judge inventing flatness.B rejected: $0.198 & 67.5k input tok/verdict (C: $0.155 / 40.9k), its −0.025 vs A2is entirely cap flips (~half contradicting the pixels on eyeball), and its@@ -459,8 +459,23 @@ written) that were accepted because the code works that way and the tests pin it14/14 naive TP measures reading comprehension of the shared gate text, not payloadquality), and B's floating lead reduces to one genuinely visual item. Flashreplicas (3.7 and 3.6): payload Δ ≈ 0 — the rig pays only at pro tier, which is- the verdict judge anyway. Underside full-res singles (the espresso catch) are- queued as the C+ experiment; its battery's numbers get appended here when it lands.+ the verdict judge anyway. The SHIPPED grouping was itself re-measured (arm Cprod, $7.2, same 42 × n=3):+ montage grouping ALONE moves the mean −0.044 vs C's accidental grouping — as large+ as the rig effect itself — because leading montage 1 with the poles ends C's+ dirty-item leniency (the 11-error windmill: C 0.51 → Cprod 0.04; clean/dirty gap+ 0.134 → 0.155) and lands ≈ A2 overall (−0.016, inside noise). A payload experiment+ must therefore measure the exact grouping it ships; this entry's operative numbers+ are Cprod's. Underside full-res singles (arm Cplus: Cprod + bottom and+ front_right_low appended as 768 px singles, $6.8): recovers the espresso-class+ underside catch (0.795 with render_artifacts flagged, vs 0.965 blind in Cprod) but+ degrades everything else — worst clean/dirty gap 0.112, highest row-σ 0.040, and+ the windmill relapses to 0.62 with all three samples dropping the floating and+ primitive reads — measured and REJECTED; a variant that replaces the semantically+ odd bottom ground-band crop instead of appending stays queued. Temperature+ (A2 @ t=0, $6.0): σ 0.037 → 0.022 with 8/42 rows fully deterministic — temperature+ is ~40 % of re-judge noise — but the mean shifts −0.037 and pearson(gate errors)+ drops −0.195 → −0.126: re-keying every σ threshold to buy half the noise is a bad+ trade; **t = 0.2 stays**.## Rejected / deferreddiff --git a/harness/docs/EVAL.md b/harness/docs/EVAL.mdindex 61747a64..55d02aa2 100644--- a/harness/docs/EVAL.md+++ b/harness/docs/EVAL.md@@ -642,12 +642,19 @@ files; old records whose stored overall contradicts their own criterion scores)production-faithful 8 views + clay · B = 14 labelled 640 px single views · C = 14-view rig+ clay through the montage machinery (5 montages + 2 crops). C mean overall 0.6835 vs A0.6427 / A2 0.6558 / B 0.6308; the only multiplicity survivor is same-cap C−A +0.038-(n=29, p=.009) — raw C−B p=.041 / C−A p=.024 do not survive Holm because cap flips make-the deltas heavy-tailed. Clean-item `untextured_flat` caps: 11(A) / 6(A2) / 8(B) / 3(C).+(n=29, SE 0.013, t≈2.98) — raw C−B p=.041 / C−A p=.024 do not survive Holm because cap flips make+the deltas heavy-tailed. `untextured_flat` cap-rule fires: 11(A) / 6(A2) / 8(B) / 3(C).B rejected at $0.198 & 67.5k tok/verdict: its deficit is entirely cap flips (~halfcontradicting the pixels), and a defect-provenance pass over every PRESENT vote showed itsdefect-hunter halo was text-quoting — interpenetration pure-view TP is 0/14 in EVERY arm(the naive 14/14 measures reading comprehension of the shared gate text), and B's floatinglead reduces to one genuinely visual item. Flash replicas: payload Δ ≈ 0 on both 3.7 and-3.6 — the rig pays only at pro tier. C ships as D47 ($0.155/verdict, 40.9k input tok);-underside full-res singles (the espresso catch) are queued as the C+ follow-up.+3.6 — the rig pays only at pro tier. C ships as D47 ($0.155/verdict, 40.9k input tok) — with the SHIPPED+grouping re-measured (arm Cprod, $7.2): grouping alone moves the mean −0.044 vs C's+accidental grouping (windmill 0.51 → 0.04; clean/dirty gap 0.134 → 0.155; ≈ A2 overall at+−0.016) — a payload experiment must measure the exact grouping it ships. Underside+full-res singles (arm Cplus, $6.8): recovers the espresso underside catch (0.795 ++render_artifacts vs 0.965 blind) but posts the worst clean/dirty gap (0.112), the highest+row-σ (0.040) and a windmill relapse to 0.62 — rejected; the replace-the-bottom-crop+variant stays queued. Temperature A2 @ t=0 ($6.0): σ 0.037 → 0.022 (8/42 rows exactly+deterministic) but mean −0.037 and pearson(gate errors) −0.195 → −0.126 — t = 0.2 stays.