[{"data":1,"prerenderedAt":395},["ShallowReactive",2],{"finding:p1-mcmc-fix3-negative":3,"finding-runs:p1-mcmc-fix3-negative":257,"finding-related:p1-mcmc-fix3-negative":334},{"meta":4,"impact":54,"sections":60},{"commit":5,"branch":6,"config_paths":7,"id":19,"title":20,"subtitle":21,"eyebrow":22,"date":23,"status":24,"category":25,"polarity":26,"axes":27,"tags":29,"task_code":37,"related_runs":38,"related_findings":50},"c8c984a48703","main",[8,9,10,11,12,13,14,15,16,17,18],"splat\u002Fconfigs\u002F2026-07-mcmc-fix3-lego-noise5e1-smoke.toml","splat\u002Fconfigs\u002F2026-07-mcmc-fix3-lego-noise5e2-smoke.toml","splat\u002Fconfigs\u002F2026-07-mcmc-fix3-lego-noise5e3-smoke.toml","splat\u002Fconfigs\u002F2026-07-mcmc-fix3-lego-noise5e1-iter2000.toml","splat\u002Fconfigs\u002F2026-07-mcmc-fix3-lego-noise5e2-iter2000.toml","splat\u002Fconfigs\u002F2026-07-mcmc-fix3-lego-noise5e3-iter2000.toml","splat\u002Fconfigs\u002F2026-07-mcmc-ablation-A-baseline-iter2000.toml","splat\u002Fconfigs\u002F2026-07-mcmc-ablation-B-scalel1only-iter2000.toml","splat\u002Fconfigs\u002F2026-07-mcmc-ablation-C-relocationonly-iter2000.toml","splat\u002Fconfigs\u002F2026-07-mcmc-final-lego-noise-relocation-30k.toml","splat\u002Fconfigs\u002F2026-07-mcmc-final-lego-relocation-only-30k.toml","p1-mcmc-fix3-negative","G3 MCMC fix3 — 5 連敗確定、scale_l1 犯人 + relocation 30k で 14.6 dB 頭打ち","fix2 (17.38 dB \u002F 30k \u002F capacity 1M 満タン) の敗因を、既実装の 3 欠陥修正版で再挑戦。calibration doc §案 A に沿った noise weight sweep (5e1\u002F5e2\u002F5e3、Lego 500 iter + 2000 iter) は完走したが PSNR 10-12 dB で頭打ち。\u003Cstrong>Component ablation (2000 iter) で scale_eigen_l1=0.001 単独で -4.4 dB の毒性を実測\u003C\u002Fstrong>、これが犯人と特定。しかし scale_l1=0 で relocation + growth schedule のみを Lego 30k で回しても \u003Cstrong>14.65 dB (2h12m, 1.3M splats)\u003C\u002Fstrong>、noise 5e2 + relocation でも \u003Cstrong>14.69 dB (2h38m, 1.8M splats)\u003C\u002Fstrong>、Phase I baseline (36.11 dB \u002F 11m48s \u002F 285k splats) に \u003Cstrong>-21 dB 大敗\u003C\u002Fstrong>。growth schedule で splats が 1.3-1.8M まで肥大化するが PSNR が付いてこない = MCMC relocation の probability-preserving 実装 (Eq.9) と本実装の SoA parameterization が結合していない可能性大。G3 は 5 連敗として卒論 §5.4 negative finding に確定、future work 送り。","P1 G3 · MCMC fix3 · 5 連敗確定 · scale_l1 犯人特定","2026-07-21","stable","negative-finding","negative",[28],1,[30,31,32,26,33,34,35,36],"mcmc","g3","fix3","5-consecutive-failure","scale-eigen-l1","ablation","lego-30k","G3 MCMC fix3 (7 月実行プラン)",[39,40,41,42,43,44,45,46,47,48,49],"lego-mcmc-fix3-noise5e1-smoke","lego-mcmc-fix3-noise5e2-smoke","lego-mcmc-fix3-noise5e3-smoke","lego-mcmc-fix3-noise5e1-iter2000","lego-mcmc-fix3-noise5e2-iter2000","lego-mcmc-fix3-noise5e3-iter2000","lego-mcmc-fix3-ablation-A-baseline-iter2000","lego-mcmc-fix3-ablation-B-scalel1only-iter2000","lego-mcmc-fix3-ablation-C-relocationonly-iter2000","lego-mcmc-fix3-final-relocation-only-30k","lego-mcmc-fix3-final-noise-relocation-30k",[51,52,53],"mcmc-3-defects","mcmc-noise-calibration","p1-axis1-phase-i-scene-adaptive",{"summary":55,"rank":56,"verdict":57,"delta_psnr":58,"delta_wallclock":59},"\u003Cstrong>G3 MCMC fix3 = 5 連敗確定\u003C\u002Fstrong>。fix2 (17.38 dB) 敗因の calibration ズレを、既実装の 3 欠陥修正版で再挑戦。\u003Cstrong>500 iter smoke で crash なし完走\u003C\u002Fstrong>、\u003Cstrong>2000 iter で 10-12 dB 頭打ち\u003C\u002Fstrong>、\u003Cstrong>ablation で scale_eigen_l1 の毒性を -4.4 dB と測定\u003C\u002Fstrong>、\u003Cstrong>scale_l1 除去した relocation + growth schedule の 30k full 2 パターンで PSNR 14.65 \u002F 14.69 dB\u003C\u002Fstrong> (Phase I baseline 36.11 dB 比 -21 dB)。犯人は複合: (1) scale_eigen_l1 実装が過剰罰化 or 符号ミス、(2) relocation の Eq.9 probability-preserving 実装が本実装の SoA parameterization と結合していない (splats 1.3-1.8M まで肥大化するが PSNR が付いてこない)。\u003Cstrong>7 月実行プラン G3 の timebox 判定基準通り即撤退\u003C\u002Fstrong>、卒論 §5.4 negative finding として 5 連敗目に確定、future work 送り。\u003Cstrong>ただし本 negative finding は従来の 4 連敗より格段に強い evidence\u003C\u002Fstrong> (component ablation で犯人を数値で切り分け、30k full の decisive failure まで到達)、卒論 measurement contribution として §5.4 で厚く記述可能。","high","5-consecutive-failure-scope-out","Phase I 36.11 dB 比 -21 dB (relocation-only) \u002F -21 dB (noise-relocation)、fix2 17.38 dB 比 -2.7 dB (更に劣化)","Phase I 11m48s 比 +11x (relocation-only 2h12m) \u002F +13x (noise-relocation 2h38m)",[61,64,69,72,172,174,181,183,225,227,236,238,244,246],{"type":62,"text":63},"lead","G3 MCMC fix3 は 3 段階 (smoke → iter2000 sweep → ablation → final 30k) を timebox 内で完遂し、\u003Cstrong>scale_eigen_l1 が -4.4 dB の毒性を持つ\u003C\u002Fstrong>と 2000 iter ablation で決定的に切り分けた。しかし scale_l1 除去版の relocation + growth schedule を Lego 30k full で回しても \u003Cstrong>PSNR 14.65 dB (Phase I 比 -21 dB)\u003C\u002Fstrong>、noise 追加版でも 14.69 dB で MCMC は本実装の scope 外と確定。5 連敗目として卒論 §5.4 に格上げ。",{"type":65,"label":66,"variant":67,"text":68},"callout","Headline (G3 MCMC = 5 連敗確定、scale_l1 犯人特定 + relocation 30k で頭打ち)","warning","\u003Cstrong>3 段階 timebox 実験の結果\u003C\u002Fstrong>: (1) 500 iter smoke sweep (noise 5e1\u002F5e2\u002F5e3) は完走 & no NaN で fix2 敗因の calibration bug は解決確認、(2) 2000 iter sweep で PSNR 10-12 dB 頭打ち、weight 依存なし、(3) ablation (A baseline \u002F B scale_l1 only \u002F C relocation only、各 2000 iter) で \u003Cstrong>B が 10.85 dB と A\u002FC の 15.2\u002F15.7 dB より -4.4 dB 劣化\u003C\u002Fstrong>、scale_eigen_l1 が犯人と特定。(4) scale_l1=0 で 30k full を 2 パターン (relocation-only \u002F noise-relocation) 実施、\u003Cstrong>両方とも 14.65-14.69 dB で Phase I baseline 36.11 dB に -21 dB 大敗\u003C\u002Fstrong>、splats は 1.3-1.8M まで肥大化するが PSNR が付いてこない。\u003Cstrong>G3 は 5 連敗として scope-out\u003C\u002Fstrong>、卒論 §5.4 で「MCMC 3 欠陥修正 + component ablation + 30k final の decisive failure」という format で future work 化。",{"type":70,"text":71},"heading","1. 実験 timeline (2026-07-21、~5h total)",{"type":73,"columns":74,"align":82,"rows":85,"caption":171},"table",[75,76,77,78,79,80,81],"段階","config","iter","PSNR","splats","wallclock","判定",[83,83,84,84,84,84,83],"left","right",[86,94,99,105,113,119,124,131,137,143,151,157,165],[87,88,89,90,91,92,93],"1. smoke","noise5e1-smoke","500","-","n\u002Fa","8s","完走 & no NaN、trajectory non-monotonic だが健全",[87,95,89,96,97,92,98],"noise5e2-smoke","11.92","9,123","完走",[87,100,89,101,102,103,104],"noise5e3-smoke","11.97","10,060","7s","完走、fix2 と同 weight で crash なし",[106,107,108,109,110,111,112],"2. iter2000","noise5e1-iter2000","2000","9.74","53,587","56s","頭打ち、weight 依存なし",[106,114,108,115,116,117,118],"noise5e2-iter2000","10.03","61,374","76s","同上",[106,120,108,121,122,123,118],"noise5e3-iter2000","10.29","54,181","59s",[125,126,108,127,128,129,130],"3. ablation","**A baseline (all-OFF)**","**15.23**","149,725","63s","**refine 経路健全性確認**",[125,132,108,133,134,135,136],"**B scale_l1 only**","**10.85**","63,358","46s","**scale_l1 単独で -4.4 dB 毒性、犯人特定**",[125,138,108,139,140,141,142],"**C relocation only**","**15.68**","205,613","75s","**relocation + growth は 2000 iter では健全**",[144,145,146,147,148,149,150],"4. final 30k","**relocation-only**","30000","**14.65**","1,315,605","2h12m","**scale_l1=0 でも 30k で頭打ち、Phase I 比 -21 dB**",[144,152,146,153,154,155,156],"**noise-relocation**","**14.69**","1,840,604","2h38m","**noise 5e2 追加も改善なし、splats 肥大化のみ**",[158,159,160,161,162,163,164],"参考","Phase I baseline (fast conv.)","12500","36.11","285,000","11m48s","本実装の勝ちパターン、比較基準",[158,166,146,167,168,169,170],"fix2 (2026-05-23 敗)","17.38","1,000,000","1h10m","capacity 満タン、scale_l1=0.001 + noise=5e3 の複合毒","5h の timebox 内で 3 段階 (smoke → iter2000 → ablation → final 30k) を実施。\u003Cstrong>ablation で scale_eigen_l1 の -4.4 dB 毒性を決定的に切り分け\u003C\u002Fstrong>、しかし scale_l1 除去でも 30k で 14.65 dB で頭打ち、5 連敗確定。",{"type":70,"text":173},"2. 敗因の分解 (2 つの独立した bug)",{"type":175,"ordered":176,"items":177},"list",true,[178,179,180],"\u003Cstrong>Bug (a): scale_eigen_l1 の実装が過剰罰化 or 符号ミス\u003C\u002Fstrong> — ablation B で 2000 iter 10.85 dB (A: 15.23 dB) と、他 component OFF でも scale_l1=0.001 単独で -4.4 dB の毒性。\u003Ccode>regularize.rs::apply_mcmc_scale_l1\u003C\u002Fcode> の勾配符号 or scaling factor の再検証が必要。論文 §3.4 は \u003Ccode>‖Σ‖_eig_L1\u003C\u002Fcode> の subgradient を \u003Ccode>sign(exp(log_scale))\u003C\u002Fcode> = 常に +1 として mean_scale が単調減少するはずだが、本実装は logit space で計算しており挙動が微妙にずれている可能性。\u003Cstrong>次に MCMC を再挑戦する場合の primary suspect\u003C\u002Fstrong>。","\u003Cstrong>Bug (b): relocation + growth schedule が Phase I に劣る\u003C\u002Fstrong> — ablation C 2000 iter で 15.68 dB (A: 15.23 dB) と一見拮抗するが、30k full まで伸ばすと \u003Cstrong>14.65 dB で頭打ち\u003C\u002Fstrong>、Phase I 36.11 dB に -21 dB 大敗。\u003Ccode>relocate.rs::apply_mcmc_relocation\u003C\u002Fcode> の Eq.9 probability-preserving (\u003Ccode>o_new = 1-(1-o_t)^(1\u002FN_t)\u003C\u002Fcode>) は unit test で pin しているが、本実装の SoA parameterization (raw_opacities, log_scales, rot_quats を独立 buffer) との rotation-aware Σ 補正が単純化されており、cluster が多数生成される 30k で品質が付いてこない。\u003Cstrong>splats 1.3-1.8M まで肥大化するが PSNR が上がらない\u003C\u002Fstrong>のはこの兆候。","\u003Cstrong>Bug (a) と (b) は独立\u003C\u002Fstrong>: fix2 (2026-05-23) の PSNR 17.38 dB は両者の合成毒 (scale_l1 -4.4 dB + relocation 収束不良) で、component ablation でようやく分離可能に。この分解結果が本 finding の卒論 §5.4 での decisive evidence。",{"type":70,"text":182},"3. 5 連敗の詳細 (卒論 §5.4 用の年表)",{"type":73,"columns":184,"align":191,"rows":192,"caption":224},[185,186,187,188,189,190],"失敗 #","日付","tag","内容","PSNR (Lego)","決着",[83,83,83,83,84,83],[193,200,205,210,217],[194,195,196,197,198,199],"1","2026-05-22","`!G#4.31.G.1`","3DGS-MCMC noise を prune 便乗経路で移植","3-5 dB 発散","commit 857012a で future work 化",[201,195,202,203,204,118],"2","`!G#4.31.G.2`","capacity 100k cap で再挑戦","OOM \u002F crash",[206,195,207,208,209,118],"3","`!G#4.31.G.3`","max_scale_cap=0.05 追加","22 dB 停滞",[211,212,213,214,215,216],"4","2026-05-23","`!G#4.31.G.4` (fix2)","3 欠陥全部入り、scale_l1=0.001 + noise=5e3","17.38 dB","calibration doc `mcmc-noise-calibration.md` 起草、次回課題送り",[218,219,220,221,222,223],"**5**","**2026-07-21**","**`!G#4.31.G.5` (fix3)**","**smoke → iter2000 → ablation → 30k final の 4 段階 timebox 実験**","**14.65 dB**","**本 finding、5 連敗確定 + component ablation で犯人特定**","MCMC 5 連敗の年表。本 fix3 (5 連敗目) は ablation で scale_eigen_l1 の -4.4 dB 毒性を切り分けた点で従来 4 連敗より格段に強い negative finding。卒論 §5.4 で「measurement-driven negative finding」の pattern として厚く記述可能。",{"type":70,"text":226},"4. 卒論 §5.4 への組み込み方針",{"type":175,"items":228},[229,230,231,232,233,234,235],"\u003Cstrong>Sub-section 5.4.X: MCMC 3 欠陥修正の実装と 5 連敗\u003C\u002Fstrong>","  \u003Cstrong>Motivation\u003C\u002Fstrong>: 論文 Kheradmand et al. 2024 の 3DGS-MCMC は brush + gsplat が未実装、本実装が axis 1 hard contribution 候補として着手。","  \u003Cstrong>実装\u003C\u002Fstrong>: `docs\u002Ffindings\u002Fmcmc-3-defects.md` に沿って `McmcConfig` + `relocate.rs::apply_mcmc_relocation` + `train_loop` 独立 tick + growth schedule を実装。unit test で Eq.9 probability-preserving を pin。","  \u003Cstrong>敗因 4 連続 (2026-05-22〜23)\u003C\u002Fstrong>: OOM \u002F crash \u002F 22 dB 停滞 \u002F fix2 17.38 dB。calibration doc で次回課題送り。","  \u003Cstrong>fix3 実験 (2026-07-21、本 finding)\u003C\u002Fstrong>: 4 段階 timebox (smoke → iter2000 → ablation → 30k final)。\u003Cstrong>component ablation で scale_eigen_l1 の -4.4 dB 毒性を測定\u003C\u002Fstrong>、scale_l1 除去した relocation + growth を 30k full で 14.65 dB。","  \u003Cstrong>敗因の分解\u003C\u002Fstrong>: (a) scale_eigen_l1 実装の勾配符号 or scaling factor に bug、(b) relocation Eq.9 と SoA parameterization の結合不足。","  \u003Cstrong>卒論 contribution\u003C\u002Fstrong>: 「MCMC 3 欠陥修正の 5 連続失敗 + component ablation + 30k final の decisive failure」という measurement-driven negative finding として、Apple Silicon native 実装の scope 境界を明示化。ROI 階層 calibration narrative (`docs\u002Fstrategy\u002F2026-06-25-next-steps-deep-dive-v2.md`) の A (Reframing) 補強 evidence。",{"type":70,"text":237},"5. Future work",{"type":175,"items":239},[240,241,242,243],"\u003Cstrong>scale_eigen_l1 の勾配符号 audit\u003C\u002Fstrong>: `regularize.rs::apply_mcmc_scale_l1` の実装を論文 Eq.11 と 1 対 1 で照合、logit space vs linear space の scaling を確認。1 日工数で fix 可能性あり、卒論 body 執筆と並走可能。","\u003Cstrong>relocation Eq.9 の rotation-aware Σ 補正\u003C\u002Fstrong>: 現状は scale 1\u002Fsqrt(N_t) の単純近似、論文補遺の完全 Σ_new (rotation 含む) 実装は 2-3 日工数。fix3 敗因 (b) の直接対処。","\u003Cstrong>msplat (rayanht) の MCMC 経路との比較\u003C\u002Fstrong>: msplat は densify + refine のみで MCMC 未実装、gsplat も同じ。\u003Cstrong>MCMC 実装は open field\u003C\u002Fstrong>で、本実装の 5 連敗 + ablation 結果は community contribution としても価値がある (E track の PR ネタ)。","\u003Cstrong>W2 (Mip-NeRF 360 実写) に workstream 切替\u003C\u002Fstrong>: G3 撤退により 7 月残り時間を全て G1 に投入可能、W2 前倒しで 7\u002F8 以前に着手検討。",{"type":70,"text":245},"6. 関連",{"type":175,"items":247},[248,249,250,251,252,253,254,255,256],"3 欠陥仕様: \u003Ccode>docs\u002Ffindings\u002Fmcmc-3-defects.md\u003C\u002Fcode>","fix2 敗因 calibration doc: \u003Ccode>docs\u002Ffindings\u002Fmcmc-noise-calibration.md\u003C\u002Fcode>","Phase I baseline (比較対照): \u003Ccode>docs\u002Ffindings\u002Fp1-axis1-phase-i-scene-adaptive.toml\u003C\u002Fcode>","7 月実行プラン (G3 timebox 判定基準): \u003Ccode>docs\u002Fstrategy\u002F2026-07-execution-plan.md\u003C\u002Fcode>","reframing narrative: \u003Ccode>docs\u002Fstrategy\u002F2026-06-25-next-steps-deep-dive-v2.md\u003C\u002Fcode>","3 実装アーキ比較 (msplat\u002Fbrush\u002Fsplat): \u003Ccode>docs\u002Farchitecture\u002Fthree-impls-arch.md\u003C\u002Fcode>","Config 一覧 (11 本、smoke×3 + iter2000×3 + ablation×3 + final30k×2): \u003Ccode>splat\u002Fconfigs\u002F2026-07-mcmc-*.toml\u003C\u002Fcode>","Runner scripts (4 本): \u003Ccode>splat\u002Fscripts\u002Frun-mcmc-*.sh\u003C\u002Fcode>","Run logs: \u003Ccode>runs-logs\u002Fmcmc-*-20260721-*.log\u003C\u002Fcode>",[258,271,278,285,292,299,306,313,320,327],{"id":45,"title":45,"subtitle":259,"date":260,"workspace":261,"tags":262,"verdict":266,"psnr":267,"psnr_unit":-1,"wallclock":268,"splats":269,"summary_url":270,"detail_path":270},"G3 MCMC fix3 smoke #1 — ablation A (all-MCMC OFF baseline) (fix2 の 1\u002F10) で 2000 iter で PSNR 20 dB 到達を判定","2026-07-01","splat",[30,31,32,263,264,265],"smoke","noise-sweep","lego","partial",15.232376098632812,"1m 2s",149725,"\u002Fruns\u002Flego-mcmc-fix3-ablation-A-baseline-iter2000\u002F",{"id":46,"title":46,"subtitle":272,"date":260,"workspace":261,"tags":273,"verdict":266,"psnr":274,"psnr_unit":-1,"wallclock":275,"splats":276,"summary_url":277,"detail_path":277},"G3 MCMC fix3 smoke #1 — ablation B (scale_l1=0.001 only) (fix2 の 1\u002F10) で 2000 iter で PSNR 20 dB 到達を判定",[30,31,32,263,264,265],10.84475040435791,"46.49s",63358,"\u002Fruns\u002Flego-mcmc-fix3-ablation-B-scalel1only-iter2000\u002F",{"id":47,"title":47,"subtitle":279,"date":260,"workspace":261,"tags":280,"verdict":266,"psnr":281,"psnr_unit":-1,"wallclock":282,"splats":283,"summary_url":284,"detail_path":284},"G3 MCMC fix3 smoke #1 — ablation C (relocation only, noise ≈0) (fix2 の 1\u002F10) で 2000 iter で PSNR 20 dB 到達を判定",[30,31,32,263,264,265],15.682197570800781,"1m 14s",205613,"\u002Fruns\u002Flego-mcmc-fix3-ablation-C-relocationonly-iter2000\u002F",{"id":49,"title":49,"subtitle":286,"date":260,"workspace":261,"tags":287,"verdict":266,"psnr":288,"psnr_unit":-1,"wallclock":289,"splats":290,"summary_url":291,"detail_path":291},"G3 MCMC final — noise 5e2 + relocation (scale_l1=0)、fix3 犯人除去版 で Lego 30k、Phase I との差分で MCMC 効果を測定",[30,31,32,263,264,265],14.684809684753418,"2h 38m 2s",1840604,"\u002Fruns\u002Flego-mcmc-fix3-final-noise-relocation-30k\u002F",{"id":48,"title":48,"subtitle":293,"date":260,"workspace":261,"tags":294,"verdict":266,"psnr":295,"psnr_unit":-1,"wallclock":296,"splats":297,"summary_url":298,"detail_path":298},"G3 MCMC final — relocation only (scale_l1=0, noise=5e0)、ablation-C 30k full で Lego 30k、Phase I との差分で growth schedule 効果を測定",[30,31,32,263,264,265],14.646121978759766,"2h 12m 29s",1315605,"\u002Fruns\u002Flego-mcmc-fix3-final-relocation-only-30k\u002F",{"id":42,"title":42,"subtitle":300,"date":260,"workspace":261,"tags":301,"verdict":266,"psnr":302,"psnr_unit":-1,"wallclock":303,"splats":304,"summary_url":305,"detail_path":305},"G3 MCMC fix3 smoke #1 — noise 5e1 (fix2 の 1\u002F100、calibration doc §案 A の下限側) で 2000 iter で PSNR 20 dB 到達を判定",[30,31,32,263,264,265],9.736172676086426,"55.95s",53587,"\u002Fruns\u002Flego-mcmc-fix3-noise5e1-iter2000\u002F",{"id":43,"title":43,"subtitle":307,"date":260,"workspace":261,"tags":308,"verdict":266,"psnr":309,"psnr_unit":-1,"wallclock":310,"splats":311,"summary_url":312,"detail_path":312},"G3 MCMC fix3 smoke #1 — noise 5e2 (fix2 の 1\u002F10) で 2000 iter で PSNR 20 dB 到達を判定",[30,31,32,263,264,265],10.031878471374512,"1m 15s",61374,"\u002Fruns\u002Flego-mcmc-fix3-noise5e2-iter2000\u002F",{"id":40,"title":40,"subtitle":314,"date":260,"workspace":261,"tags":315,"verdict":266,"psnr":316,"psnr_unit":-1,"wallclock":317,"splats":318,"summary_url":319,"detail_path":319},"G3 MCMC fix3 smoke #1 — noise 5e2 (fix2 の 1\u002F10) で 500 iter loss trajectory 観察",[30,31,32,263,264,265],11.921218872070312,"8.01s",9123,"\u002Fruns\u002Flego-mcmc-fix3-noise5e2-smoke\u002F",{"id":44,"title":44,"subtitle":321,"date":260,"workspace":261,"tags":322,"verdict":266,"psnr":323,"psnr_unit":-1,"wallclock":324,"splats":325,"summary_url":326,"detail_path":326},"G3 MCMC fix3 smoke #1 — noise 5e3 で 2000 iter で PSNR 20 dB 到達を判定",[30,31,32,263,264,265],10.288583755493164,"59.43s",54181,"\u002Fruns\u002Flego-mcmc-fix3-noise5e3-iter2000\u002F",{"id":41,"title":41,"subtitle":328,"date":260,"workspace":261,"tags":329,"verdict":266,"psnr":330,"psnr_unit":-1,"wallclock":331,"splats":332,"summary_url":333,"detail_path":333},"G3 MCMC fix3 smoke #1 — noise 5e3 で 500 iter loss trajectory 観察",[30,31,32,263,264,265],11.971680641174316,"7.16s",10060,"\u002Fruns\u002Flego-mcmc-fix3-noise5e3-smoke\u002F",[335,359,380],{"id":53,"title":336,"date":337,"status":24,"polarity":338,"category":339,"axes":340,"tags":341,"task_code":350,"related_runs":351,"delta_psnr":354,"delta_wallclock":355,"rank":56,"verdict":356,"impact_summary":357,"detail_path":358},"Phase I scene-adaptive iter budget — **STRONG Pareto improvement** 確定 (8 scene mean +0.077 dB at -24% wallclock vs Phase D)","2026-05-26","positive","design",[28],[342,343,344,345,346,347,348,349],"p1-axis1","phase-i","scene-adaptive","pareto-front","stop-iter","sh-progressive","universal-improvement","calibration","P1 Phase I",[352,353],"chair-phase-i-adaptive-12500","materials-phase-i-adaptive-12500","**+0.077 dB** (8 scene mean vs Phase D 30k)、+0.701 dB (vs brush)、-0.031 dB (vs G.3 30k = within noise)","**-24.4%** (vs Phase D)、**-31.6%** (vs G.3 30k)、both Pareto dimensions improved","scene-adaptive-pareto-confirmed","Phase H で Lego stop_iter=12500 が Pareto sweet spot だが ficus\u002Fmic で大幅 fail (-4.12 \u002F -6.03 dB) と scene-dependent 確定後、**Phase I で chair \u002F materials @ stop_iter=12500 も fast converger と確認** (G.3 30k 比 -0.05〜-0.07 dB のみで essentially same)。これにより \u003Cstrong>scene-adaptive iter budget\u003C\u002Fstrong> (Lego\u002Fchair\u002Fmaterials @ 12500、ficus\u002Fdrums\u002Fhotdog\u002Fmic\u002Fship @ 30000) を構成、\u003Cstrong>既存 8 scene 全データ点が揃った状態で 8 scene mean を算出\u003C\u002Fstrong>。\u003Cstrong>結果\u003C\u002Fstrong>: scene-adaptive 8 scene mean **33.561 dB** \u002F total wallclock **3h 50m 10s**、Phase D 30k 比 \u003Cstrong>+0.077 dB \u002F -24.4% wallclock\u003C\u002Fstrong>、G.3 30k 比 \u003Cstrong>-0.031 dB (within noise) \u002F -31.6% wallclock\u003C\u002Fstrong>、brush 比 \u003Cstrong>+0.701 dB\u003C\u002Fstrong>。\u003Cstrong>Pareto front の両軸で improvement\u003C\u002Fstrong>: Phase D を quality + speed 両方で dominate、G.3 30k quality を 1\u002F3 短時間で達成。**axis 1 contribution の最終形**: kernel-level 直叩きではなく \u003Cstrong>scene-adaptive iter budget + sh_progressive + opacity_decay の組み合わせ\u003C\u002Fstrong>が Apple Silicon native Metal 最適化の universal Pareto improvement。卒論 §5.4.7 末尾 + §6 future work で本 Phase I を明示、scene-adaptive を新 universal default として推奨。","\u002Ffindings\u002Fp1-axis1-phase-i-scene-adaptive\u002F",{"id":52,"title":360,"date":212,"status":24,"polarity":26,"category":361,"axes":362,"tags":363,"task_code":368,"related_runs":369,"delta_psnr":376,"delta_wallclock":-1,"rank":56,"verdict":377,"impact_summary":378,"detail_path":379},"A.2 MCMC 検証で発覚した noise gate 不整合と L1 全滅 segfault","experiment",[28],[364,30,365,349,366,263,367],"phase-5","sgld-noise","regression","segfault","A.2",[370,371,372,373,374,375],"lego-mcmc-30k","mcmc-l1-only-smoke","mcmc-noise-sh3-smoke","mcmc-combo-iter-bisect","mcmc-combo-500","mcmc-l1-500","2.5 dB (sh=3 + 全部入り、50 iter で発散)","investigative","SGLD gate を paper 式に揃えた結果 mean_noise_weight が ~50-150x スケールズレし、calibration 補正 (5e5→5e3) でも iter 240 前後で SIGSEGV。Bisect smoke で真因が L1 全滅 → refine prune → 空 buffer crash というアルゴリズム順序問題と判明。Calibration ≠ correctness。","\u002Ffindings\u002Fmcmc-noise-calibration\u002F",{"id":51,"title":381,"date":195,"status":382,"polarity":383,"category":384,"axes":385,"tags":386,"task_code":368,"related_runs":391,"delta_psnr":-1,"delta_wallclock":-1,"rank":392,"verdict":377,"impact_summary":393,"detail_path":394},"A.2 MCMC 法の完全実装 — 3 設計欠陥の整理 (spec)","draft","neutral","spec",[28],[364,30,387,384,388,389,390],"sgld","relocation","scale-l1","opacity-l1",[370],"mid","本実装の MCMC が論文と乖離している 3 箇所 (5% incremental growth 欠如、λ_Σ\u002Fλ_o covariance\u002Fopacity 正則化欠如、relocation が refine prune に便乗) を整理し、A.2 の修正項目と検証条件を確定させた spec。","\u002Ffindings\u002Fmcmc-3-defects\u002F",1790227326287]