[{"data":1,"prerenderedAt":229},["ShallowReactive",2],{"run:multiscene-materials-30k-q":3,"run-findings:multiscene-materials-30k-q":191},{"run":4,"config":20,"metrics":100,"curve":106,"assets":190},{"id":5,"title":5,"subtitle":6,"eyebrow":7,"date":8,"workspace":9,"commit":10,"tags":11,"verdict":19},"multiscene-materials-30k-q","静穏再実行 (materials, G.3-alone 30k recipe, n=1, load-controlled)","Run summary · P1.M5 multi-scene parity (低 wallclock + 低 splats 版)","2026-05-25","splat","63266e4",[12,13,14,15,16,17,18],"p1-d","multi-scene","brush-compat","premultiplied","opacity-decay","phase-d-multi","materials","partial",[21,24,27,30,33,36,39,42,45,48,51,54,57,60,63,66,69,72,75,78,81,83,86,89,91,94,97],{"key":22,"value":23},"dataset","\u002FUsers\u002Fotkrickey\u002Fdev\u002F3dgs-workspace\u002Fdatasets\u002Fnerf_synthetic\u002Fmaterials",{"key":25,"value":26},"gt_convention","Premultiplied",{"key":28,"value":29},"iterations","30,000",{"key":31,"value":32},"seed","42",{"key":34,"value":35},"capacity","1,000,000 splats",{"key":37,"value":38},"sh_degree","3",{"key":40,"value":41},"loss","L1Ssim",{"key":43,"value":44},"lambda","0.200",{"key":46,"value":47},"ssim","window=7 sigma=1",{"key":49,"value":50},"backend.backward","Simd",{"key":52,"value":53},"backend.loss_path","Gpu",{"key":55,"value":56},"backend.emit_pairs","Original",{"key":58,"value":59},"backend.forward","F32",{"key":61,"value":62},"backend.dispatch","PerStage",{"key":64,"value":65},"backend.radix_scan","GpuHierarchical",{"key":67,"value":68},"backend.radix_scatter","SimdBallot",{"key":70,"value":71},"backend.binning","Resident",{"key":73,"value":74},"backend.tile_sort","GlobalRadix",{"key":76,"value":77},"backend.submission","Chained",{"key":79,"value":80},"backend.target_cache","true",{"key":82,"value":80},"backend.shader_fast_math",{"key":84,"value":85},"os_build","26.6.2 (25G83)",{"key":87,"value":88},"backend.ssim_clear","HostMemset",{"key":90,"value":88},"backend.raster_grad_clear",{"key":92,"value":93},"backend.backward_uniform","HostBuffer",{"key":95,"value":96},"backend.lazy_c3_scratch","false",{"key":98,"value":99},"backend.ssim_walk","PerChannel",{"psnr":101,"wallclock":102,"wallclock_regress":103,"splats":104,"final_loss":105},29.99651336669922,"5m 19s",false,354917,"5.361909e-3",{"loss":107,"splats":172},{"iters":108,"values":140},[109,110,111,112,113,114,115,116,117,118,119,120,121,122,123,124,125,126,127,128,129,130,131,132,133,134,135,136,137,138,139],1,1000,2000,3000,4000,5000,6000,7000,8000,9000,10000,11000,12000,13000,14000,15000,16000,17000,18000,19000,20000,21000,22000,23000,24000,25000,26000,27000,28000,29000,30000,[141,142,143,144,145,146,147,148,149,150,151,152,153,154,155,156,157,158,159,160,161,162,163,164,165,166,167,168,169,170,171],0.6085893511772156,0.047659359872341156,0.021587636321783066,0.014302583411335945,0.011365393176674843,0.009930282831192017,0.009014909155666828,0.008568278513848782,0.007927220314741135,0.00767429219558835,0.007335001602768898,0.007072465494275093,0.006792791653424501,0.006800057832151651,0.0065331538207829,0.006343063898384571,0.0062012141570448875,0.006058741360902786,0.005889873951673508,0.00592569075524807,0.005782597698271275,0.005663007497787476,0.005676166154444218,0.005576420575380325,0.005502097308635712,0.005389836151152849,0.005409201607108116,0.0053529292345047,0.00531209260225296,0.005300597753375769,0.005361909046769142,{"iters":173,"values":174},[109,110,111,112,113,114,115,116,117,118,119,120,121,122,123,124,125,126,127,128,129,130,131,132,133,134,135,136,137,138,139],[175,176,177,178,179,180,181,182,183,184,185,186,187,188,189,104,104,104,104,104,104,104,104,104,104,104,104,104,104,104,104],5207,7239,158203,272293,328935,351495,362192,367190,368387,367178,365239,363050,360379,358296,356369,{"has_renders":103,"has_splat":103},[192],{"id":193,"title":194,"date":195,"status":196,"polarity":197,"category":198,"axes":199,"tags":201,"task_code":213,"related_runs":214,"delta_psnr":223,"delta_wallclock":224,"rank":225,"verdict":226,"impact_summary":227,"detail_path":228},"p3-final-stack-v2","M7 v2 — 最終スタック独立再計測 (post S4b) は session 汚染で +6.4%、paired control 122.77s を維持。多 scene (ficus\u002Fmaterials\u002Fdrums) で退行なし確認","2026-08-31","stable","neutral","experiment",[109,200],3,[202,203,204,205,206,207,13,208,209,210,211,18,212],"m7","m7v2","seminar05-milestones","final-stack","msplat-parity","pareto-v3b","regression-check","measurement-methodology","session-contamination","ficus","drums","2026-09-seminar05-milestones M7 v2 (最終スタック独立再計測の再実施 + 多 scene 退行チェック)",[215,216,217,218,219,220,221,5,222],"finalv2-12500-r1","finalv2-12500-r2","finalv2-12500-r3","multiscene-ficus-30k","multiscene-materials-30k","multiscene-drums-30k","multiscene-ficus-30k-q","multiscene-ficus-30k-q2","Part1: finalv2 self-eval n=3 mean **36.257dB** (σ0.10%)、brush cross-eval q8 n=3 mean **36.289dB** (σ0.11%、M7比+0.04dB)。Part2 (§6、load汚染下 n=1): ficus -0.020dB \u002F materials -0.027dB \u002F drums -0.010dB (いずれも a-10 σ±0.32dB の1\u002F10以下、退行なし)。PSNR は load非依存の量なので §14 の静穏再実行でも §6 の値 (ficus 34.261dB \u002F materials 29.998dB) から変更しない。§14 の追加観測: materials は baseline\u002F§6\u002F静穏の3 runでPSNR range **0.028dB** (安定)。ficus は §6\u002Fq1\u002Fq2 の3 runで range **1.443dB** (34.261\u002F32.836\u002F34.279、a-10 のLego range 0.885dBの1.6倍) — config\u002FRNG起因は排除済み、a-10のGPU非決定性の一事例と推測されるが、実用上の帰結は「a-10のσ±0.32dBをficusの退行判定に転用できない」というgate閾値の限界","Part1: as-measured 12.5k n=3 mean **130.657s** (σ1.727%、session汚染のためmilestones非採用)。paired control (S4b on-arm 122.768s \u002F rnd off-arm 122.6467s、相互0.13%一致) を現在値として維持。msplat等価gap **0.745×** (旧0.755×)、throughput **0.916×** (旧0.926×)。Part2 (§6、load汚染下): drums (load概ねクリーン) 3936s→687.85s=**5.72×**。ficus\u002Fmaterials は load汚染下で方向感のみ (≥3.7×\u002F≥4.4×)。§14 静穏再実行 (in-runサンプラーで確認): ficus 349.63s→**254-256s** (-27.3%〜-26.7%、q1\u002Fq2で再現)、materials 421.90s→**319.54s** (-24.3%)","high","measurement-methodology-finding-controls-retained-multiscene-no-regression","2026-09-seminar05-milestones M7 のフォローアップを実施した。**Part 1 (12.5k n=3 再計測)**: `2026-08-30-finalv2-12500-r{1,2,3}.toml` (final-12500-r{1,2,3}.toml の複製、backend flag 無指定で S4b flip 後の default = ssim_reduce=Gpu を含む全採用施策込み) を実行。各 run の `[backend] BackendConfig {...}` ログ行で `ssim_reduce: Gpu, radix_scatter: SimdBallot, refine_scratch_persist: true, refine_direct_write: true` を確認、AC 電源・pgrep 空・binary sha256\u002Fgit rev 不変 (`c804b349...`\u002F`9bb4ea5`) を series 前後で確認。**as-measured mean は 130.657s (σ1.727%)** — S4b flip で速くなるはずが M7 (124.39s) より遅いという逆方向の結果になった。追加調査で、(1) 10s 間隔 load sampler が r1 window 内で 5.77 まで load1 上昇 (iter6000-7000 相当のタイミング)、r2 window 内でも 4.22 まで上昇、(2) old M7 (r2-r4) と new (r1-r3) の 1000-iter segment ごとの ms\u002Fiter を突き合わせると iter0-3000 は new が -4〜-9% 速い (S4b の効果と整合) のに iter3000-12500 は new が一貫して +1.8〜+12.0% 遅い、(3) 実行後の `ps aux` で Finder (147%CPU)・別プロジェクト (otksynth) の Rust バイナリ・音声処理プロセスの実存を確認、という 3 系統の証拠が「同時実行された無関係なユーザープロセスによる session 汚染」を裏づけた。既存の汚染判定規約 (`p3-tier1-bit-exact.toml` §5「load 証拠 AND 自 arm 内 3σ 超」) はこの事例 (3 run 全体が相関して底上げされ、単一の outlier が無い) を検出できない構造的な穴であることも判明した (§2 に規約拡張案を記載)。**同日の paired control 2 本 (S4b gate on-arm 122.768s、`p3-refine-nondrain` off-arm 122.6467s — 後者は refine_gpu_snapshot という no-op field 追加のみの差で本 session の HEAD と事実上同一コード) が相互に 0.13% で一致**しており、この 2 点を「現在の真値」の推定として採用し、milestones.md の「現在地」は上書きしない。brush cross-eval q8 PSNR (final.ply への `splat eval` 実行、load 非依存) は 3 run 平均 **36.289dB** (M7 の 36.249dB から +0.04dB、ノイズ帯)。avg splats (trapezoid 積分による自前算出) は 321,011 (M7 の 321,637 と 0.2% 差で一致、算出法の再現性を裏づけ)。msplat 等価 gap は control 値 (122.768s) ベースで **0.745×** (旧 0.755×)、throughput は **0.916×** (旧 0.926×) とわずかに改善。Pareto v3b (`docs\u002Ffigures\u002Fpareto-v3b-msplat-lego-final-stack-v2.png`) は 12.5k 点の x 座標を control 値、y 座標を finalv2 の brush cross-eval 実測に更新 (両者は異なる run 由来なので script docstring に明記)。**Part 2 (多 scene breakage screen)**: 2026-08-30 に採用された M4 (radix simd_ballot)・A7 (refine 3-pass 削減)・S4b (ssim_reduce=gpu) の 3 施策は全て Lego 12.5k gate のみを根拠にしており、他 scene での検証が無かった。`p1-axis1-phase-g-pareto-landscape.toml` §2 (G.3-alone 30k recipe、2026-05-25、[trainer]\u002F[refine]\u002F[adam] 節が現行 config と完全一致することを diff で確認済み) を baseline に、ficus (227k splats)・materials (355k splats)・drums (998k splats、Lego 407k の 2.5 倍で 3 施策が splat 数 non-uniform に依存する経路を初めてテストする scene) の 3 scene を n=1 で再計測した。**ΔPSNR は ficus -0.020dB \u002F materials -0.027dB \u002F drums -0.010dB で、いずれも `a-10-variance-baseline` の σ±0.32dB (Lego 30k での run-to-run noise floor) の 1\u002F10 以下** — 3 scene とも退行なし。M4\u002FA7\u002FS4b はいずれも理論上 PSNR に影響しない施策 (M4: sort parity bit-exact、A7: 書き込みロジック不変、S4b: gradient 非変更) のため、この結果は「今回の 3 施策の改善効果」ではなく「壊れていないことの確認」として設計通り。Δsplats も ficus +0.32% \u002F materials +1.50% \u002F drums -0.31% と小さい。wallclock は ficus (load 9.6-32.6 で終始汚染)・materials (load 7-27 で終始汚染) は方向感のみ (旧比 ≥3.7×\u002F≥4.4× 高速化) だが、**drums は iter 27000 まで load 0.8-5.5 とほぼクリーンで 3936s→687.85s = 5.72× 高速化を実測**、Lego の 2.5 倍規模の scene でも大幅改善が維持されることを確認した。","\u002Ffindings\u002Fp3-final-stack-v2\u002F",1790227325086]