0000000000000000000000000000000000000000 d3e6da7802486f0af6f649a198debc8f3495d169 c0decafe <220505+c0decafe@users.noreply.github.com> 1775092786 +0200 commit (initial): Add proposal feasibility review and scope summary d3e6da7802486f0af6f649a198debc8f3495d169 d3e6da7802486f0af6f649a198debc8f3495d169 c0decafe <220505+c0decafe@users.noreply.github.com> 1775095321 +0200 reset: moving to HEAD d3e6da7802486f0af6f649a198debc8f3495d169 e82236624b8aa7c591511a07a69c8d0a110daf9d c0decafe <220505+c0decafe@users.noreply.github.com> 1775095335 +0200 commit: Refactor builder lifecycle and derivation-backed training pipeline e82236624b8aa7c591511a07a69c8d0a110daf9d b0c33d867dd1b1bdb6bc954a4eea47c07359bfbf c0decafe <220505+c0decafe@users.noreply.github.com> 1775095432 +0200 commit: Add repo bootstrap and media inventory tooling b0c33d867dd1b1bdb6bc954a4eea47c07359bfbf cb277126c6fe497887facb5f871719b5ee8a0b25 c0decafe <220505+c0decafe@users.noreply.github.com> 1775095442 +0200 commit: Add benchmark materials and reference planning docs cb277126c6fe497887facb5f871719b5ee8a0b25 9c8b61033e214df0b1e8173d2f37d97a56061e6e c0decafe <220505+c0decafe@users.noreply.github.com> 1775095455 +0200 commit: Add internal planning checkpoints and market research 9c8b61033e214df0b1e8173d2f37d97a56061e6e 93cbdf1317cee2695127c477bedcfe38fda6b872 c0decafe <220505+c0decafe@users.noreply.github.com> 1775096231 +0200 commit: Wrap all tool scripts as derivation-backed commands 93cbdf1317cee2695127c477bedcfe38fda6b872 dc5b643a9c276d988229ed5de96acc6c87ecd05a c0decafe <220505+c0decafe@users.noreply.github.com> 1775096920 +0200 commit: Rename direct builder commands to match task namespaces dc5b643a9c276d988229ed5de96acc6c87ecd05a 28872260257de6047f40cf87f2b8da6f84ab9f3f c0decafe <220505+c0decafe@users.noreply.github.com> 1775168278 +0200 commit: Prune non-canonical builder image workflows 28872260257de6047f40cf87f2b8da6f84ab9f3f c0ce04b61c21a243583b6ce67b39932777bf983d c0decafe <220505+c0decafe@users.noreply.github.com> 1775169733 +0200 commit: Prune non-canonical training and builder entrypoints c0ce04b61c21a243583b6ce67b39932777bf983d 68b30a6b7a0dbbae8620d9984b68470367039636 c0decafe <220505+c0decafe@users.noreply.github.com> 1775170126 +0200 commit: Trim unused direct commands and add lfs task 68b30a6b7a0dbbae8620d9984b68470367039636 45620251c03d0ef72cab89357f3530b91cbb2068 c0decafe <220505+c0decafe@users.noreply.github.com> 1775241964 +0200 commit: Pin model inputs and add store-backed ingest workflow 45620251c03d0ef72cab89357f3530b91cbb2068 747e913a8c3494b4301a618cd923e94e5791e50d c0decafe <220505+c0decafe@users.noreply.github.com> 1775249290 +0200 commit: Add transcript and story bible derivations for step 2 747e913a8c3494b4301a618cd923e94e5791e50d df63d6672cc40f39825054df33e9931f9f843940 c0decafe <220505+c0decafe@users.noreply.github.com> 1775259056 +0200 commit: Archive session 50f08833: root cause analysis and continuation plan df63d6672cc40f39825054df33e9931f9f843940 a003cacc90da7b694a0adb960217235ebd0b1084 c0decafe <220505+c0decafe@users.noreply.github.com> 1775261330 +0200 commit: Make transcript review a GPU pipeline with ASR + Omni disambiguation a003cacc90da7b694a0adb960217235ebd0b1084 b994d3e5f111162e101e78ee47b7d93ab5c82b9a c0decafe <220505+c0decafe@users.noreply.github.com> 1775262136 +0200 commit: Improve transcript disambiguation quality and audit detail b994d3e5f111162e101e78ee47b7d93ab5c82b9a b5147d19f7dd73e3220c2d3bc6114257a5a58ff6 c0decafe <220505+c0decafe@users.noreply.github.com> 1775307432 +0200 commit: Fix Omni generation crash: use Thinker (B) with VL fallback (C) b5147d19f7dd73e3220c2d3bc6114257a5a58ff6 a18b4ec4441084f961fad020d68d3f144b5f1b7b c0decafe <220505+c0decafe@users.noreply.github.com> 1775308706 +0200 commit: Strip unsupported processor kwargs before generate() a18b4ec4441084f961fad020d68d3f144b5f1b7b ac7bb9bd00c762d04f827af15645d7ae819d1130 c0decafe <220505+c0decafe@users.noreply.github.com> 1775309208 +0200 commit: Detect model architecture from config.json for correct class selection ac7bb9bd00c762d04f827af15645d7ae819d1130 a6c3f0e65cad851227a6db307ecd47a1dc4dc7a2 c0decafe <220505+c0decafe@users.noreply.github.com> 1775311415 +0200 commit: Make JSON parser robust against malformed model output a6c3f0e65cad851227a6db307ecd47a1dc4dc7a2 10d3fbbc4152cb41ac73215ae93ff914eca7c8ff c0decafe <220505+c0decafe@users.noreply.github.com> 1775315317 +0200 commit: Maximize remote builder utilization and add progress monitoring 10d3fbbc4152cb41ac73215ae93ff914eca7c8ff f609e3e5e198ee93cfe0974fb13bc82ce68208b2 c0decafe <220505+c0decafe@users.noreply.github.com> 1775318748 +0200 commit: Don't force CPU steps to GPU server; build them locally f609e3e5e198ee93cfe0974fb13bc82ce68208b2 50f7f396e73b69d580cd8f5e5243409a815e0a3f c0decafe <220505+c0decafe@users.noreply.github.com> 1775318868 +0200 commit: Make flash_attention_2 best-effort, not required 50f7f396e73b69d580cd8f5e5243409a815e0a3f d7025704891df818847c8fa3de1968223d2b0184 c0decafe <220505+c0decafe@users.noreply.github.com> 1775323882 +0200 commit: Add ffmpeg to step2-reviewed nativeBuildInputs d7025704891df818847c8fa3de1968223d2b0184 cf651b3b4ba583388545edd26b794e2108fbdf6b c0decafe <220505+c0decafe@users.noreply.github.com> 1775328960 +0200 commit: Fix permission error in step2-reviewed: chmod after cp from store cf651b3b4ba583388545edd26b794e2108fbdf6b 8075281f3500238cea40be9b9fa360b5621dee46 c0decafe <220505+c0decafe@users.noreply.github.com> 1775329021 +0200 commit: Add dontUnpack to step3 packaging derivations (no src needed) 8075281f3500238cea40be9b9fa360b5621dee46 8f65a19345da8a1eb3969d0553a9cc5a376b29ea c0decafe <220505+c0decafe@users.noreply.github.com> 1775334125 +0200 commit: Fix wan21 packaging: mkdir work before cp, use /* glob 8f65a19345da8a1eb3969d0553a9cc5a376b29ea 0c77dd6e0f6e9be535ac7f1cc5cdea7c0e985540 c0decafe <220505+c0decafe@users.noreply.github.com> 1775339508 +0200 commit: Fix wan21 zip: don't cd work (persists to installPhase) 0c77dd6e0f6e9be535ac7f1cc5cdea7c0e985540 9d60ea387fc3b8404fc39e1baaee6871c4f82b84 c0decafe <220505+c0decafe@users.noreply.github.com> 1775339588 +0200 commit: Add progress report 2026-04-04 9d60ea387fc3b8404fc39e1baaee6871c4f82b84 b8b3fc505ff992d82dbe5c0376430e9e93eefb55 c0decafe <220505+c0decafe@users.noreply.github.com> 1775340475 +0200 commit: Split src filter per-derivation, add flash-attn, document dual-builder pattern b8b3fc505ff992d82dbe5c0376430e9e93eefb55 ad5a182d2b87bed64d15385871bc669e0c9b7bb1 c0decafe <220505+c0decafe@users.noreply.github.com> 1775340531 +0200 commit: Remove flash-attn from gpuPython (marked broken in nixpkgs) ad5a182d2b87bed64d15385871bc669e0c9b7bb1 87bb9a2b6da4e4ac58a1e0b3e4fa05f88be50879 c0decafe <220505+c0decafe@users.noreply.github.com> 1775362060 +0200 commit: Make model input ensure lazy: don't crash on GC'd paths at eval time 87bb9a2b6da4e4ac58a1e0b3e4fa05f88be50879 a10f2f6ac26a71ac13e59e8ab5c44a1edc3267b4 c0decafe <220505+c0decafe@users.noreply.github.com> 1775362233 +0200 commit: Skip GC'd model paths in pipeline-inputs-local.nix generation a10f2f6ac26a71ac13e59e8ab5c44a1edc3267b4 42d873801cc87213c90c30a7ce2352a0b0760ad7 c0decafe <220505+c0decafe@users.noreply.github.com> 1775428891 +0200 commit: Register model GC roots under /nix/var/nix/gcroots/ so Nix respects them 42d873801cc87213c90c30a7ce2352a0b0760ad7 07a2851885f697c79c636e19fec2f930ab0038a4 c0decafe <220505+c0decafe@users.noreply.github.com> 1776267367 +0000 commit: Fix direnv setup; add LFS coverage for reference MP4s; ignore whisper binaries 07a2851885f697c79c636e19fec2f930ab0038a4 1f957afd23ba9b5c694530cdd3b773ccadaeb299 c0decafe <220505+c0decafe@users.noreply.github.com> 1776267377 +0000 commit: Add fandom wiki research: database, scrape/query/clean tools, correction audit 1f957afd23ba9b5c694530cdd3b773ccadaeb299 387dc6577d55432d03cdbdf18e105a1fed8c53dc c0decafe <220505+c0decafe@users.noreply.github.com> 1776267387 +0000 commit: Revise proposal scope to image-first cycle at EUR 7.5K/month budget 387dc6577d55432d03cdbdf18e105a1fed8c53dc a47c4df963fe26b5e3afe6ad14e182ed38c5a061 c0decafe <220505+c0decafe@users.noreply.github.com> 1776267409 +0000 commit: Add S7 animation research: style analysis, benchmark packs, model selection a47c4df963fe26b5e3afe6ad14e182ed38c5a061 ec8d22dd095c79a549d588b6313866811e9241f9 c0decafe <220505+c0decafe@users.noreply.github.com> 1776267420 +0000 commit: Add vision analysis results for official Vimeo trailer ec8d22dd095c79a549d588b6313866811e9241f9 4ae70926967dec57fd2001570e190a997c2de2b3 c0decafe <220505+c0decafe@users.noreply.github.com> 1776267472 +0000 commit: Add youtube-s7-validation benchmark set: frames, packs, eval results, bible 4ae70926967dec57fd2001570e190a997c2de2b3 6d615444675aa5c051a51dcb001fa65533e79649 c0decafe <220505+c0decafe@users.noreply.github.com> 1776267667 +0000 commit: Add official S7 YouTube reference corpus (104 videos via LFS) 6d615444675aa5c051a51dcb001fa65533e79649 51875189aff9280f3df2e92581e87971dba35553 c0decafe <220505+c0decafe@users.noreply.github.com> 1776267794 +0000 commit: Session checkpoint b032b830: S7 research, fandom corrections, image-first proposal 51875189aff9280f3df2e92581e87971dba35553 6d2a78544b9e6c4eef80469eed17593bc6e98941 mnm 1776445304 +0000 commit: S7 bible: cross-check + normalization + wiki lockdown + consolidated docs 6d2a78544b9e6c4eef80469eed17593bc6e98941 5e21bf7573a4d5222947cef24ce41a035524e2fc mnm 1776445530 +0000 commit: Bible final normalization: locations→18 categories, outfits→5 types, gadgets mapped 5e21bf7573a4d5222947cef24ce41a035524e2fc 3a473aec91ea86e7ebb9a4fd9254f841ce255da9 mnm 1776449044 +0000 commit: Bible deep iteration: per-episode gadget mapping, location refinement, catsuit variants 3a473aec91ea86e7ebb9a4fd9254f841ce255da9 5046b552fd0ca391c8f13807fe940eef12e641c9 mnm 1776462058 +0000 commit: CRITICAL: wiki cross-check reveals character IDs are unreliable in bible 5046b552fd0ca391c8f13807fe940eef12e641c9 16b685cc37d4ea0b3e1eab2b8329683e07a96b24 mnm 1776509991 +0000 commit: Character re-identification: 2,835 shots re-IDed with wiki hair/suit anchors 16b685cc37d4ea0b3e1eab2b8329683e07a96b24 4d1cd0fb0d62a03fe55f50af35e7d94c8a85a557 mnm 1776722683 +0000 commit: Bible final refinement: cleaned IDs, outfit stats corrected, catsuit colors verified 4d1cd0fb0d62a03fe55f50af35e7d94c8a85a557 eda4468d6b7039bbe85a0be6aa83ef9f49e27a0b mnm 1776723298 +0000 commit: Add bible gap analysis doc: 12 gaps, priority ratings, fix estimates eda4468d6b7039bbe85a0be6aa83ef9f49e27a0b ab5a354b2c1777780017840ee1c0ae03f8bf27d8 mnm 1776726901 +0000 commit: Close 9 gaps: gadgets, locations, props, villains, scene types, S1-6 comparison ab5a354b2c1777780017840ee1c0ae03f8bf27d8 955010d9f3610224e9dfcea6f489dc0ce1dc9cec mnm 1776755517 +0000 commit: devenv: add spies-diarize task for Gap 4 (speaker diarization via pyannote-audio) 955010d9f3610224e9dfcea6f489dc0ce1dc9cec cb293489bd1c589fed2e3bf5fcec915769e1b677 mnm 1776755589 +0000 commit: Gap 2 complete: full outfit re-extraction for all 1639 trio shots cb293489bd1c589fed2e3bf5fcec915769e1b677 2092cef0cf4435f1017801260c16f6e5d87a1381 mnm 1776781005 +0000 commit: spies-diarize: auto-load HF token from pi-secrets, fallback to ~/.config 2092cef0cf4435f1017801260c16f6e5d87a1381 0f5e3fe517bd7c92c61d9fff6c38f96b2852c433 mnm 1776802417 +0000 commit: Gap 4 complete: speaker diarization across all 13 S7 episodes 0f5e3fe517bd7c92c61d9fff6c38f96b2852c433 3ff3ac38a26fa6b512c6c09cd7100637c8da13dc mnm 1776808840 +0000 commit: Gap 4: map SPEAKER_N labels to character names 3ff3ac38a26fa6b512c6c09cd7100637c8da13dc fa1625d0ba369a19e1461f77271514bbcc4af664 mnm 1776811356 +0000 commit: docs: comprehensive update after all gap passes complete fa1625d0ba369a19e1461f77271514bbcc4af664 de1f97f271c7160abbe20d60361c497190d4fb06 mnm 1776965558 +0000 commit: Diarization complete: all 13 episodes with PLDA de1f97f271c7160abbe20d60361c497190d4fb06 6d8cea6a8d99468c64863c8cbf446dd7f20edf13 mnm 1776969461 +0000 commit: Training readiness: 1,551 clips + manifest + structured captions 6d8cea6a8d99468c64863c8cbf446dd7f20edf13 75a99b19f7801bceceb573f8ac60dd3dd3d0d729 mnm 1776971669 +0000 commit: Add enriched VLM caption job (tools/vlm_caption_job.py) 75a99b19f7801bceceb573f8ac60dd3dd3d0d729 b4ded5577262d8d1f07a47efc6b502d9c92cff16 mnm 1777017835 +0000 commit: VLM captions complete: 1,551/1,551 (100%) b4ded5577262d8d1f07a47efc6b502d9c92cff16 ca66274467efb3775f55cbe444220e5c9dbabd62 mnm 1777018389 +0000 commit: Add meeting talking points doc ca66274467efb3775f55cbe444220e5c9dbabd62 a45bb2f0ec95615ffc9ed81713d264286bc48e3d mnm 1777018758 +0000 commit: Meeting doc: add image-only budget fallback section (FLUX.1-dev LoRA) a45bb2f0ec95615ffc9ed81713d264286bc48e3d 9d3bb9a9b6a2c7a1f92baf9c0904b71a555306f3 mnm 1777020095 +0000 commit: Add model-licensing.md + update meeting talking points 9d3bb9a9b6a2c7a1f92baf9c0904b71a555306f3 099a9991406f6f287ccb0098f47ab598e3cc05ec mnm 1777021327 +0000 commit: Add meeting executive summary (meeting-executive-summary.md) 099a9991406f6f287ccb0098f47ab598e3cc05ec 96bee82f24c667c500874ecd7263be12e77948f8 mnm 1777099248 +0000 commit: Add conservative VLM re-pass script (tools/vlm_repass_conservative.py) 96bee82f24c667c500874ecd7263be12e77948f8 38c812ba1d10ffa13d00d4ae629bc4efc8fcbb84 mnm 1777123808 +0000 commit: Re-pass complete: conservative VLM captions, 1,551/1,551 38c812ba1d10ffa13d00d4ae629bc4efc8fcbb84 70cdd08d7e7fd5a54fe3b36581146280d1e49cac mnm 1777150723 +0000 commit: Expand gadget + villain DBs, fix matching, apply to manifest 70cdd08d7e7fd5a54fe3b36581146280d1e49cac e35e46dbbc276856aaa7c4ed59924011de640eec mnm 1777151725 +0000 commit: Isolate gadget false positives: context-token stop + distinctive requirement e35e46dbbc276856aaa7c4ed59924011de640eec 3f314dacc58266c900d2862d178cf68d1b9044c1 mnm 1777206475 +0000 commit: Tier 1 improvements: character names, merged captions, location resolution 3f314dacc58266c900d2862d178cf68d1b9044c1 f921caabccef9d77fe41324cd01d94329724fd04 mnm 1777208814 +0000 commit: Location resolution complete: 0 Unknown (was 454/29%) f921caabccef9d77fe41324cd01d94329724fd04 14c322ad24fd8bed981f7c25fd05721456ce9d3b mnm 1777214637 +0000 commit: Tier 2 improvements: outfit data, villain IDs, caption consistency 14c322ad24fd8bed981f7c25fd05721456ce9d3b d62954c32d17bbdfd91440c39e5557262d7ff27a mnm 1777231478 +0000 commit: Continued improvements: hair-only char inference, villain keywords, scene types d62954c32d17bbdfd91440c39e5557262d7ff27a a6dcd64e131efefff4b2bbdade3f8867e3f5a371 mnm 1777233873 +0000 commit: Final dataset improvements: attribution, dataset rebuild, acceptance a6dcd64e131efefff4b2bbdade3f8867e3f5a371 7faebc7397d501676d692c5f8a3b8d9f67c461d1 mnm 1777285590 +0000 commit: Update pipeline to Wan2.2, add dataset-state doc, rebuild dataset files 7faebc7397d501676d692c5f8a3b8d9f67c461d1 55fccac0c5a22e38782a28c9e689573d0e09a464 mnm 1777313363 +0000 commit: Docs: update s7-current-understanding, add dataset-state + GPU runbook 55fccac0c5a22e38782a28c9e689573d0e09a464 6d3094e54332ad6f7c832f7f4ba5249223e5679b mnm 1777320940 +0000 commit: Force-alignment: word-level timestamps via faster-whisper 6d3094e54332ad6f7c832f7f4ba5249223e5679b 2577dad147f2a69dd3f6a1cad6ff5f094440ac59 mnm 1777323019 +0000 commit: Final cleanup: scene types, integrity check, model ingest, pipeline validation 2577dad147f2a69dd3f6a1cad6ff5f094440ac59 bf63ca0f5497bfb11855874e3ff5743893fb9750 mnm 1777330572 +0000 commit: Complete documentation pass bf63ca0f5497bfb11855874e3ff5743893fb9750 d95e4782d969dab06838b217351bd9e26c69730a mnm 1777384079 +0000 commit: Fix training blockers: DiffSynth-Studio format + correct training script d95e4782d969dab06838b217351bd9e26c69730a da635918bb71b1c6b64314f88e82901491fad6b1 mnm 1777395205 +0000 commit: Rename wan21 → wan22 throughout pipeline da635918bb71b1c6b64314f88e82901491fad6b1 da635918bb71b1c6b64314f88e82901491fad6b1 mnm 1777805150 +0000 reset: moving to HEAD da635918bb71b1c6b64314f88e82901491fad6b1 c925c454cb58b897b5411c487baf7c6e35dcdbb5 mnm 1777805150 +0000 commit: Add PLDA flag to benchmark episode diarizations c925c454cb58b897b5411c487baf7c6e35dcdbb5 da635918bb71b1c6b64314f88e82901491fad6b1 mnm 1777805174 +0000 reset: moving to HEAD~1 da635918bb71b1c6b64314f88e82901491fad6b1 da635918bb71b1c6b64314f88e82901491fad6b1 mnm 1777805178 +0000 reset: moving to HEAD da635918bb71b1c6b64314f88e82901491fad6b1 e6c4a046050ba5332b331550dae53c7db158850b mnm 1777805178 +0000 commit: Add PLDA flag to benchmark episode diarizations e6c4a046050ba5332b331550dae53c7db158850b f221a54be1757573afc3817df3d6c68222b06506 mnm 1777805187 +0000 commit: Add wiki ground truth pages and training dataset