comma-lab

comma-lab is a public research and engineering workspace for the comma.ai video compression challenge. The README describes it as a community and historical-record repository, not a live leaderboard page, and it does not make an arXiv or preprint commitment. Its main reusable package is tac, a Python Task-Aware Compression library and algorithmic engine. The comma_lab package is the lab operations layer for custody, state projection, and reporting, not a second compression engine. The repo covers public-frontier intake, archive byte anatomy, exact replay custody, writeup drafts, and OSS tooling. The README defines terminology boundaries: TAC means task-aware compression, a codec is a narrower encoder/decoder or wire format, and the historical pact alias appears only in internal paths and docs. It itemizes inherited substrate from contest PRs #130, #135, and #133 and says learned semantic renderer and pose carrier artifacts are byte-identical after decode, with a separate SHA-256 receipt document for what is borrowed versus original. Score-bearing claims are read through evidence grades: A++/A for ranked rows, A-negative for scoped negative results, empirical for byte/smoke/loss evidence, derivation/prediction for roadmap hypotheses, external for community context, and invalid for non-ranking evidence. Workflows include public-frontier intake and exact replay custody; tac packet compilers, entropy coders, scorer-aware losses, renderers, quantizers, sensitivity maps, master-gradient consumers, and procedural byte derivation; comma_lab state projection, strict preflight adapters, public-intake hygiene, and release hygiene; separated contest-CPU and contest-CUDA auth eval; and release, paper, and OSS hygiene for tac. The README documents a CUDA versus CPU auth eval split finding from 2026-05-08, stating that the upstream scorer produces two authoritative score axes, the public leaderboard ranks by the CPU score, and every shippable archive gets dual-eval on contest-compliant Linux x86_64 hardware. The package map lists src/tac, src/comma_lab, tools/scripts/experiments, docs, .omx/research, reverse_engineering, submissions, and a pinned upstream snapshot. A related standalone package adpena/tac is described as a curated production extract with an import-compatible library surface. Quick start steps install with uv, run tac CLI commands, and run a terminology guard. Historical training profiles WILDE, SHIRAZ, and GREEN are retained for reproducibility, though newer lanes may use other approaches. The repo also preserves historical context such as H.265 and CNN post-filter baselines, asymmetric warp renderer and pose TTO work, a gradient obstruction fix, public PR replay/deconstruction, and post-deadline lanes. Methodology treats compression as a small compiler for contest archives and keeps research state in durable files. The paper in docs/paper covers the asymmetric warp architecture, gradient obstruction bug discovery, Fridrich-informed loss design, and rank-1 radial zoom warp derivation. Requirements list Python 3.11+, PyTorch 2.0+, ffmpeg, and recommended CUDA GPU with MPS/CPU development support. License is MIT.