FFMPEG-V2 / reportsView source ↗

Experiment and report archive

All tracked historical reports are retained, including regressions, rejected candidates and remaining gaps. Dates and baselines belong to their own measurement epochs. They are not pooled into a universal speedup claim. Original reports are mostly Bengali; the English benchmark guide explains the latest outcome.

Phase Focus Outcome and documents
Initial / 2 Typed workflows, native progress, Metal correctness Phase 2 results, design, initial validation
3 Reused supervision/capability work Results, design
4 Larger corpus, policies, GPU paths Results, research
5–7 Recovery, long jobs, consumer/lifecycle checks Acceptance, design
8 M4 Max scoped thread-pool tuning Results, research
9 Same-tuning HD/raw gaps and failed lookahead candidate Results, research
10 Progress synchronization and per-worker/load/cache profiling Results, research
11 Conditioned cache/load replay and experimental PGO Results, research, English comparisons

Machine-readable evidence

Phase 11 summary, direct paired analysis and complete-workflow paired analysis include exact accepted block counts, times, estimates and intervals. The verification index directory retains redacted summaries for every development epoch.

Public index JSON files are redacted historical summaries, not valid local release or activation manifests. Their publication field records the original private file's SHA-256. Rewriting private paths changes public JSON bytes; original local evidence remains retained privately. Raw media, PGO profiles, runtime queues and large native artifacts are not bundled. This source publication does not claim an independent public replication of the full local lab.

Readiness and source

Known public limitations · historical gaps · readiness · native source · x264 patch source · upstream pin.

The public website renders the report archive with a searchable list. Every report also remains directly readable in GitHub, without requiring the website or JavaScript.