Experimental optimization measurements
For the released product's 22.61% / 28.48% less elapsed time comparison against the prior Stock workflow preset, start with released workflow results. The research candidate below is a separate experiment and is not the basis of those claims.
Public 0.6.1 is a source/portability release. The following are historical Phase 11 measurements of an experimental PGO engine, not a rerun of the new public packaging. The candidate was not promoted.
Exact Phase 11 comparisons
Negative change means less elapsed time. Seconds are accepted-case medians; paired time change is the geometric mean of whole-block ratios, not their median ratio. The release requires nine valid whole blocks in every workload/load group and an upper paired 95% ratio below 1 against both references, plus exact quality.
| Pipeline / workload | Reference | Valid / required | Candidate s | Reference s | Paired elapsed change | Ratio 95% interval | Group eligible |
|---|---|---|---|---|---|---|---|
| direct / hd60/load0 | original | 1 / 9 | — | — | Insufficient | — | No |
| direct / hd60/load0 | stock | 1 / 9 | — | — | Insufficient | — | No |
| direct / hd60/load2 | original | 1 / 9 | — | — | Insufficient | — | No |
| direct / hd60/load2 | stock | 1 / 9 | — | — | Insufficient | — | No |
| direct / raw30/load0 | original | 3 / 9 | 2.3782 | 2.4107 | -1.654% | 0.980260–0.986062 | No |
| direct / raw30/load0 | stock | 3 / 9 | 2.3782 | 2.4121 | -1.801% | 0.978012–0.984411 | No |
| direct / raw30/load2 | original | 7 / 9 | 2.3285 | 2.3743 | -2.033% | 0.977730–0.981932 | No |
| direct / raw30/load2 | stock | 7 / 9 | 2.3285 | 2.3755 | -1.994% | 0.977327–0.982943 | No |
| workflow / hd60/load0 | original | 9 / 9 | 5.6018 | 5.6954 | -1.764% | 0.979057–0.985468 | Yes, bounded |
| workflow / hd60/load0 | stock | 9 / 9 | 5.6018 | 5.7016 | -1.743% | 0.979896–0.985302 | Yes, bounded |
| workflow / hd60/load2 | original | 9 / 9 | 5.6307 | 5.7251 | -1.532% | 0.982220–0.987248 | Yes, bounded |
| workflow / hd60/load2 | stock | 9 / 9 | 5.6307 | 5.7376 | -1.791% | 0.979719–0.984871 | Yes, bounded |
Complete HD workflow passed its two bounded groups. Direct HD accepted only 1/9 each, raw 3/9 and 7/9. Overall performance/promotion gate failed. Original native 0.6.0 was retained. A small raw interval below 1 does not override the preregistered sample requirement.
Workload and controls
- Apple M4 Max, 10P + 4E, 36 GiB, macOS 27.2.
- Big Buck Bunny 1920×1080 at 30fps; HD 60–120s, raw YUV420p 60–90s (2,799,360,000 bytes).
- Matched libx264 veryfast/CRF20, encoder/decoder/filter 8/2/2, workflow validation 8, HD AAC160k.
- load0 has no added owned CPU helper; load2 has two 50%-duty helpers. Other host activity is observed/gated, not absent.
- Warm cache and temporary input/runtime residency checks; parent-side child-lifetime timers and actual counters. No fixed CPU affinity, frequency or thermal state.
- Direct: 688 completed cases including 12 excluded warmups, 180 attempted blocks. Workflow: 144 including 6 excluded warmups, 23 attempted blocks.
- Rejected blocks remain in the evidence; they are excluded from accepted timing. No pooling of warmups, interrupted runs or different epochs.
- The original lab used an explicit task-scoped system-memory waiver, nice≥10, 12 GiB owned-job cap, disk/mount/output/deadline guards.
Quality and compression
| Source | VMAF mean | PSNR-Y dB | SSIM | Metric delta vs both references |
|---|---|---|---|---|
| hd60 | 94.781527 | 47.011742 | 0.998632 | 0 |
| raw30 | 94.920178 | 46.621946 | 0.998291 | 0 |
Strict full-frame decoded video identity and HD audio identity passed on the target outputs. Fresh native media/CLI/MCP checks (14) and six codec-mode checks passed. Scheduling-dependent codec modes had strict decode/assertion acceptance rather than an exact-byte equivalence claim. Private container files differ by 14 metadata bytes; this is not a compression-ratio improvement.
CPU work and remaining acceptance
Descriptive instruction counts decreased approximately 4.46–4.85% versus original native and 4.78–5.25% versus Stock across completed cases. Those counts include invalid-environment blocks and are not elapsed-time acceptance. Phase 10 reduced progress-query lock traffic about 98.4%; its wall-clock gate still failed.
Remaining gaps: enough stable whole blocks in every direct workload/load, independent wider-source/hardware confirmation, and fresh new-prefix FATE/native/extended/GPU/fault/soak/consumer/install-update-rollback acceptance. Old receipts cannot accept a new engine.
Earlier epochs and reproduction
Phase 8 reduced complete-workflow time 22.31% (BBB) and 30.77% (ToS) against the old native workflow with different concurrency. Its same-tuning Stock comparison still included HD slowdowns; see the full report. Phase 9 lookahead calibration failed independent confirmation (39.94% slower), and Phase 10 instruction improvements did not pass elapsed-time confidence. These are distinct experiments, not one pooled speedup.
Use all reports, Phase 11 method, full Phase 11 report, summary JSON, direct analysis and workflow analysis. Analyses retain exact block ratios and original source-report hashes; private paths are redacted. Raw corpus/profiles/full process receipts are not bundled. The scripts are auditable research recipes, not a turnkey reproduction without those inputs.
For a new local bounded test, first review ffv2 benchmark-plan INPUT --output-dir DIR; that command is separate from the preregistered historical Phase 11 experiment. Publish full quality, failures, environment and uncertainty with any new result.