Same media.
Measured on one Mac.

Latest validated SDK · 4 November 2026

WebM playback, with less overhead.

Three alternating fresh-process runs per player on the exact same VP9 WebM bytes. Mac mini · Apple M6 · 12 CPU cores · 24 GB unified memory · macOS 27.0.1 · Release builds.

The latest BOSS route used 81.7% less process CPU time and 43.9% less peak RSS than the previous BOSS route on this fixture. Other Player still used less CPU time and memory. All nine runs produced playback witnesses.

Local and HTTP acceptance passed 19 cases and 43 checks, plus eight focused unit tests. Coverage includes VP9/VP8, Opus/Vorbis, six-channel audio submission, pause/resume, seeking, EOF, sequential HTTP, cancellation, authorization revocation, changed-media rejection and real timeout reporting. Unsupported video keeps the compatibility route.

This is one five-second fixture with three seconds of observation per run and a warm filesystem cache. Clock startup includes reaching 0.5 seconds of playback; it is not first-visible-frame latency. CPU time is not energy use. Peak RSS is process memory, not total system memory. No compiler or archive was running when the paired benchmark started. Audio-frame submission is not a receiver or listening certification.

Reproducible comparison

112 playback-clock checks.
100 seek-clock landings.

Mac mini · Apple M6 · 12 CPU cores · 24 GB unified memory · arm64 · macOS 27.0.1 (26A434) · Swift 6.4 · Release builds · AC power.

The same generated byte files were used by both engines, with the upstream recipes and encoding settings preserved. The complete media SHA-256 manifest includes HLS segments. Unpublished author encodes cannot be declared byte-identical without their originals and hashes.

Original baseline: local formats · median of three repetitions

Values measure load invocation to a playback clock beyond 0.5 seconds, with a 50 ms polling interval. They are not first visible frame latency. Smaller figures are not always better if the required output is missing. The first matrix measures clock progress and seek landings; it does not contain per-run visual-output checks.

Original baseline: network, AVI and paced live · median of two repetitions

HTTP files support ranges. Live TS is paced at the upstream test's byte rate and preroll, delivered through HTTP rather than Other Player's custom test reader. This is a common public-API comparison, not a reproduction of every upstream internal test assertion.

Each valid timing cell also shows peak process RSS and process CPU time for the complete observed session. BOSS used more memory on most of these small local fixtures; its original VP9 route measured about 150 MiB versus Other Player's 59 MiB. The newer measured VP9 route is reported at the top of this page. There is no universal performance winner.

A separate BOSS AVI witness reports 178 displayed pictures and an active 160×90 VLC video output. The ordinary AVFoundation-layer detector does not see VLC's renderer.

Exact HDR10+ fixture

Other Player's original two-frame, 0.2-second HDR10+ fixture was kept unchanged. Its official metadata probe identifies ST 2094-40. BOSS's owned route submits both frames but rejects a target-luminance field of 5,000,000/1, above its accepted range. A separate control changing only that field to 500/1 attaches a 64-byte HDR10+ payload and renders both frames. The original matched fixture is unchanged. This explains the detection/validation difference without weakening validation; it is not a certified display-consumption test.

What could not be matched

Other Player's real FEL tests require fel-clip-a.mkv, fel-clip-b.mkv and fel-clip-c.mp4, which are not included in the repository. We did not substitute other films. Published A12/M2 FEL figures are not comparable to this Mac run. The measured Other Player default build has its optional DualLayerDolbyVision and TrueHD traits disabled.

The comparison uses one pinned upstream source revision and its fixture recipes. Public result files retain exact shared-media hashes and the measurement method. The comparator is labelled Other Player.

Corrected BOSS build · reported separately

16 playback controls passed.
Video, seeking and retirement.

Finite, continuous, unencrypted MPEG-TS HLS media playlists with range-capable segments. No whole-media download or persistent movie/library file. Encoded video and metadata are unchanged. Adaptive/master, encrypted, byte-range, discontinuity, partial/low-latency and live graphs retain existing routing and are outside this correction.

Plain and +1000 s timestamp-offset HEVC HLS now render video and audio. Submitted video timestamps follow seeks to 12, 90 and 4 seconds. H.264 HLS stays native, including a server that ignores range requests. MP4, MKV and AVI controls also pass.

All 38 seek clocks landed. Header-scoped playback had zero header-gate failures; revocation stopped playback with the expected -20002 error. Joined teardown completed within 0.11 seconds. Three focused parser, PMT-CRC and virtual range-server regressions passed.

These are BOSS acceptance runs after the paired baseline, not a new paired speed ranking. An initial per-segment size probe adds setup requests on long playlists. No new consumer app build was installed or uploaded by this task.

Earlier broad format matrix · retained for comparison

Earlier BOSS versus Other Player format results.

In the 13 format cases with valid video in both paired runs, median load-to-clock time was 0.732 s for BOSS versus 0.746 s for Other Player. Median seek-clock latency was 0.053 s versus 0.078 s. BOSS used more process memory and CPU: 88.9 versus 68.8 MiB median lifetime peak RSS, and 0.182 versus 0.153 CPU seconds per observed process session. These short-window results are not an energy, photon-latency or universal speed claim.

The 84-run rendering audit recorded 39/42 BOSS and 42/42 Other Player witnesses. The three missing BOSS witnesses were the same in-band AVC MKV. Its native remux declared a playable video track but no out-of-band SPS/PPS. The correction routes only that unqualified prepared output to owned decode; all three corrected repeats render, and 11/11 final playback controls have actual video witnesses. Four focused regression tests pass. Failed original runs remain visible in the raw audit.

Each value is the median of three runs. The AVC MKV row uses BOSS's separately labelled post-fix repeats against Other Player's prior audit; it is not a newly paired run. The other rows are from the same 84-run audit. Numeric layer counts are used only as presence witnesses; VLC uses displayed-picture counters.

The original paced live TS adapter reached the 0.5 s clock at 4.02 / 4.15 s for BOSS versus 12.02 / 16.02 s for Other Player (5 s / 10 s GOP). This uses the same media and pacing recipe through our shared HTTP adapter; it does not execute every original custom-reader regression assertion. The exact 20-minute HEVC HLS fixture now renders in both players: BOSS 1.009 s, Other Player 0.743 s. BOSS submitted frames near the 12, 90 and 4 s seek targets.

Final authorization controls: 116 origin requests, zero header-gate failures, and revocation stops with failed/-20002. Native qualified H264/HEVC, MKV and AVI controls still render. The SDK patch and evidence have been handed to Odyssey TestFlight for integration and its app-update workflow; this BOSS benchmark thread has not archived, installed or uploaded Odyssey.

All recorded metrics and limits ↓ · 84-run rendering audit ↓ · Final AVC/control results ↓ · 20-minute HLS results ↓

Latest reviewed SDK sources

What changed in BOSS.

The standalone development SDK has been aligned to the reviewed 4 November Odyssey source cohort. Its original SDK is preserved. The complete native macOS SDK compiles with the matching Mac disc slice. The latest source also includes the validated WebM routing improvement above. Standalone BOSS Player 1.0 (4) has completed Apple processing and is marked Testing for the BOSS Internal group on iOS/iPadOS, tvOS and macOS. Physical-device playback acceptance remains separate from this release.

Structured website and SDK update plan ↓

Next in priority order

What remains.

Scope and limits

• No cold-cache, statistical significance, energy, thermal or system-wide memory claim.
• CPU seconds cover the complete observed process session; peak RSS is lifetime process memory, not all OS services.
• AV presentation-layer counts may count the same layer through overlapping view/layer traversal; only presence is used, never the numeric count.
• All 112 clocks advanced and all 100 seek clocks landed, but this does not mean 112 video passes: four HEVC HLS runs have confirmed missing video.
• The initial 84-run matrix has no per-run presentation-layer witness. AVI was checked separately through VLC displayed-picture statistics.
• HDR10+ exact sample is 2 frames / 0.2 seconds. It is not lengthened to satisfy the 0.5-second clock metric.
• No iPhone, iPad, Apple TV, HDMI analyzer or receiver qualification was performed. No active Odyssey app was modified or installed.

Existing capability labels describe their stated working scope. This comparison does not extend them to all media, devices, Dolby profiles or receiver configurations.