Skip to content

SonKu P7.5: quarantine real-music cancellation, measured centered MID + protected stereo SIDE - #689

Closed
masarray wants to merge 10 commits into
integration/sonku-p74-windows-field-20261114from
fix/sonku-p75-phase-safe-music-20261114
Closed

masarray wants to merge 10 commits into
integration/sonku-p74-windows-field-20261114from
fix/sonku-p75-phase-safe-music-20261114

Conversation

@masarray

Copy link
Copy Markdown
Owner

Incident

Physical listening owner reports P7.4 audio destroyed / severe cancellation, despite green synthetic CTests. Treat P7.4 sound as FAILED. Do not promote or claim dearVR parity.

Research-only fix

  • Opt-in ONLY ARMIX_SONKU_EXPERIMENTAL_MUSIC_SAFE=ON, requires native XYZ experiment; public/default and P7.4 unchanged.
  • Measured KEMAR single centered MID HRTF rather than combining two offset filters with coherent material. Original musical SIDE is preserved unconditionally as an orthogonal L/R axis with bounded width, without adaptive polarity-gated rescue, never creating anti-phase silence.
  • Early and late ROOM process MIDI-independent mono MID only to avoid collapsing phase-inverted stereo reflections. Same native callback, same SPSC/ACK/state; direct FIR does half the multiply-adds. No WebAudio DSP/threads/allocations.
  • Dedicated original signal regression: anti-phase, coherent bass with inverted high notes, correlated mono spread invariance, moving XYZ/ROOM partition parity 44.1/48k. Full experimental Windows JUCE build + portable tester workflow with exact source/notice.
  • IMPORTANT tradeoff: SIDE remains head-locked, and hard-panned mono vs stereo source behaviour plus actual timbre and externalization are NOT qualified. Experimental only; not product-grade dearVR parity.

Release veto

Wait for exact CI (source core + real Windows JUCE + native ACK v3 + portable), then headphone source/processed captures, physical ASIO and owner signoff. Do not merge to main, do not replace old artifact without evidence. Detailed incident/FAT and Issue #683.

@cloudflare-workers-and-pages

cloudflare-workers-and-pages Bot commented Nov 14, 2026 •

Copy link
Copy Markdown

Deploying arsonmix with  Cloudflare Pages  Cloudflare Pages

Latest commit: 170cfec
Status: ✅  Deploy successful!
Preview URL: https://0b58314f.arsonmix.pages.dev
Branch Preview URL: https://feat-sonku-p80-transparent-b.arsonmix.pages.dev

View logs

Copy link
Copy Markdown
Owner Author

P7.5 exact-SHA experimental Windows audio candidate — CI PASS, headphone sound NOT YET PASSED

Owner’s P7.4 physical field verdict remains FAILED: severe audible cancellation. Do not endorse the old artifact as good audio.

Exact SHA 089520f43c7e9d4900d70ee9ea5b10f83d167226 · Actions PASS · Windows portable research artifact (ArSonMix-SonKu-P75-Music-Safe-Windows-x64-43591146247, artifact ID 14112507871, 15,056,166 bytes; retention through 2026-12-05).

Verified

  • Native L/R → orthogonal MID/SIDE P7.5 source regression, 44.1/48 kHz: anti-phase at SPREAD 0/60/180, coherent MID spread invariance, bass+inverted high stereo content, and moving 3D room 64/257 chunk parity. Fast preflight PASS.
  • Real Windows x64 ARMixConsole JUCE 8.0.15 built with both explicit research flags ARMIX_SONKU_EXPERIMENTAL_NATIVE_XYZ_RT=ON, ARMIX_SONKU_EXPERIMENTAL_MUSIC_SAFE=ON. Runtime plugin and focused P7.5 test share the compile profile, not a separate simulated engine.
  • Actual Windows MSVC Release CTest 4/4 PASS, including JUCE v3 ACK and v1/v2 compatibility.
  • Approved embedded GUI image/HTML/CSS/JS byte hash + PE icon/source SHA + package provenance + archive integrity PASS; mandatory MIT KEMAR notice and field FAT guide hash-verified in ZIP.
  • New GUI labels protected research MID/SIDE mode only after exact ChannelId/instance/deviceEpoch/graphRevision-accepted snapshot; public GUI stays unchanged.

Still an UNSIGNED audio-quality gate

No validated physical audio improvement yet. New mode preserves stereo side but head-locking/tonal issues, intermediate MIX coloration, mono hard-left sources, multi-insert ASIO XRUNs, actual Save/Open .smix, and independent headphone dearVR A/B remain open. Source-only numeric tests cannot prove that the owner hears natural audio. Do not merge, publish as a release, or claim dearVR parity. PR remains DRAFT.

Physical test and failure reproduction guide.

Copy link
Copy Markdown
Owner Author

New real-listening report: P7.5 still FAILS tonal and room fidelity (14 Nov 2026)

After the original severe P7.4 cancellation and P7.5 music-safe source tests, the owner now reports P7.5 still has prominent/nasal mids, attenuated bass, unnatural consolidated stereo, and a flat/in-head room versus dearVR PRO 2. This supersedes any impression that P7.5 numeric anti-phase PASS means the audio quality problem is solved. The old P7.5 ZIP remains a CI-verified research build, not an owner-approved acoustic result.

New engineering RFC: SonKu P8 — official dearVR benchmark and measured acoustic architecture based on Dear Reality official manual and Sennheiser Clarity technical explanation, then verified against exact P7.5 code. Key issues:

  • Forced stereo MID→single centered HRIR plus direct head-locked SIDE is a safety prototype, not a believable two-object stereo 3D renderer.
  • No measured full-band timbre correction/reference, scene-dependent low-frequency shelf, ER with simple scalar surface absorption, and synthetic low-density eight-line diffuse return leave audible problems.
  • dearVR's BASS BOOST is separate from ROOM, so record Bass Boost state instead of assuming the room alone provides the extra low frequency. Its CLARITY control trades HRTF spectral coloration against localization; do not copy patented implementation.
  • P8 gates: (0) capture dry/direct/ER/late measured transfer + third-octave and stereo phase/band retention; (1) coherent stereo positional matrix + measured spectral correction; (2) calibrated six-surface acoustics + bandwise RT60/direct/reverb ratio; (3) exact Windows/ASIO/headphone/dearVR A/B.
  • Do not patch with guessed bass gain, reverb multipliers, fixed wet boosts or another mono fold. Do not claim dearVR parity from CI, do not merge/release P7.5 as an audio success.

Public defaults unchanged. Need captured dry/wet or controlled matched listening to validate the cause and perceptual improvement. Keep #683 open and #689 draft.

Copy link
Copy Markdown
Owner Author

MILESTONE A: closed WITHOUT MERGE — historical audio research. P7.5 audio-safety investigation was an experimental response to a physically rejected P7.4 candidate; its earlier voice-guard scope is not authority to replace accepted P9.9. Production #733 selectively ported the accepted latest sound, not the full chain. Preserved original source: fix/sonku-p75-phase-safe-music-20261114 @ 170cfecc26279f95ee27f9b8f58a18cee10199db; original target integration/sonku-p74-windows-field-20261114. All 47 PR heads are recorded in docs/governance/MILESTONE_A_PR_BRANCH_RECONCILIATION_20261117.md (chore/milestone-a-pr-history-20261117). No source branch or checkpoint deletion, no DSP changes, no retroactive claim of listening acceptance. New work may reopen via scoped approval.

@masarray masarray closed this Nov 17, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant