Skip to content

DualSense rear-channel underrun — 2026-09-21

Historical record. This preserves the source, tests and limitations recorded at the time; it is not current setup guidance. See the documentation index.

Evidence and limits

External USB/IP source PCM and physical Bluetooth HID captures were compared against a frozen, independent public WDL sinc64 oracle. The comparison fixes one global block offset, preserves authored zero samples, and does not remove quiet carriers to obtain a passing result. Host USB submission/completion is not proof of over-air receipt or actuator behavior.

The same pre-gate candidate produced different outcomes across controlled runs:

  • Capture candidate-clock-02 inserted one all-zero carrier between source blocks 203 and 204. Every nonzero block still appeared, but the strict first-anchor comparison failed: 31,129 scalar mismatches. Matching only nonzero hashes would have hidden this sequence error. The source-data timestamp preceded the quiet carrier, but the corresponding ISO completion followed it; neither timestamp identifies actual helper availability.
  • Capture candidate-clock-03 passed all 66,176 scalars in 1,034 complete source blocks at one fixed offset. Its full accounting is 249 leading zero carriers, 1,034 source blocks, and 1,063 trailing zero carriers. The remaining 30 oracle stereo frames are zero and do not form an emitted complete block. All source- window reports have successful matched host USB completions; only the final idle-tail carrier's completion is outside the capture.

Both runs used the scheduler-corrected broker, SHA-256 44EEBCDA8672E32EA9CF5B386F3E5AFFA1044194074A83F32D239762644C4786, but neither used the new rear underrun gate. Capture 03's pass therefore does not validate this change or erase capture 02's failure. Its captured source-data complete → host USB submission median was 29.340 ms; this is not a latency-parity claim.

A concurrent read-only ring observer collected 1,097 valid retained slots with no rejected or overwritten slots. Its actual polling was approximately 15.5 ms, not the requested 2 ms sleep. The producer-supplied enqueue timestamp marks the start of publication, potentially before capacity waiting; only observed write- sequence changes bound committed publication. Sampling cannot establish the helper's instantaneous empty decision, and invalid observations must not be interpolated.

Bounded correction

Only steady V5/native speaker presentation is eligible. Before claiming the front packet, the presenter probes for an owned current-generation rear block. If the ring is empty after a recently committed source block, it may defer once for at most one 32-frame / 3 kHz block interval (approximately 10.667 ms), also bounded by the prior publication-attempt timestamp's approximately 32 ms recent window. Notifications cannot renew that absolute deadline. Real end-of-source therefore resumes ordinary silence after one bounded wait, without replaying the last waveform. Authored zero blocks remain ordinary owned PCM.

The tradeoff is a possible bounded delay to the unclaimed front-audio carrier; there is no extra startup buffer, gain change, resampling change, packet removal, or fabricated waveform. Compact/paired transports and startup priming are not gated. Eligible native commands and microphone transitions keep their existing priority; a Busy native credit is rechecked at the existing 1 ms retry scale during deferral. No physical I/O runs under the state lock.

Ring protocol version 2 adds a separate data-available event, not reuse of the producer's space notification. Preparation does not release a slot: only an accepted media write commits it. Lifecycle changes reset both prepared ownership and the gate. Existing ordered helper Stop semantics remain unchanged. The null-by-default internal test hook has no CLI, IPC, profile, or settings route.

Software validation and remaining gate

The actual-helper regression first failed with one extra rear interval between authored A and B while front audio and the native trigger still progressed. The same test then passed. The initial expanded six-test run also passed, covering authored silence, finite end-of-source waiting, Busy physical ownership, generation clear, microphone control, and front/trigger order.

Receipts under isolated_results/haptics-ab-20260920/ are:

  • rear-underrun-red-01/rear-underrun-red.trx: 1 failed, preserved.
  • rear-underrun-green-01/rear-underrun-first-green.trx: 1 passed.
  • rear-underrun-expanded-01/rear-underrun-expanded.trx: 6 passed.
  • candidate-clock-wdl-timing-02/ and candidate-clock-wdl-timing-03/: immutable external-capture sequence/timing audits and input hashes.
  • realtime-ring-observation-01/: bounded read-only observation and quality metadata, separate from the external capture clock.

The first combined 36-test run found a further real failure: a Busy native write granted itself one front-media fairness debt while that same media was held by the rear gate. The corrected guard grants new debt only when media can actually progress; it preserves existing microphone/native debt. The same 36 tests then passed unchanged. Both receipts are retained, not replaced.

Final receipts under DS4WindowsTests/TestResults/ cover ordered Stop preservation, native Busy→ready wake and fairness, prepared-age accounting, and the null internal test seam:

  • sony-timing-gate-focused.trx: 35 passed, 1 failed, preserved.
  • sony-timing-gate-focused-02.trx: 36 passed, 0 failed, including zero-allocation checks.
  • sony-timing-expanded-final.trx: 1,780 passed, 0 failed, 3 opt-in skips.
  • sony-timing-full-final.trx: 6,766 passed, 0 failed, 12 opt-in skips; SHA-256 831C4EAB536725E525C273C754E6CD8B7860DA2D243088A6A8B9128DBC869BC8.

Independent review found no remaining source blocker in the bounded gate. The new gate has not yet been validated on hardware. Fresh paired capture must verify the complete fixed sequence, all zero carriers, front/control behavior, and timing; no release or user-visible improvement is claimed here.

Portable deployment checkpoint

Build07 was independently verified as a complete 554-file portable package. Its mapper DLL SHA-256 is 1A2F8462F653F62CFA4CCA78CEDB9EC24BE3F7EB38C04F3A85B3210C07D8497A. The lab directory explicitly substitutes the already-tested clock/WDL broker 44EEBCDA8672E32EA9CF5B386F3E5AFFA1044194074A83F32D239762644C4786 and its license notice; the unchanged ZIP retains the public broker and is not being offered as the private timing candidate.

At 05:36:10 UTC the reviewed same-path Desktop replacement completed. The full old tree is retained separately. All private data matched byte-for-byte before launch, authentication and HidHide were unchanged, and the new mapper logged backend-ready and controller-search startup. Installed files, drivers, and the game were not changed. Evidence is sony-timing-deployment-01/result.json.

Hardware verification remains pending. The old, pre-gate session had already lost the Bluetooth controller at 05:03:46 UTC with read error 1167 and a charge-battery message after reporting 0%. It removed its virtual import and helper. Thus the new build was validated for idle startup only; no connected controller, helper, converter echo, or new-gate physical quality is claimed.