Skip to main content
Help/Pressure/Research Notes

Research Notes: Pressure

Last updated: 2026-05-26 · Plugin v1.6.0

Current Test Status

  • Survived The Gauntlet™ (1,000,000 parameter mutations, 0 crashes)
  • Category 5 stress tests passing
  • Zero audio artifacts detected in stress testing
  • Pluginval clean (strictness 10)
  • Zero memory leaks over 1 hour marathon test

Validation Depth

801

Total Tests

ASan · UBSan · TSan · RTSan

Sanitizer Lanes

10

Pluginval Strictness

Production latency: ~85 µs/block on M4 Pro at 48.0 kHz / 512 samples

Audio Quality Specs

Measured at −12 dBFS, 60% pressure

EngineTHD RangeCharacter
MIX (VCA)0.02–0.06%Clean bus control
INST (FET)0.05–3.0%Fast, assertive control
VOCALS (Opto)0.4–0.7%Smooth vocal-focused leveling
DRUMS (VCA)0.01–0.08%Transparent dynamics
MASTER (VariMu)0.06–1.9%Wide-knee density control

Swipe sideways to compare

Engine Response Characterization

Measured 2026-05-31. Controlled lower-layer characterization at approximately 6 dB raw gain reduction.

These measurements are response fingerprints, not universal quality rankings or release thresholds. They do not claim preset, AUTO-mode, lookahead, oversampling, or full-plugin behavior.

ModeVoiceTime to 90% engagementRecovery to 10%Observed tail shape
MIXClean bus control12.0 ms481 msMulti-stage observed
INSTFast assertive control0.1 ms534 msNo multi-stage flag
VOCALSSmooth vocal-focused leveling11.8 ms5,704 msMulti-stage observed
DRUMSTransient-conscious control20.3 ms593 msNo multi-stage flag
MASTERWide-knee density control2.8 ms338 msNo multi-stage flag

Swipe sideways to compare

VOCALS: The final 1% recovery crossing remained above the measurement floor after the 20 s observation window.

What AUTO Does

AUTO is an engine-aware compression system. PRESSURE drives a tailored curve for the selected voice rather than applying one generic compressor setting.

  • Coordinates threshold, ratio, attack, release, and knee through per-voice curves.
  • Adds an input-aware threshold adjustment while preserving the selected voice's identity.
  • Uses TONE to shape engine-specific detector emphasis.
  • Keeps DRUMS-specific program-dependent release distinct from the shared AUTO curve system.

Internal characterization confirms that AUTO produces materially different behavior by voice and event duration. The published response table above remains the cleaner controlled comparison because it measures all five voices at approximately 6 dB raw gain reduction.

AUTO is not a universal material-adaptive attack/release algorithm, and these observations are not a five-engine AUTO timing table.

Framework Improvements

Issues identified through mutation testing and their resolutions.

  • DC offset detection added to buffer validation

    Mutation testing revealed that buffer validation did not yet cover DC offset. Detection was added with a −60 dBFS threshold, and all plugins were re-validated.

    Resolved 2026-02-01

  • Sample-level discontinuity detection added

    Mutation testing identified that the framework lacked sample-by-sample gradient analysis. A discontinuity detector with a 6 dB inter-sample threshold is being integrated to extend coverage to transient-level anomalies.

    In progress

Why Public Research Notes?

Mutation testing is how we verify the test suite itself. By intentionally injecting faults into the plugin, we identify where the testing framework needs to improve—and then we improve it.

This page documents that cycle: what we found, what we fixed, and how the framework got stronger. Engineering rigor, documented in the open.

If you encounter unexpected behavior our tests haven't covered, we want to hear about it.

Learn more about our testing methodology: Testing Suite Documentation

Found a Bug?

If you crash it or hear artifacts our tests missed, email me:

keith@bellweatherstudios.com

Include: DAW + OS version, what you did before it crashed, crash log if available. Best bug reports get credited in release notes.

These are research notes, not marketing claims. We publish the engineering process because trust is built on evidence, not messaging.