Docs

Evidence

Evidence comes in three tiers that are never mixed on the site: offline automated checks, device reports and acoustic measurement.

docs/EVIDENCE.md · edit on GitHub

Tiers#

TierToday
1. Offline, automatedreleases/<v>/acceptance.json, produced by build/build.py. The frozen 7.0.0 file has 19 checks per build: 108 pass, 0 warn, 6 not run (cpu). CI rebuilds with 24 per build, adding the release gates and stats.derived.
2. Device-reporteddocs/evidence/device-tests.json. All six v7.0.0 builds have an author-reported listening pass.
3. Acoustic measurementNone yet; the site keeps saying "calculated".

Device results are the author's own listening tests, not instrument measurements.

Tier 1: offline checks#

Each build is checked for file validity, VDC banks and sections, IRS format and limits, EEL static lint, response, limiter and runtime behaviour. Status is pass, warn, fail or not run.

CheckResult
response.*pass. The shipped VDC times IRS reproduces each stored correction curve to 0.03 dB RMS on all six builds.
irs.peak_limit, irs.l1, irs.tail, vdc.pole_radiusRelease gates, see Limits. Enforced on every build by CI.
eel.compile, limiter.*, tail, nan_soakpass. Run through tests/eel (eel2py).
vdc.sr_consistencypass: the 44.1 kHz bank is within 0.3 dB of the 48 kHz bank.
stats.derivedWarns if stored profile stats drift from the files on a fine grid. A warning is not a failure.
cpunot run. Needs a device.

The response convention is documented in the build/response.py docstring: predicted is raw plus correction plus a constant level alignment.

eel2py is an independent interpreter, not the WDL/JSFX runtime. Its assumptions are listed in tests/eel/eel2py.py. Treat it as strong offline evidence, not device proof.

Tier 2: device reports#

One entry per build, device, date and result, in docs/evidence/device-tests.json.


{"release","build","device","date","test","result":"pass|fail|untested","notes"}

Tier 3: acoustic measurement#

There are no acoustic measurements. Every curve on the site is calculated.

Speaker cap#

The v7 speaker chains have the same response as the earlier builds, so these figures still apply.

  • Applied peak boost is +7.3 to +8.2 dB, against +10.7 to +11.6 dB needed.
  • The mean shortfall over 100 to 2000 Hz is about -2.0 dB.
  • The worst single point is -11 to -12 dB.

True-peak limits#

Measured with eel2py at 8x oversampling, as overshoot above the ceiling. The shared v7 limiter core gives a worst case of about +0.25 dB on tones, clicks and +6 dB broadband noise, against a gate of 0.5 dB. The v5 limiter overshot by +1.7 to +1.95 dB on noise.

Sample peaks never exceed the ceiling because of a hard clamp. These figures come from a subset interpreter, not the WDL runtime; the device check is the final authority.

Raw data#