Scientific policy

Audivane prepares two codec chains from one decoded source, then runs a fixed-trial blind ABX comparison. Audio stays local. The desktop app runs native codecs in a private worker; the browser app uses local WASM codecs. This page is the public statement of the rules the app enforces.

Preparation and delivery

• Mono/stereo layout and the explicit session-rate policy are preserved. Both chains start from the same decoded source, including when that source is lossy.
• The aligned-gain-matched-v1 preparation aligns the pair, applies constant gain matching and a common peak trim, and records its actual cropping, gains and trim.
• Alignment, loudness arithmetic order and PCM16 quantization semantics are frozen. A change to scientific meaning requires deliberate policy and schema versioning.
• Audivane measures the final delivered PCM and verifies its exact bytes, sample rate, channels and frame count independently before binding it to a listening run.
• Pair relation, preparation validity and delivered validity are versioned separately. Historical evidence cannot authorize current scoring.
• Over-limit or malformed audio is refused. A decoder-capped prefix is never accepted as a complete source. See audio formats and limits.

Listening and scoring

Prepare test verifies the pair and stages trial 1 paused. Play is explicit. An A/B/X set must be complete and verified. An advancing trial has no answer authority or selected label until its explicit pick switches to that set. Pause and End remain effective during pending work.

The plan locks when the first trial is staged. Exactly the planned trial count is scored; an incomplete or permanently ended run is not a final result. Exports taken while a run is still going are redacted. Completed, sealed and void runs each retain audit evidence under their distinct lifecycle rules, and stay visibly distinct in the app.

Calibration is off by default. When you request it, it uses a separate first trial of at most ten seconds of prepared side A, which is excluded from the scored trial count. Canary material, listener calibration, preparation validity and null gates are separate checks; none can silently substitute for another.

The experiment engine owns blinding, the fixed-N binomial decision, confidence intervals and run evidence. The interface does not infer answers or score from playback labels.

What the evidence establishes

Audivane’s checks are designed to be verifiable: intended behavior is validated with analytical fixtures, applicable independent references and exact delivered audio. Agreement between implementations or passing existing tests can preserve the same defect, so neither is treated as proof on its own.

Two boundaries follow from this, and they are stated wherever results appear:

• Validity is not audibility. A passing preparation and a completed run establish that the comparison was constructed and scored as described. They do not certify audibility, codec conformance, tamper-proof operation or exclusive/bit-perfect output.
• Digital capture is not the physical chain. Digital checks verify bounded routing and playback behavior. Physical headphones, device wiring, drivers and OS or Bluetooth processing need separate listening checks — your hardware and listening conditions remain part of the experiment.

A result from Audivane is evidence about your files, your machine and your hearing on the day of the test. It is not a general verdict about a codec.