Technical Quality
CORPUS is being built in the open. Some of what you read here is live, some is still design intent. Expect it to evolve.
Every upload gets an automated technical report: one file in, a plain-language damage report out, and the points base of the score derived from it. This is the measured half of quality: the report does not judge music, style, or taste. It answers two questions: is this signal intact, and does it really carry what its format claims?
The base: measured format (20–110)
| What the signal contains | Base points |
|---|---|
| Lossless, 24-bit, 44.1/48 kHz | 100 |
| Lossless, genuine high resolution above 48 kHz | 110 |
| Lossless, 16-bit content | 90 |
| Near-transparent lossy (MP3 ≥ 256 kbps class) | 60 |
| Standard lossy (MP3 128–192 kbps class) | 50 |
| Heavily lossy | 40 |
The measurement decides, not the label: a WAV that carries a repackaged MP3 scores as an MP3, a "96 kHz master" that turns out to be an upsampled CD earns no high-resolution bonus, and a 24-bit file with 16-bit content scores as 16-bit. Concealment itself carries no extra penalty; misrepresentation is handled as an integrity matter in The Contribution Process.
Deductions
Findings reduce points only when the evidence is strong. Ambiguous cases (glitch textures, analog noise aesthetics, artistic filters) appear in the report as information with reduced confidence and never deduct automatically. What does deduct is weighed by how much training value is lost, not by how bad it sounds: an abrupt ending costs nothing because the remaining audio is fully usable, while clipping costs a lot because the distortion sits in every loud frame.
- Damage baked into every frame (hard clipping, bad resampling, comb filtering) costs the most, up to 30 points.
- A steady noise carpet costs up to 17; historic transfers stay valuable.
- Localized damage (clicks, dropouts, bad splices) costs by the seconds affected. The report timestamps it, so training can skip exactly those parts.
Deductions cap at half the base, and no scored file drops below 20 points.
What costs nothing
Production choices and trivially correctable properties: mains hum, mono content in a stereo container, DC offset, an abrupt ending, loudness (a fact, not a defect), and every lo-fi or filtered aesthetic. The style-relative reading of the mix lives in The Sound Quality Reading and stays advisory.
Multichannel
A 5.1, 7.1, or Atmos/ADM delivery gets the full technical report but no points. Stereo masters are the ingest currency; submit the stereo master for scoring.