Public evaluation record
Make the comparison inspectable.
This page records how Prism will be assessed, what listeners can inspect, and what we will not claim before validated outputs exist.
Current status
No validated Prism results are published yet.
There is no model version, benchmark score, or A/B result to report on this page. We will add results only after the outputs, listening protocol, and source details can be checked together.
The evaluation frame
A result needs its listening conditions.
A useful comparison must make the source, the candidate output, the listener task, and the limits visible at the same time.
- 01
Document the source
Name the recording, licence, excerpt boundaries, track grouping, and every preparation step before comparison begins.
- 02
Freeze the outputs
Publish the exact Prism build, settings, output files, and any level matching used for the listening set.
- 03
Blind the listening task
When matched candidate outputs exist, listeners will compare anonymised clips against a defined prompt. We will not simulate this step with reference stems.
- 04
Publish the record
Report the protocol, source set, version, outputs, and limitations alongside any measured or listener-reported result.
Reference material
Compare perspectives, not models.
Use the player to move between the licensed studio reference tracks and the original mix. It is a listening aid for the material under discussion, not an A/B test and not a Prism output.
These are licensed reference studio multitracks from the source recording. None is a Prism output, and this player does not present a blind A/B result.
About this excerptWhat is recorded now
The evidence currently available.
- Source recording
- Karma — Tom McKenzie, licensed CC BY 4.0 through AlbumDB / Zenodo. The linked credits contain the full source and licence details.
- Excerpt and tracks
- A 30-second excerpt from 150–180 seconds, supplied as studio multitracks and grouped into six reference listening buses plus the original mix.
- Prism version
- No validated Prism output or version is published for this reference material.
- Blind listening
- Not run. A blind A/B requires matched, anonymised candidate outputs; this page contains only reference material.
- Results
- Not published. Results will appear with the exact files, protocol, and limits used to produce them.
Limits
What this reference cannot establish.
- One licensed 30-second excerpt is a reference example, not a representative benchmark set.
- The six listening buses are a source-specific studio grouping and do not define Prism evaluation categories.
- Piano and Other use additional audition gain so quiet details can be heard; that makes this unsuitable for level-based comparison.
- No numerical, comparative, or listener-preference result is implied by these tracks.