How these recordings are made

This page describes how every recording in this catalogue is made. It is written for listeners, for librarians, and for the review teams at the services that carry this catalogue. Nothing is left out to make the result sound more impressive than it is.

The works

Every work published here is in the public domain. We use a conservative test rather than the earliest date we could argue for: the composer must have died in 1925 or earlier, and the work must have been first published in 1930 or earlier. If either date is unknown, the work is not used.

The printed edition we read the notes from must itself be in the public domain, published in 1930 or earlier. We take the composer's own markings and the notes; we do not reproduce a modern editor's fingerings, interpretive marks or critical apparatus.

From printed page to notes

A public-domain printed edition is converted into a machine-readable score. The result is checked by software before anything else happens: every bar must add up to its time signature, no note may have zero length, no voice may collide with itself, repeats must be written out, ornaments must be resolved according to the conventions of the work's period.

To catch transcription errors, two independent public-domain editions of the same work are converted separately and compared. Only the places where the two disagree need a human to look at the original page and decide. That keeps the error rate low without anyone having to proofread hundreds of bars by eye.

There is no human performer

This is the part most easily misunderstood, so it is stated without hedging: nobody plays these recordings.

A plain score is a list of notes, all equally loud, all exactly on the beat. That is not music. What turns it into music here is a set of written rules applied by software — phrase shaping in both time and volume, a slight easing at the end of a phrase, weighting of the melodic line against its accompaniment, pedal changes placed at harmonic changes, chords spread by a few milliseconds, and small irregularities in the timing of each note.

Every one of those rules is deterministic, and every one is driven by numbers in a parameter file: a tempo range, a permitted depth of rubato, a dynamic range, a pedal density. The same input always produces a byte-for-byte identical result. There is nothing improvised and nothing learned from anyone's playing.

Those parameter files differ by period — Baroque, Classical, Romantic — and by intended use: a version meant for close listening is shaped differently from one meant to be played quietly in the background for hours.

Reference ranges, not imitation

Where a work has been measured, the permitted ranges come from several mature commercial recordings of it, with only the interval kept — a fastest and a slowest tempo, a widest and a narrowest dynamic range. Nothing about any individual performance is retained. There is no file anywhere in this system that could hold one pianist's timing, because the data structure only has room for a minimum and a maximum. Until a work has been measured this way, the tempo falls back to the marking printed in the score itself.

The purpose is to stay inside what the work can bear, and to catch gross errors. It is not to approach, imitate or reproduce any performer's interpretation.

The instrument

The performance data is played by a commercially licensed virtual piano, used under a licence that permits commercial release of the resulting recordings. We keep the invoice, the licence terms and the written confirmation on file for every release.

The resulting audio is then levelled: a fixed tone adjustment, a gain change to reach a loudness target, a resample, and dithering. No dynamic processing, no limiting, no effects. The master is ours, made from scratch; we never re-release anyone else's recording.

Where AI is used, and where it is not

The claim "this is not AI-generated music" is only worth making if we are precise about where AI does appear. So:

• The audio contains no generative AI and no machine-learning model. Not in the notes, not in the performance data, not in the instrument, not in the mastering. This is enforced by automated tests that fail the build if any such library is introduced into the audio path.
• The software that performs the score was written with the help of an AI coding assistant. That assistance produced source code. The code is deterministic rule-based software, and the assistant takes no part in producing any recording.

The distinction is between a tool that helped build the machine and a model that generates the output. Here it is the first, and never the second.

Quality control

Every recording is checked by software against fifteen technical tests, among them clipping, loudness and loudness range, dropped or stuck notes, pedal held through a harmonic change, abrupt tempo jumps, a truncated start or end, an unnatural reverb tail, channel faults, and metadata errors. Anything that fails does not reach a human; the report names the parameter to change, and the recording is made again.

What does reach a human is a short review of automatically selected excerpts — the opening, the densest passage, the quietest and loudest moments, the tempo changes, the end. The reviewer answers a fixed list of pass-or-fail questions about whether the recording meets its written specification. They do not score it, rank it, or say whether they liked it. A failure has to cite which rule was broken and at what moment, or it is not recorded as a failure.

What this catalogue is not

• Not performances by a human pianist, and not presented as such.
• Not reissues, remasters or transfers of existing recordings.
• Not generative AI music.
• Not associated with, endorsed by, or named after any performer, living or dead.

Questions

If something here is unclear, or you are reviewing this catalogue for a service and need documentation, write to us — the address is in the footer of every page. We keep the production record for every release: the score source, the parameter file, the instrument licence, and the test results.