Parent FlowStrument demo page

<aside>

This is the promotion gate for the energy-aware retrained universal flow. It is still one 153,601-parameter model with no instrument-family gates or fallback routes. All output clips are source-level matched. Compare Accepted flow + exact energy anchor against Energy-aware retrained flow, focusing on attack, stable sustain, timbre identity, and f0 dynamics.

</aside>

Objective results

Across 25 shifted acoustic notes, the candidate wins 14, loses 11, and improves mean LSD by 0.0364 dB. No row is more than 0.0431 dB worse than the no-formant R3 carrier. On 14 trackable in-the-wild transfers, candidate RPA@50c is 0.9732 versus 0.9762 accepted, 0.9727 R3, and 0.9004 HyperGANStrument.

<aside>

The decisive risk rows are Brass +2, the largest loss versus the accepted model though still better than R3, and out-of-domain acoustic guitar A5, where RPA@50c is 0.959 versus 1.000 accepted. Please judge whether either has an audible pitch or timbre regression.

</aside>

Priority risk checks

Brass acoustic 016, +2 st

Largest candidate loss versus accepted flow

Largest candidate loss versus accepted flow

In-the-wild acoustic guitar, A5

<aside>

Reference status: no real same-instrument A5 performance was recorded. This is a single-prompt out-of-domain transfer from a guitar clip measured near MIDI 51.93 to A5 (MIDI 81). The disabled reference slot below is intentional.

</aside>

Largest observed out-of-domain pitch-accuracy dip; real target explicitly marked unavailable

Largest observed out-of-domain pitch-accuracy dip; real target explicitly marked unavailable

Listener-confirmed regression guards

Guitar acoustic 010, -3 st

Previously failed formant-preservation case

Previously failed formant-preservation case

Keyboard acoustic 004, -4 st

Previously attack-sensitive case

Previously attack-sensitive case

Mallet acoustic 056, +18 st

Previously attack-sensitive wide-shift case

Previously attack-sensitive wide-shift case

String acoustic 056, +3 st