Reusing a listening correction#

In brief. When someone listens to a Stitch trial and says "bar two lands late", that note is stored against the exact audio and time span they heard. This page shows how to find those notes again later, filtered by version, span or topic, without stretching them to cover anything they didn't say.

record_feedback stores a listener's words against one exact output hash, a local frame interval and a scope such as bar_phase, timing or tonal_overlap. query_feedback retrieves these records without broadening their meaning or resolving disagreements on the listener's behalf.

from pocket_music.feedback import query_feedback

result = query_feedback(
    "/path/to/trial",
    variant_id="v01",
    start_frame=48000,
    end_frame=96000,
    scope="bar_phase",
    limit=20,
)

Coordinates are frames in the selected output file, not source-recording or arrangement time. Supply both interval endpoints or neither. Filtering uses interval overlap with an exclusive end; returned notes keep their full original interval and text. Two contradictory claims are returned as two claims. Bar-position feedback never becomes timing, tonal or full-mix approval.

The query verifies the sealed trial and feedback records and current selected output hashes. A changed output or tampered note fails explicitly. output_sha256 can constrain the exact recording in addition to variant_id; an identity outside the trial is an error. Empty notes means no stored matching feedback, not a favorable musical verdict.

The result includes explicit pagination (total, offset, limit, next_offset), the applied filter, and musical_verdict: null. limit defaults to 20 and is capped at 100. The same public function is exposed as the MCP tool stitch_feedback_list and through a JSON argument file:

pocket stitch feedback-list --spec /path/to/feedback-query.json

Keep real listener words and trial audio outside Git. Generated test claims exercise this contract without claiming a person auditioned the fixture.