Preserve fractional-millisecond decoder evidence in timing evaluator

This commit is contained in:
karl committed 2026-10-13 20:52:35 +02:00
1 parent 2c3c91179b
commit f15ee07ed0
2 files changed
+35 -5

No files matched your search

@@ -138,3 +138,17 @@ missing source windows prohibit numeric accepted timing scores. Historical
full-source gated DTW **6/8** is context, never an inherited result or an isolated
causal comparison with this end-to-end hybrid run. No new timing score exists
until this exact mode executes and all checks finish.
The first attempt at `.searchi/qa/speech-live-timing-v1` failed during source
preflight, before any model load or inference. Its immutable `failure.json`
SHA-256 is `9ba49fd5cdb69e57c267f579818f226949b02eafbc0e592161fad7593f00c6aa`.
The bark recording lasts 2.038594 seconds: its rounded 2039 ms window implies
32624 samples, while aligned decoding correctly retains 32618 real samples.
The six-sample difference is 0.375 ms. The evaluator now accepts the existing
one-millisecond coverage bound (16 samples), without padding, trimming or
changing source bytes, labels, ranking, or DTW points. Its regression test
accepts that real tail and rejects a 17-sample discrepancy and invalid bounds.
Seven Rust tests, strict all-target Clippy and five timing-adapter tests passed.
A retry must use a fresh output directory after the corrected evaluator is
frozen and the CPU inference slot is granted again; the failed attempt remains
preserved and has no quality score.
@@ -60,6 +60,18 @@ fn file_hash(path: &Path) -> Result<String> {
}
Ok(format!("{:x}", state.finalize()))
}
fn validate_sample_coverage(samples: usize, duration_ms: u32) -> Result<()> {
let expected = duration_ms as usize * 16;
ensure!(
samples > 0
&& samples <= 480000
&& expected > 0
&& expected <= 480000
&& samples.abs_diff(expected) <= 16,
"Fixture source clock coverage differs beyond existing one-millisecond bound"
);
Ok(())
}
fn write_new(path: &Path, value: &Value) -> Result<()> {
let mut file = fs::OpenOptions::new()
.write(true)
@@ -174,16 +186,13 @@ fn run_case(
let mut source_windows = Vec::new();
for window in windows(metadata.audio_start_ms, metadata.audio_end_ms, max_windows)? {
let pcm = decode_audio_aligned(&assets.tools, &path, window, &limits)?;
ensure!(
pcm.len() == window.duration_ms as usize * 16,
"Fixture source clock coverage differs"
);
validate_sample_coverage(pcm.len(), window.duration_ms)?;
let sha = hash(
&pcm.iter()
.flat_map(|value| value.to_le_bytes())
.collect::<Vec<_>>(),
);
let record = json!({"asset":id,"start_ms":window.start_ms,"duration_ms":window.duration_ms,"samples":pcm.len(),"pcm_sha256":sha});
let record = json!({"asset":id,"start_ms":window.start_ms,"duration_ms":window.duration_ms,"samples":pcm.len(),"nominal_samples":window.duration_ms as usize*16,"coverage_difference_samples":pcm.len() as i64-i64::from(window.duration_ms)*16,"pcm_sha256":sha});
expected.entry(sha).or_default().push(record.clone());
source_windows.push(record);
}
@@ -487,6 +496,13 @@ mod tests {
assert!(values.iter().all(|window| window.duration_ms <= 30000));
}
#[test]
fn fractional_millisecond_tail_preserves_real_samples() {
assert!(validate_sample_coverage(32618, 2039).is_ok());
assert!(validate_sample_coverage(32607, 2039).is_err());
assert!(validate_sample_coverage(0, 1).is_err());
assert!(validate_sample_coverage(480001, 30000).is_err());
}
#[test]
#[ignore = "Requires the five hash-verified real media fixture files; no inference"]
fn timing_mode_preflight_checks_all_real_frozen_sources_without_models() -> Result<()> {
let root = Path::new(env!("CARGO_MANIFEST_DIR")).join("../../../fixtures/media");