Parity, measured

Generated by tools/parity_table.py. Every row is a comparison the test suite enforces; gate is the threshold that fails the build, measured is what this run actually observed. A row that says not measured was not run in this environment. It is left in rather than dropped, because a table quietly missing a row reads as a table with nothing to hide.

The rows are measured from Python: the weight-free rows against the shared fixtures in tests/data/conformance/, the weighted rows against the reference dumps in tests/data/reference/ and the release checkpoint. A row that names the four ports is held for them by their own harnesses over the same fixture files: go test ./..., cargo test, npm run test:all and swift test, which the parity job in CI runs. Never hand-type a number here; run the generator with --checkpoint on a machine that holds the weights.

Environment: Python 3.12.12, torch 2.13.0, onnxruntime 1.29.0, loudkit 0.1.1

stage	compared against	gate	measured
Philox 4x32-10	published known-answer vectors	exact	✓ 3/3 vectors
LR-SAMPLER-v1	shared fixture	exact	✓ 3/3 cases
Speech funnel	shared fixture (Python, Swift, Go, Rust, JS)	exact	✓ 217/217 cases
Long-form splitting	shared fixture (Python, Swift, Go, Rust, JS)	exact	✓ 48/48 cases
Postprocess detectors	shared fixture (Python, Swift, Go, Rust, JS)	exact	✓ 37/37 cases
EOS peak observation	shared fixture (Python, Swift, Go, Rust, JS)	step exact, probability rtol 1e-09	✓ 3/3 cases
Seed derivation	shared fixture	exact	✓ 6/6 cases
Token generator, teacher-forced	reference implementation	top-1 >= 99%, median KL < 1e-4	✓ top-1 100.00% (195/195), max median KL 1.76e-06
Token generator, free-running	committed reference stream	exact	✓ 446/446 tokens
Mel decoder, fixed tokens	reference implementation	corr >= 0.999	✓ corr 0.999999–0.999999 over 3 sentences
Vocoder, fixed tokens	reference implementation	corr >= 0.98	✓ corr 0.9951–0.9999 over 3 sentences
Re-render, same seed and build	itself	bit-identical	✓ identical
ONNX renderer	shared fixture	mel corr >= 0.999, wave corr >= 0.95	✓ mel 0.999991, wave 0.9992 (worst of 2)
CoreML renderer	shared fixture	mel corr >= 0.999, wave corr >= 0.95	✓ mel 1.000000, wave 1.0000 (worst of 2)
loudr-1-turbo ONNX renderer	shared fixture	mel corr >= 0.999, wave corr >= 0.999	✓ mel 1.000000, wave 0.9995 (worst of 2)
loudr-1-turbo CoreML renderer	shared fixture	mel corr >= 0.999, wave corr >= 0.999	✓ mel 1.000000, wave 0.9995 (worst of 2)

Exactness of a free-running stream is per architecture and per BLAS: the sampling law is the same everywhere, and the logits under it move at the last bit between machines, which can flip a sampled token. The figure in the row above is what this machine (arm64) reproduces.

Why each gate is where it is

• Philox 4x32-10: the RNG is checked against a standard, not against itself
• LR-SAMPLER-v1: min_p in logit space, gumbel-argmax, ties to the low index
• Speech funnel: invisibles, symbols, footnotes, punctuation, Polish respelling
• Long-form splitting: where the reader breathes; a different split is a different reading
• Postprocess detectors: where a chunk ended; a different verdict is a different cut
• EOS peak observation: audible despite never feeding back: two detector rules threshold on it
• Seed derivation: one user seed, independent per-stage streams
• Token generator, teacher-forced: the only generator comparison free of sampling chaos (EXP-010)
• Token generator, free-running: same law, same seed; pins generator and sampler stability, against the committed free-running reference
• Mel decoder, fixed tokens: same tokens and same injected noise, so a difference is arithmetic
• Vocoder, fixed tokens: gated loosely on purpose: predicted phase decorrelates, spectrum does not
• Re-render, same seed and build: identity class I-2: determinism within one backend
• ONNX renderer: a second backend does not get a second, looser bar
• CoreML renderer: a second backend does not get a second, looser bar
• loudr-1-turbo ONNX renderer: a second backend does not get a second, looser bar
• loudr-1-turbo CoreML renderer: a second backend does not get a second, looser bar