Software
Four open-source Python libraries, all on PyPI, all MIT-licensed, all installable. They are deterministic checkers rather than models: each one reads the artifacts a machine-learning run already produces — a trainer state, a batch of synthesised audio, a set of speaker trials, a validator’s own output — and reports what is wrong with them, the same way every time.
They exist because of a single recurring failure. A run, a synthesis, a verification or a check can fail in a way that produces a plausible success: a falling loss curve over corrupted data, a clean pass over a run that died hundreds of steps ago, a verdict returned over evidence that was never examined. A fault that raises an error is fixed that afternoon. A fault that returns a good-looking number is shipped.
The rule they share is that an absent or unmeasurable value must never be able to read as a good one. Where a check cannot run, these libraries report that it did not run — they do not quietly skip it and pass.
3 articles in this section · all 25 articles, newest first →
The libraries
Versions, release counts and test counts here are harvested from PyPI and the GitHub Releases API rather than typed by hand.
ttsproof
v0.4.0activeAutomated failure-mode QA for text-to-speech.
trainproof
v0.21.0activeA deterministic linter for ML training runs.
notchecked
v0.2.0activeCoverage accounting for validators: what was checked, what could not be checked, and what was never in scope.
spkproof
v0.3.0activeDeterministic checks for speaker-verification studies.
What each one catches, and what it does not
Three of the four libraries have an article describing the failure that produced them. Every one of those articles is also honest about the failures the library cannot see.
Loss Curves Lie: Building a Deterministic Linter for ML Training Runs
July 2026A deterministic linter for training runs. A run learning pure noise reduced its loss by 62% and looked textbook-healthy. The article covers what can be caught from a trainer state and is explicit about what cannot — and a later release exists because the linter returned a clean pass on a run that had been dead since step 11.
Your TTS Model Sounds Great — Until It Says "GPUB"
July 2026Automated failure-mode QA for text-to-speech: the model that sounds excellent until it reads an abbreviation aloud as a word. Backed by a published 390-sample study.
The Same Bug, Four Times, Three of Them Mine
August 2026Coverage accounting for validators. A validator that reports a verdict without reporting its coverage is asserting something it did not measure. The same defect was found in four systems, three of them the author’s own, before it was admitted to be one problem.
spkproof has no article of its own. The study it came out of is the fourteen-encoder panel in Benchmarks.