Collapse Signatures: What Broken Training Data Actually Looks Like
June 6, 2026 · 7 min read
By John “Virus”, Machine Learning Engineer at Genovo Technologies
Collapsed data has a look. After enough corpora, the signatures become as recognizable as a stack trace: variance that is suspiciously well-behaved, tails that end too early, categorical distributions with the rough edges sanded off. The first eight hours of every Synthos validation are spent matching a corpus against this catalog before any proxy model trains.
The catalog exists because generator models are systematic in how they fail. A model sampling from its own learned distribution under-represents whatever it was uncertain about — and uncertainty concentrates in the tails, the rare classes, and the awkward correlations. Recursive generations amplify exactly those omissions.
Five dimensions, five failure styles
Each scored dimension corresponds to a distinct way data goes wrong:
- Distribution fidelity — marginals drift toward the mode; tail mass evaporates a few percent per generation.
- Feature correlation — relationships tighten unnaturally as the generator learns a simplified joint distribution; real-world messiness reads as noise and gets cleaned away.
- Temporal consistency — seasonality and burstiness smooth out; synthetic time series are too polite.
- Outlier structure — anomalies are the first casualties; a fraud corpus without weird transactions is a fraud corpus in name only.
- Schema compliance — the quiet killer: units drift, encodings shift, a column’s meaning changes between shards while its name stays put.
Signatures buy speed, cascades buy certainty
Signature matching alone would over-flag unusual-but-fine data — some real corpora are just strange. That is why pre-screening only shapes the plan: it decides where the proxy cascade should look hardest, and the cascade delivers the verdict. Pattern recognition sets the hypothesis; controlled training runs test it. Skipping either half is how validation products end up wrong in both directions.