Capture
A run or evaluation fails. The failure, its inputs and its outputs are recorded.
ML · ships today
Learn from
model failures
Turn graded failures into the next candidate, inside a bounded loop with a person at the end.
Ships as training/retrain-cycle
A model that misses its bar is not the end of the run. Its failures are graded and mined, a new candidate is trained and compared on the same held-out set, and the loop repeats for at most six cycles before a person decides.
The pipeline
Mine
Graded failures become preference pairs or a corrected dataset revision, with lineage.
Retrain
Each cycle trains from the previous cycle's adapter on the pinned data, inside the budget.
Compare and decide
The candidate is scored against the approved model on the same test. After six cycles without a pass, the loop stops and asks a person.
What the record keeps
Every cycle's inputs, adapter, scores and spend; the cycle count; and the final decision.
Bounded by design
max_iterations is literal and mandatory; exhaustion asks a human, it never fails silently.
Same test every time
Cycles are compared on the same held-out set, so gains are real.
Approved model stays live
Nothing replaces it until a candidate beats it and a person signs off.
Ships as training/retrain-cycle. Trigger wiring depends on your deployment.
See how failure can trigger a reviewed improvement cycle →Run it on your own infrastructure.
Apache 2.0 licensed, self-hosted. cairn pipeline run training/retrain-cycle