Fetch the baseline
The shipped model's scores, so the comparison is against what users have today.
Agentic · ships today
Compress with a
quality budget
Fetch the shipped model's scores, quantize, re-evaluate, and gate the quality delta under a regression budget.
Ships as compress/quantize
The agentic form of model compression: an agent fetches the baseline scores, runs the quantization, re-evaluates the candidate and presents the delta. A small drop ships; an over-budget regression is blocked at the promote gate.
The pipeline
Quantize
Run the compression on your GPUs inside the budget.
Re-evaluate
Score the candidate on the same suites.
Gate the delta
A regression inside the budget passes to a person; over budget is blocked.
What the record keeps
The baseline scores, the compression settings, the candidate, the delta, and the promote decision.
Regression budget
You say how much quality you will trade; the gate enforces it.
Cross-pack
compress, model/eval and eval/delta compose without a domain pack.
See also
Compress a model in ML workflows covers distillation and side-by-side students.
Ships as compress/quantize. Supported model families depend on the installed operators.
See how failure can trigger a reviewed improvement cycle →Run it on your own infrastructure.
Apache 2.0 licensed, self-hosted. cairn pipeline run compress/quantize