Problem
We report the distillation cost as "about 2% of the teacher's Gini." That is honest and precise, and it is also not the unit anyone approves budget in.
A CRO asks: what does that cost in expected loss on a book of this size? Without that translation, the trade-off cannot be evaluated against the alternative — which is usually not "no model" but "deploy the teacher with a full MLOps stack."
Proposed approach
A helper that turns a Gini delta into portfolio economics under stated assumptions:
loss_impact(artifact, teacher_latent, y, X, exposure, approval_rate, lgd=...) →
- bad rate at the same approval rate, teacher vs artifact
- incremental expected loss per year at a given book size
- swap-set analysis: which accounts the artifact approves that the teacher declines, and vice versa
Assumptions must be explicit and caller-supplied. This is a calculator, not a claim.
Acceptance criteria
Files
New helper in compileml.tune or a small compileml.economics, plus docs
Scope
Small to medium. Straightforward analytics; the care goes into not overclaiming.
Problem
We report the distillation cost as "about 2% of the teacher's Gini." That is honest and precise, and it is also not the unit anyone approves budget in.
A CRO asks: what does that cost in expected loss on a book of this size? Without that translation, the trade-off cannot be evaluated against the alternative — which is usually not "no model" but "deploy the teacher with a full MLOps stack."
Proposed approach
A helper that turns a Gini delta into portfolio economics under stated assumptions:
loss_impact(artifact, teacher_latent, y, X, exposure, approval_rate, lgd=...)→Assumptions must be explicit and caller-supplied. This is a calculator, not a claim.
Acceptance criteria
Files
New helper in
compileml.tuneor a smallcompileml.economics, plus docsScope
Small to medium. Straightforward analytics; the care goes into not overclaiming.