ON-POLICY DISTILLATION · A DERIVATION MONOGRAPHThe measure
The measure
beneath the loss.
“On-policy” chooses where the prefixes come from. KL direction chooses what happens after a prefix is held fixed. Follow both coordinates through one exact two-token receipt.
OUTERsource measure×INNERlocal divergence
independent choices
EXPERIMENTAL HYPOTHESIS · UNPROVEN
Factoring the objective this way may help a mathematically inclined reader stop conflating on-policy sampling with reverse KL. This artifact contains no learner-outcome evidence.
IMMUTABLE TECHNICAL KERNELOPD-CORE-v1
9AAD0A2560BFF8696A65B49F5AA72058A75C102AA2C9E17535C40471FFD0E697Verified before build. All probabilities are invented teaching inputs. No model, dataset, paid API, or runtime CDN is used.