Recovery methods
From pure distillation failure to anchored transfer.
The method changed only when immutable evidence justified the next bounded condition. The product keeps direct labels, upstream teacher labels, and distribution-design labels separate.
Independent distillation
768 → 59
Strict teacher filtering left 5 of 16 archetypes represented. No target training occurred.
Terminal negativeDistribution matching
768 → 719 of 768
4 of 48 duplicate auto-refund outputs passed; 44 were exact policy mismatches.
Blind spot isolatedAnchored transfer
214 + 10
224 unique records trained fresh OLMo for 272,568 processed tokens.
Completed conditionLabel lineage
Ten direct anchors are not the whole cost.
Anchored transfer consumes ten original labels directly at the target, but its teacher and distribution design rely on earlier labeled data.
Original labels directly used by target
10
Exact strict teacher outputs used by target
214
Unique target training records
224
Original labels used upstream to train teacher
224
Labeled records used to design distribution
224
What stayed fixed
Model target
Fresh pinned OLMo
Training budget
≤272,643 whole-sequence tokens
Evaluator
Parser 0.1.0 · metrics v0
Selection
Safety-first frozen checkpoint rule