GitHub View workflow

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.

01

Independent distillation

768 → 59

Strict teacher filtering left 5 of 16 archetypes represented. No target training occurred.

Terminal negative
02

Distribution matching

768 → 719 of 768

4 of 48 duplicate auto-refund outputs passed; 44 were exact policy mismatches.

Blind spot isolated
03

Anchored transfer

214 + 10

224 unique records trained fresh OLMo for 272,568 processed tokens.

Completed condition

Label 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.

Accurate description: ten original anchors plus 214 teacher outputs, with 224 upstream teacher labels and 224 labeled records used for distribution design.

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