Reusable Predicate Efficiency / V65
Reusing a frozen predicate reached the 95% development floor with 64 new examples, but the best from-scratch control reached it with 32. The locked fourfold data-efficiency gate failed; the sealed set remained closed and no candidate was promoted.
CONTROL COMPARISON
Development sample efficiency at the 95% floor, across three data draws and bounded fitting budgets.
The locked gate compares reuse against the best scratch control and requires a fourfold advantage. That gate failed; the descriptive worst-draw comparison does not replace it.
- Reuse · worst draw
- 64 examples
- Scratch · best draw
- 32 examples
- Factorized scratch · worst draw
- 128 examples
Method
Compare a reused 354-parameter predicate plus a new 50-parameter combiner against from-scratch factorized and flat models on three-effect interval cancellation. Six training sizes from 16 to 512, three draws, and nine fit budgets per cell yield 486 fits. The gate requires reuse to reach 95% with at most 128 examples and the best scratch control to require at least four times as many.
Results
Reuse reached the floor at 64 examples in the worst recorded draw; the best scratch control reached it at 32. The required ratio failed. A comparable worst-draw factorized scratch condition needed 128, a descriptive twofold difference that does not satisfy the locked fourfold claim. The flat model stayed below 72% across sizes. Sealed scoring was not performed and no candidate was promoted.
Interpretation and limitations
Factorization accounts for much of the gain over the underfit flat model. A library-level fourfold efficiency advantage is not established. Earlier predicate acquisition used 822 labels, a separate cost that must be retained when interpreting reuse.
SOURCE PROVENANCE
EMMA V65: does reusing a frozen module pay in data?
LABORATORY REPORT / 2026-10-07SOURCE CHECKSUM / SHA-256
189460544e01f9c9f509e6af6950518d10fcc797803f1311c71e9b431c1ffe1dPublic journal edition reviewed 2026-10-08. Source documents and saved evidence were inspected; experiments were not rerun for this edition. Proprietary implementation code, model binaries, private infrastructure, and detailed machine records are not published here. Journal identifiers are editorial references. Catalog inclusion does not imply qualification or runtime promotion.