Nested Composition / V95
A composition used as one member of another answered a new question at 84.0% on 1,000 sealed runs with only three trained values, above both frozen members (78.0% and 80.0%) and above boosted trees trained on the same runs (82.6%). All four gates passed.
SCOPE + LIMITATIONS
- Most of the gain is the outer combiner recalibrating one member to a new question (weight -1.05 on the failure composition against 0.07 on loss_halves); the second member contributes little here. Two members, one question, a linear combiner, one seed.
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
Ask whether a training run trains well: makes progress and halves its loss within 200 steps. The outer composition has two frozen members, the earlier failure composition (itself two predicates and a combiner, locked by checksum) and the transferred capability loss_halves, and owns only a 3-value combiner. Fit the combiner on 1,000 fresh executed runs no member had seen, then score once on 1,000 more, of which 40.4% train well.
Members unchanged, at least two points above the better member (+4.0), only the combiner learned, and restoration from its manifest in a fresh process all passed. Composition now works one level deeper: a composition is a member of another.
- Outer nested composition · 3 trained values
- 84.0%
- Failure composition alone
- 78.0%
- loss_halves alone
- 80.0%
- Boosted trees on raw features, same 1,000 runs
- 82.6%
Question and method
Ask whether a training run trains well: makes progress and halves its loss within 200 steps. The outer composition has two frozen members, the earlier failure composition (itself two predicates and a combiner, locked by checksum) and the transferred capability loss_halves, and owns only a 3-value combiner. Fit the combiner on 1,000 fresh executed runs no member had seen, then score once on 1,000 more, of which 40.4% train well.
Recorded results
| Condition | Recorded result |
|---|---|
| Outer nested composition · 3 trained values | 84.0% |
| Failure composition alone | 78.0% |
| loss_halves alone | 80.0% |
| Boosted trees on raw features, same 1,000 runs | 82.6% |
Comparative standing and locked gates
Members unchanged, at least two points above the better member (+4.0), only the combiner learned, and restoration from its manifest in a fresh process all passed. Composition now works one level deeper: a composition is a member of another.
Interpretation and limitations
Most of the gain is the outer combiner recalibrating one member to a new question (weight -1.05 on the failure composition against 0.07 on loss_halves); the second member contributes little here. Two members, one question, a linear combiner, one seed.
SOURCE PROVENANCE
V95: a nested composition, a composition used as one member of another
LABORATORY REPORT / 2026-10-21SOURCE CHECKSUM / SHA-256
55129ea549f5a1940db907e389701e4125c7f23975dd7a9d452ec6372ea59670Public journal edition reviewed 2026-10-22. 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.