Linked Decisions with Replaceable Weights / V78–V78.2
Separating operation matching from applicability improved development results, but both linked-Decision versions failed qualification. V78.2 reached 74.2% on held-out operations versus V78’s 67%; wrong-operation routing remained 18.2%, above the eight-percent ceiling. Sealed questions stayed unopened.
CONTROL COMPARISON
Give Match and Applies separate slots and weight sets, then compose their answers. Initial comparisons pair lab marker heads with Laya open weights and single-head controls. V78.2 fine-tunes Match with varying offered operations while training the lab Applies head longer. Comparisons use the same spent development questions; no independent sealed result exists.
Both versions failed fit and decision-quality gates despite gains over single-head controls. V78.2 is retained best for the linked Decision, not qualified. Its Match weights are registered independently, but the complete composition remains unreliable. Earlier weight sets remained bit-identical after new ones were registered and used.
- V78 complementary pairing · held-out operations
- 67.0% correct
- V78.2 fine-tuned Match plus lab Applies
- 74.2% correct; 63.7% required declines; 18.2% wrong operation
- Lab single-head reference
- 41.2% held-out accuracy
- Open single-head reference
- Approximately 39.3% held-out accuracy
- V78.2 fine-tuned Match alone · answerable requests
- 93.0% first choice on untrained operations
- Match-only registration probe
- 95.0%; does not qualify the whole Decision
Question and method
Give Match and Applies separate slots and weight sets, then compose their answers. Initial comparisons pair lab marker heads with Laya open weights and single-head controls. V78.2 fine-tunes Match with varying offered operations while training the lab Applies head longer. Comparisons use the same spent development questions; no independent sealed result exists.
Recorded results
| Condition | Recorded result |
|---|---|
| V78 complementary pairing · held-out operations | 67.0% correct |
| V78.2 fine-tuned Match plus lab Applies | 74.2% correct; 63.7% required declines; 18.2% wrong operation |
| Lab single-head reference | 41.2% held-out accuracy |
| Open single-head reference | Approximately 39.3% held-out accuracy |
| V78.2 fine-tuned Match alone · answerable requests | 93.0% first choice on untrained operations |
| Match-only registration probe | 95.0%; does not qualify the whole Decision |
Comparative standing and locked gates
Both versions failed fit and decision-quality gates despite gains over single-head controls. V78.2 is retained best for the linked Decision, not qualified. Its Match weights are registered independently, but the complete composition remains unreliable. Earlier weight sets remained bit-identical after new ones were registered and used.
Interpretation and limitations
One teacher authored the operations and requests. Match accuracy is conditioned on answerable questions and must not be substituted for overall safe routing. Applies remains the weak link. V78’s pairing also trails the earlier 90% and 95% references on the separate sixty-question test, illustrating task-dependent standing. Reserved questions and the sealed theme remain unopened.
SOURCE PROVENANCE
V78: linked Decisions with plug-in weights (Match and Applies)
LABORATORY REPORT / 2026-10-09SOURCE CHECKSUM / SHA-256
97c401660125a22ce1f24e87b52845c52d6ae390390bf6f2ef5a585a2ff2be27Public journal edition reviewed 2026-10-10. 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.