Consolidation into the Library Weight Tree / V94
Three units learned in operation (915 values) were distilled into one consolidated module of 609 values that scored 83.2% on 3,000 fresh runs, against 82.9% for the unit stack and 78.3% frozen. The module matched the stack’s decisions on 96.8% of configurations against a 98% fidelity gate, so V94 failed that gate; every other gate passed.
SCOPE + LIMITATIONS
- Distillation smoothed the boundaries between the units’ sub-regions, which is where the decisions differ; on balance those changes are improvements. Consolidation is built and in use, and the earlier states stay restorable. One distillation, one family of tasks.
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
Distil the three promoted units of V82.5, applied in promotion order, into one module using 14,307 configurations the system wrote itself inside the units’ scopes; none was executed. Record the base composition, the archived units and the consolidated node in the library weight tree. Evaluate all three states on 3,000 fresh executed configurations, 500 per regime.
Fidelity failed at 96.8% against 98%. Accuracy within half a point of the stack, unchanged answers outside the scopes, fewer values, fresh-process restoration from the tree, and lineage with members archived rather than deleted all passed.
- Unit stack
- 82.9%
- Consolidated module
- 83.2%
- Frozen
- 78.3%
- Decisions equal to the stack
- 96.8%; gate 98%
- Outside the units’ scopes
- 1,471 configurations bit-identical to frozen
- Size
- 609 values against 915
Question and method
Distil the three promoted units of V82.5, applied in promotion order, into one module using 14,307 configurations the system wrote itself inside the units’ scopes; none was executed. Record the base composition, the archived units and the consolidated node in the library weight tree. Evaluate all three states on 3,000 fresh executed configurations, 500 per regime.
Recorded results
| Condition | Recorded result |
|---|---|
| Unit stack | 82.9% |
| Consolidated module | 83.2% |
| Frozen | 78.3% |
| Decisions equal to the stack | 96.8%; gate 98% |
| Outside the units’ scopes | 1,471 configurations bit-identical to frozen |
| Size | 609 values against 915 |
Comparative standing and locked gates
Fidelity failed at 96.8% against 98%. Accuracy within half a point of the stack, unchanged answers outside the scopes, fewer values, fresh-process restoration from the tree, and lineage with members archived rather than deleted all passed.
Interpretation and limitations
Distillation smoothed the boundaries between the units’ sub-regions, which is where the decisions differ; on balance those changes are improvements. Consolidation is built and in use, and the earlier states stay restorable. One distillation, one family of tasks.
SOURCE PROVENANCE
V94: consolidating units learned in operation into the library weight tree
LABORATORY REPORT / 2026-10-21SOURCE CHECKSUM / SHA-256
9545841ad7863a8c536f8867e65a3551c049435e232a1edc77be16da46525b7aPublic 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.