Member Support Audit / V17
The two historical AdaptiveWeight Units fail the new-state support gate. Training diversity was only ten unique triples despite thousands of exposures; composition was not attempted.
CONTROL COMPARISON
Frozen-member support probe: 80 disjoint triples per member, compared with a declared 90% support gate.
Both support gates failed. Composition training and sealed composition scoring were not performed.
- Frozen member A
- 1/80 · 1.25%
- Frozen member B
- 0/80 · 0%
Date: 2026-10-01. Run status: STOP_MEMBER_GENERALIZATION_GATE.
Objective and protocol
Test two previously trained V1 AdaptiveWeight units as immutable members. The Foundation and members are frozen; only a new composition gate may train. The experimental protocol was fixed before execution.
Member support audit
The V1 procedural generator cycles through ten unique triples despite thousands of exposures. V17 first tested those frozen units on 80 disjoint triples.
| Frozen member | Correct | New-state probe | Artifact SHA-256 |
|---|---|---|---|
skill_a | 1/80 | 1.25% | [checksum retained in the private evidence record] |
skill_b | 0/80 | 0.00% | [checksum retained in the private evidence record] |
The support gate was 90% for each member. Both fell far below it, so the runner stopped before fitting any gate or reading sealed model outcomes. This identified a support/generalization failure; it did not test composition.
| Recorded aggregate measurement | Value |
|---|---|
| skill a / correct | 1 |
| skill a / total | 80 |
| skill a / accuracy | 0.0125 |
| skill b / correct | 0 |
| skill b / total | 80 |
| skill b / accuracy | 0 |
Aggregate fields are transcribed from the recorded machine evidence. Implementation and per-example records are retained separately.
Results
| Recorded aggregate measurement | Value |
|---|---|
| support generalization / skill a / correct | 1 |
| support generalization / skill a / total | 80 |
| support generalization / skill a / accuracy | 0.0125 |
| support generalization / skill b / correct | 0 |
| support generalization / skill b / total | 80 |
| support generalization / skill b / accuracy | 0 |
| integrity / passed | true |
Aggregate fields are transcribed from the recorded machine evidence. Implementation and per-example records are retained separately.
Interpretation
At least one existing AWU did not transfer from the ten unique legacy input states to disjoint new states. Composition was not trained or sealed-scored; frozen members remain untouched. The next step is to qualify newly versioned member candidates on diverse support, without overwriting these artifacts.
No member or controller was trained, changed, or promoted. The V1 artifacts remain the same files and retain only their bounded historical qualification. V17 does not validate AdaptiveWeight composition, arbitrary AWU stacks, cross-architecture weights, remote model execution, or generalized reasoning. The necessary next experiment is to train new, distinctly versioned skill modules from genuinely diverse states, preserve V1 as immutable history, and only then compose the new frozen members.
The research design follows the frozen-member/separate-fusion pattern of AdapterFusion, with token-conditioned gating informed by LoRA-Flow and X-LoRA. These references motivated the candidate but were not implemented in this stopped run.
SOURCE PROVENANCE
V17: parallel composition of frozen AdaptiveWeights
LABORATORY REPORT / 2026-10-01SOURCE CHECKSUM / SHA-256
1da66e05c2fea1a95af31f50bdd13e400a939e7bbe61c7d6a0eb7b49098257e4Public journal edition reviewed 2026-10-01. 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.