Member Generalization / V18
Two new AdaptiveWeight Unit candidates trained on diverse support reach only 16/70 and 9/70 development cases. Sealed scoring and promotion are stopped.
Date: 2026-10-01. Status: DEVELOPMENT_MEMBER_GENERALIZATION_GATE_FAILED. No member was promoted.
Purpose
V17 exposed that the historical V1 generator produced only ten unique inputs. V18 created separate v2 candidate modules from new unique states and preserved V1 unchanged. This run does not test composition.
Data and training
Foundation SHA-256: [checksum retained in the private evidence record]. Dataset manifest SHA-256: [checksum retained in the private evidence record]. The train/dev/sealed/composition-reserve partitions exclude old V1 and all V17 states.
Implementation code and detailed machine records are retained separately. This public edition presents the study methods, aggregate results, and qualification boundaries.
Evaluation
| Recorded aggregate measurement | Value |
|---|---|
| development / a foundation only / correct | 0 |
| development / a foundation only / total | 70 |
| development / a foundation only / accuracy | 0 |
| development / a v1 old unit / correct | 1 |
| development / a v1 old unit / total | 70 |
| development / a v1 old unit / accuracy | 0.014285714285714285 |
| development / a v2 candidate / correct | 16 |
| development / a v2 candidate / total | 70 |
| development / a v2 candidate / accuracy | 0.22857142857142856 |
| development / b foundation only / correct | 0 |
| development / b foundation only / total | 70 |
| development / b foundation only / accuracy | 0 |
| development / b v1 old unit / correct | 0 |
| development / b v1 old unit / total | 70 |
| development / b v1 old unit / accuracy | 0 |
| development / b v2 candidate / correct | 9 |
| development / b v2 candidate / total | 70 |
| development / b v2 candidate / accuracy | 0.12857142857142856 |
| development gate / passed | false |
| training generalization diagnostic / members / skill a v2 / training autoregressive exact / correct | 213 |
| training generalization diagnostic / members / skill a v2 / training autoregressive exact / total | 230 |
| training generalization diagnostic / members / skill a v2 / training autoregressive exact / accuracy | 0.9260869565217391 |
| training generalization diagnostic / members / skill a v2 / training teacher forced token accuracy | 0.9753623188405797 |
| training generalization diagnostic / members / skill a v2 / development exact accuracy | 0.22857142857142856 |
| training generalization diagnostic / members / skill b v2 / training autoregressive exact / correct | 160 |
| training generalization diagnostic / members / skill b v2 / training autoregressive exact / total | 230 |
| training generalization diagnostic / members / skill b v2 / training autoregressive exact / accuracy | 0.6956521739130435 |
| training generalization diagnostic / members / skill b v2 / training teacher forced token accuracy | 0.8971014492753623 |
| training generalization diagnostic / members / skill b v2 / development exact accuracy | 0.12857142857142856 |
| sealed gate / passed | false |
| fresh process restore / scores / skill a v2 / correct | 16 |
| fresh process restore / scores / skill a v2 / total | 70 |
| fresh process restore / scores / skill a v2 / accuracy | 0.22857142857142856 |
| fresh process restore / scores / skill b v2 / correct | 9 |
| fresh process restore / scores / skill b v2 / total | 70 |
| fresh process restore / scores / skill b v2 / accuracy | 0.12857142857142856 |
| integrity / passed | true |
Aggregate fields are transcribed from the recorded machine evidence. Implementation and per-example records are retained separately.
Composition and sealed scoring were not justified by the failed member gates. A follow-up would investigate token alignment, activation and gradient flow, block placement, and task encoding before a materially revised v3 training plan with fresh training and development data. V1 and V2 artifacts remain controls; the evidence does not support repeating the unchanged 1,600-step FFN setup.
Runbook: V18 protocol. Machine data, candidate weights, and hashes are in ignored [retained internal evidence].
SOURCE PROVENANCE
V18: AdaptiveWeight member generalization
LABORATORY REPORT / 2026-10-01SOURCE CHECKSUM / SHA-256
80290268e3563e01735a3e96e515737b8de08be469713b145291ae83735f0401Public 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.