Date: 2026-10-05. Status: REJECTED_DEVELOPMENT_GATES. No active promotion.

Question and changed mechanism

V56 qualified selective import and native inference, not learning. This pilot asks whether a separately owned linear readout can exploit its frozen full-document scientific representation. The native classifier remains untouched. Only the new 3,075 weight/bias parameters train; feature statistics are fit-only non-trainable buffers.

Results

Locked gates: {"balanced_admitted_80": false, "balanced_all_80": false, "coverage": true, "gain_memory_2": true, "gain_parent_5": true, "raw_high_confidence_wrong_2": false, "retention_95": false}.

The original NLI parent uses a different, shorter input contract and no article title. Comparisons with it are system comparisons, not isolated model-only effects. The native scientific classifier and matched memory receive the same representations/input information as this pilot.

Data, ownership and persistence

362 fit /51 selection /46 calibration cases reuse the verified agent-trace source-component split. Development183 was already open;187 qualification cases remain closed. No gold rationale enters inference. All 642 full-document representations are persisted;51 selection representations are reused rather than recomputed. The frozen in-memory member hashes match before and after new extraction.

Candidate: [retained internal artifact] inside the run directory. Artifact SHA-256: [checksum retained in the private evidence record]. Frozen parent SHA-256: [checksum retained in the private evidence record]. Exact ownership, fit IDs, source/trace hashes, optimizer and calibration outputs are in the adjacent manifest and locked runbook.

Four focused provenance tests pass. The candidate restores from its own artifact/statistics in a fresh process. Private source/data/state/catalog/artifact closure follows this report; that gate must complete before claiming remote recovery.

Primary research and interpretation

MultiVerS supports joint scientific document/evidence representations; official checkpoints distinguish weak scientific preparation from target-fitted models. The imported publisher training overlap remains unknown.

Linear-probing then fine-tuning language-model research and author code motivate separating readout learning from representation updates. V57 is a fixed linear-probe diagnostic, not their full LP-FT reproduction. Pinned source code and hashes are retained under primary-source; it is not executed. The repository has an MIT license, while the inspected trainer retains an Apache-2.0 source header and upstream attribution. No third-party implementation is vendored into runtime.

Calibration research motivates temperature scaling. The author demonstration is unmaintained; this pilot uses the existing independently implemented diagnostic and fixed calibration-only temperature grid. Calibration never replaces the raw-confidence gate.

AdapterFusion supports non-destructive composition, but a categorical router cannot exceed the measured native-classifier union oracle on the same selection cases. A new representation classifier is a distinct hypothesis.

A failed linear probe does not prove that no useful information exists in the encoder. Inspect class-specific errors and rationale quality before selecting query-conditioned sentence-state access or a separately owned contextual module. Do not repeat this unchanged probe with arbitrary extra epochs/seeds. Priority2 and successive distinct acquisitions in priority3 remain unqualified until their complete gates pass.

Locked runbook.