Date: 2026-10-05. Imported external research member; no EMMA weights trained, active promotion or semantic-learning qualification.

The non-executing V55 metadata interpreter identified544model tensors totaling1,751,085,088 bytes; the other3,502,683,862checkpoint bytes were not needed for model execution. V56 selectively retrieved the model storages with source ETag, exact range/byte checks and ZIP CRC verification; each storage and the resulting safetensors artifact have recorded SHA-256 identities. No provider pickle or provider Python model code executed.

Derived model artifact SHA-256: [checksum retained in the private evidence record]. Import integrity: 199/199prior research artifacts unchanged.

Frozen inference evidence

Existing open selection partition: 51/51complete inputs, balanced accuracy 63.49%. Elapsed mount/inspection/inference time 32.3s; peak allocatedGPU1930621440 bytes. Fresh-process maximum probability difference: 0.0; all frozen in-memory tensors match before/after.

Architecture:24Longformer layers,1024hidden,16attention heads,50,275token embeddings, independent stance/rationale heads. Pinned provider tokenizer/config and ten additional tokens are checked. Native REFUTES/NEI/SUPPORTS indices are explicitly converted to EMMA CONTRADICT/SUPPORT/NO_EVIDENCE. Claim/special tokens use global attention; every sentence retains its source index. Overlength input refuses rather than silently truncating.

Limits and next decision

This is an imported frozen capability reference, not an acquired EMMA capability. Provider metadata is consistent with stage-one FEVER/PubMedQA/Evidence Inference preparation but does not independently prove training history or corpus disjointness. Selective transfer removes unused training-checkpoint content; all 24trained Transformer layers still execute. This is not a demonstration of sparse computation or arbitrary layer skipping. The checkpoint retains a scifact_n_train configuration field even though its datasets field excludes SciFact; neither field alone is proof of target training or target exclusion. The full publisher checkpoint SHA is unknown; the derived model artifact is independently hashed and privately recoverable.

The51selection cases were already opened and are diagnostic. The187qualification pairs remain unscored. Provider input includes article title, which prior NLI controls did not use; this is not a model-only causal comparison. No annotations or reference rationales enter inference. An accuracy result alone does not satisfy semantic acquisition, retention, calibrated confidence, memory advantage or successive-learning gates.

Initial file staging exposed slow CRT zero expansion on the exFAT volume. The exact owned process was stopped before tensor transfer; the original header/zero-only body was validated, the protected-before inventory preserved and native file allocation substituted. Failed sparse marking and the correction are documented, not concealed.

Initial strict loading also refused the legacy encoder.embeddings.position_ids buffer. The installed backend no longer declares it; an explicit persistent compatibility buffer now accepts the original tensor and validates it against the deterministic position sequence. Strict loading remains enabled. The failed attempt is preserved and model tensor values were not changed.

The next learning mechanism must be justified by this frozen member’s compatibility, cost and headroom and must train only a separately owned child. Existing Foundation, parent and qualified modules remain protected. No unchanged V53 repeat, relaxed gates or claim that imported weights constitute learning.

Locked import runbook, V55 contract, MultiVerS primary paper.

Standardized private source/data/state/catalog/artifact closure follows this report. The overall goal remains active.