EVIDENCELOCK — state-only LoRA

This research adapter studies four evidence states: SUPPORTED, MISSING_EVIDENCE, CONFLICT and OUT_OF_SCOPE. It is a state classifier evaluated through next-token option scores; complete JSON response drafting and citation generation are a separate task.

Training and decision protocol

The adapter completed 864 genuine optimizer updates using 864 curated exposures from 99 original synthetic training cases and 396 distinct case/option-rotation prompts. LoRA rank is 4, alpha 8, dropout 0, with q_proj and v_proj targets. The frozen state protocol uses four option rotations, scores the next-token options, takes a strict-majority decision and abstains without a strict majority. The source preserves the exact observable-ledger encoding and rotation protocol in gpu/classification_protocol.py, gpu/compact_ledger.py and the frozen scoring driver.

Measured frozen English results

Comparator	Correct state decisions	Abstentions	Conflict recall
Base model with the same state protocol	39/99	29/99	See preserved per-case records
State-only LoRA	77/99	1/99	9/9

The adapter correctly identified 20/27 missing-evidence cases, but called six missing-evidence cases SUPPORTED and abstained on one. It achieved 39/54 supported and 9/9 out-of-scope decisions. These measurements concern 99 synthetic held-out English scenarios, not broad task or language generalization. No held-out result was used for further tuning. State accuracy does not establish drafting, citation or deployment reliability.

The evidence benchmark uses coarse family/template-lineage separation. Development validation was reused during earlier experiments. Original factual-question versus draft-policy ambiguity is disclosed in the research records. The small translation-review subset was owner-approved with model assistance; independent native-speaker validation is not established. No multilingual classification result is asserted for this adapter.

Attribution and artifact terms

Upstream model creator: Cohere Labs / Cohere. Base: CohereLabs/tiny-aya-fire, revision 73500c80fcb6c8bea3e830c21d5d2920249a0e1a. EVIDENCELOCK contributors produced the LoRA adaptations. The original adapter weight bytes are preserved; only the public configuration's base identifier and revision metadata were normalized. Original configurations and exact before/after changes accompany each artifact.

These adapter deltas are distributed for noncommercial research under CC-BY-NC 4.0 with Cohere's acceptable-use addendum and the Cohere Labs Acceptable Use Policy. Retain attribution and modification notices. Attribution does not imply endorsement. Base weights and tokenizer assets are not redistributed; obtain the base separately under its access conditions. Original EVIDENCELOCK source has a separate MIT license.

Reproduction and records

The final Linux training and evaluation stack is Python 3.12.16, torch 2.11.0+cu130, Transformers 5.22.1, PEFT 0.21.2, TRL 1.7.1, Accelerate 1.15.0, bitsandbytes 0.50.2, CuratorKIT 0.1.22, AlignTune 0.1.28 and AuditKIT 1.1.3. Exact dependency locks and protocols are in the source repository; historical environments are recorded in their original reports. Frozen final inference uses FP16 compute, NF4 double quantization and the pinned model revision. Successful loading alone does not reproduce the measured scoring protocol.