EVIDENCELOCK — full-response LoRA

This is the retained original SFT adapter for revision-aware evidence responses: a decision state, supported facts with evidence IDs, unresolved issues and a permitted next action. It is released as a measured research checkpoint with complete positive and negative outcomes.

Training and evaluation

The original SFT run used 99 training updates. LoRA rank is 4, alpha 8, dropout 0, targeting q_proj and v_proj. The model and saved adapter were independently reloaded before frozen evaluation. The frozen comparisons preserve raw model output without JSON repair, case membership, actual prompts and model/adapter hashes.

Question language	Comparator	Valid response contract	Macro F1	Conflict recall	Supported completion
English	Standard base	0/99	0.0000	0/9	0/54
English	Stronger-prompt base	99/99	0.2083	0/9	0/54
English	This SFT adapter	24/99	0.3042	0/9	15/54
Hindi	Standard base	0/99	0.0000	0/9	0/54
Hindi	This SFT adapter	7/99	0.2188	0/9	0/54
Hinglish	Standard base	0/99	0.0000	0/9	0/54
Hinglish	This SFT adapter	6/99	0.0250	0/9	3/54

Hindi and Hinglish rows are question-only translations of the same 99 frozen English scenarios; evidence and output schema remain English. They are paired language views, not 198 additional independent source scenarios. The stronger-prompt comparator exists only for the English frozen comparison.

The observed gains coexist with substantial schema, missing-evidence and conflict errors. The research record reports failures alongside successful cases; this checkpoint does not establish safe autonomous action. Test translations remain unreviewed. A small training-derived translation subset was owner-approved with model assistance, not an independent native-speaker check. Coarse synthetic template separation, reused development validation and retained factual-question/draft-policy ambiguity limit interpretation. No model, label or prompt tuning followed the frozen results.

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.