i-am-shaurya05/shadowwork-shift-learner-corrective

Learner corrective

Role: Frozen selected model. This is a frozen LoRA adapter derived from the custom Plain SFT base, itself fine-tuned from CohereLabs/tiny-aya-global.

Exact identity

• Original audited adapter-manifest digest: 680659a3c6fce48e18f4ce8087b3fbb2c5806d4427e90c2b5cc6c49a524bb441
• Unmodified LoRA safetensors SHA-256: e01c7fd74b6fd627ece63cf170a3c60349723a0b59500b90964a5cf84509d6f1
• Public adapter_config.json changes only base_model_name_or_path from the local Kaggle path to i-am-shaurya05/shadowwork-shift-plain-sft. Original config and manifest are preserved under provenance/.
• Quantized inference: NF4, double quantization, FP16 compute, greedy decoding, up to 112 new tokens per next action, maximum 14 tool actions.
• Requires the original ShadowWork simulator, tool schemas and authorization/contract/completion gates for meaningful workflow evaluation.

Evidence and limitations

Development (56 cases): frozen parent 20/56, rehearsal control 24/56, learner-corrective 56/56, engineering early-finish 54/56. Independent deterministic simulator holdout (70 cases): parent 42/70, learner-corrective 70/70, engineering controller 69/70. These scores are on a synthetic, narrow SQLite workflow simulator; they are not general real-world reliability claims. The rule-based controller nearly matches the learned model.

License

Derived from Cohere Labs Tiny Aya, released under CC-BY-NC and subject to Cohere's Acceptable Use Policy. Attribution to Cohere Labs is required. Non-commercial use only under the upstream terms; verify the original model card for the complete conditions.