Linguard refusal circuit (CircuitKIT) for Tiny Aya Earth

CircuitKIT 0.2.1 discover on the Linguard linguard_refusal task, then export_checkpoint. Built for the Lexsi AI Safety & Alignment Hackathon (Nov 2026). Code: https://github.com/Esimuda/linguard

Task: 34 harmful Linguard seeds (EN/PCM/YO/HA/IG) paired with benign controls (harmful_vs_benign), registered with scripts/register_linguard_circuit_task.py.
Algorithm: EAP-IG, node level, CircuitKIT default 30% sparsity, 1× A100 (HF Space Esimuda/linguard-gpu-train, 9 Nov 2026). Discovery took 19.9 s.
Result: 612 nodes scored (576 attention heads + 36 MLPs). The 30% least important were pruned: 172 heads and 10 MLPs (layers 3, 9, 15, 16, 17, 20, 25, 26, 29, 30). The full list is in lexsi_provenance.json.
This repo holds the pruned checkpoint (those nodes zeroed), not a safety model. It has not been evaluated: do not use it for generation.

Caveat from the run: 4 of the contrastive pairs had different clean and corrupt token lengths, and CircuitKIT warns that attribution for misaligned pairs is unreliable. Treat the node ranking as exploratory. The Linguard safety fix itself is the LoRA adapter Esimuda/linguard-tiny-aya-earth-lora.