| # LoudInk v2 shadow candidate | |
| ## Status | |
| This published bundle is a shadow candidate. It is not eligible for default rollout until `gate_loudink_v2` passes on the private, decontaminated MLX workload suite. | |
| ## Architecture | |
| - One 4-bit LFM2.5 1.2B base model, with separate writer and compact-IR LoRAs. | |
| - Deterministic versioned IR compiler for supported native apps, Google services, Maps, and generic search. | |
| - Confirmation is required for login-sensitive, destructive, send/submit, purchase, and unsupported workflows. | |
| ## Evaluation reporting | |
| The local release dashboard deliberately separates: direct neural output; model path including parser/repair; and end-to-end compiler-assisted success. Do not describe end-to-end compiler coverage as raw model accuracy. | |
| ## Limits | |
| This is not a screenshot-guided arbitrary-site agent. It supports service workflows plus generic search, and falls back to a clear confirmation for work outside that scope. | |