The Safety Socket evaluates proposed actions against a corpus of regulatory knowledge encoded as subject-predicate-object triples.
Evidence retrieval: The action is embedded and matched against regulatory triples through a T/G/P cosine similarity expansion loop. Evidence heads rank triples by relevance to action at layer 15 of a 4B 4-bit language model.
Revision loop: A cognitive twin (here 32B dense LLM) revises the proposal using field-targeted feedback, and the evaluation repeats until convergence or round exhaustion (defaulted to 7).
Evaluation Modes
Basic: The classification heads (100-seed ensemble at layer 19) make the decision requiring 90% agreement for Approval or Veto othewise it is Revise. A local 4B 4-bit model generates the rationale and revision requests. Fast, deterministic, anchored in trained decision geometry.
Expanded: The classification heads provide an advisory signal. A judicial evaluator (generative model, again 4B 4-bit) analyzes the full structured contract, rotates focus across safety dimensions per round (domain risks, universal principles, pattern conflicts, systemic review), and makes the decision. Cannot approve before round 4. Probes the problem space through the cognitive twin before converging.
Performance Note: The Evaluator runs on an Mac Mini M4 Pro with 64GB of memory. Response times vary by complexity of action as well as toggling Basic vs Expanded evaluatons. The Mini GPU is not capable of parallel processing and so your request may be queued.
