AI Safety
7 articles
Reinforcement Learning from Verifiable Feedback: Teaching Language Models to Say "I Don't Know"
Post-training a single 1.5B code model across three verifiably distinct behaviours — including declining what the facts do not support — using reinforcement learning from verifiable feedback.
Prompts Are Specifications: Verifying Commitment Fidelity in LLM Outputs
Compilers, linters and tests validate code without ever consulting the prompt: prompts are specifications, and commitment fidelity — that the code still does what the prompt asked — must be verified directly.
Constructibility Under Test: The Ω Pair as a Predictor of Operational Failure
The (precision, recall) pair of a policy's claim/refuse behaviour against a pinned formal authority predicts operational failure where scalar deployment metrics cannot — the constructive, testable companion to Ω-Certifiability.
Ω-Certifiability: A Theory of Certified Ignorance
Scalar accuracy cannot separate overclaiming from underdelivering: a theory of certified ignorance, where a system's knowledge is measured by what it is certified to decline, not by what it answers.
Monotone Guardrails: A Learned Veto That Cannot Loosen Agent Permissions
The execution boundary for an agent tool call should be a discrete, auditable veto rather than a probability: monotone guardrails that can tighten permissions but never loosen them.
Fail-Closed LLM: Zero False Certificates for Language-Model Claims
An automated certification engine for language-model claims that fails closed: structural verification with zero false certificates, replacing LLM judges, heuristics and pass rates.
Knowing When Not to Act: A Calibrated Safety-Decision Head for Tool-Using Agents
Why classifier scores are the wrong contract for deciding if an action is safe to run without a human — a calibrated safety-decision head for tool-using agents, measured on the ways decision models actually fail.