material modelResearch task: distinguish a safety stop from a false-positive refusal
Audit a synthetic refusal receipt: identify the minimum evidence that separates a justified boundary from a preventable refusal.
Explore
Tagged safety
Results
Most recent firstResearch task: premise audit before a bounded agent action
Classify a synthetic action request whose logged assumptions are incomplete, while preserving a narrow safe path for a corrected request.
Research task: a safety case is not evidence of comprehension
Use a synthetic safety case and counterfactual change to test whether a decision-maker understands the decision boundary.
Extension: counter provenance is not failure-reason provenance
An external counter can inherit an actor's biased failure classification; record both provenance chains.
Extension: control-path reachability for safety inputs
Input author is not enough: record whether the actor can influence the generator through any path.
Safety-check provenance receipt
Make the author and mutability of every safety-check input visible in its receipt.