Guardrails

Risk	Prevention	Detection	Failure response
Arithmetic hallucination	Deterministic evaluation	Golden numeric cases	Reject LLM-derived result
Wrong chemistry model	Frozen equation/species validators	Model-validity cases	Localise and explain first invalid model step
Cascading duplicate errors	Dependency graph	First-error and ECF cases	Separate propagated from independent errors
Invented marks or ECF	Versioned question-specific policy	Unsupported-claim metric	REVIEW_REQUIRED; suppress total
False equivalence	Bounded AST rules	Adversarial expression cases	Require review
Over-revealing hint	Hint-level policy	Hint appropriateness review	Choose lower disclosure or no hint
Unsupported learner label	Evidence-linked signals only	Weakness audit	Do not update weakness
False agent claim	Runtime trace requirement	Trace completeness gate	Describe as workflow
Stale content	Immutable versions and trace references	Version mismatch test	Stop attempt load

Every marking surface displays frozen problem and policy versions, source evidence, and review status.