Claims and argument
Claim, safety case, likelihood term, risk level, confidence adjustment, residual risk and risk acceptance determination, marginal versus absolute risk basis, limitation.
A safety case is the structured argument that a model's risk is acceptably low -- the connective layer between evaluation evidence and a go/no-go decision that the UK AI Safety Institute's own safety-case vocabulary and the 2026 International AI Safety Report both treat as central. This track defines safety case, claim, risk level, residual risk determination, and the marginal-versus-absolute risk basis a claim is measured against, as used by Anthropic, OpenAI, Google DeepMind, Meta, xAI, and NVIDIA.
- LimitationProposedrecord-type
A proposed term for a stated reason a safety evaluation's conclusions may not hold or generalise, such as elicitation strength or evaluation awareness.
- Marginal versus Absolute Risk BasisProposedcontrolled-value
A proposed controlled value naming the baseline, such as a no-AI world or other developers' current models, against which a stated risk is measured.
- Residual Risk and Risk Acceptance DeterminationProposedrecord-type
A proposed record of the risk remaining after mitigations, paired with the decision and decision-maker accepting that residual risk for a stated scope.
- Confidence AdjustmentProposedrecord-type
A proposed term for a recorded move that raises a stated risk level or treats a capability threshold as reached due to uncertainty, with a required rationale.
- Risk LevelProposedcontrolled-value
A proposed term for the overall graded risk assigned to a model, domain, or company activity as a whole, distinct from a capability threshold or level.
- Likelihood TermProposedcontrolled-value
A proposed controlled ladder of ordinal or numeric terms for the probability a harm or claim holds, including an unanchored value for undefined scales.
- Safety Case (Assurance Argument)Proposedrecord-type
A proposed structured, reviewable argument, built from claim records and supporting evidence, that a model's risks are acceptably low in a stated context.
- ClaimProposedrecord-type
A proposed definition of a claim as an atomic assertion in a safety argument, carrying its own evidence and a rule for combining it with sibling claims.
