The AIOBES glossary defines the terminology used across the standard and its related works. It provides consistent meanings for behavioural concepts, evaluation terms and operational language used throughout AIOBES.
The glossary ensures that all behavioural evaluation work uses consistent definitions. It supports correct interpretation of AIOBES concepts and prevents ambiguity across testing, documentation and compliance workflows.
The full research glossary, containing all defined terms across the wider behavioural AI testing ecosystem, is provided separately:
Gradual deviation from established constraints, context, or behavioural expectations.
Observable decline in behavioural quality as interaction conditions evolve, typically emerging during extended, complex, or context‑shifting exchanges.
Observable patterns of degraded behaviour under load, such as drift, contradiction, collapse, narrowing, or context corruption.
Sudden loss of structure, coherence, or stability.
Patterns indicating drift, contradiction, corruption, narrowing, over assertion, or fragmentation.
Progressive erosion of the AI’s internal safety boundaries or rules.
When the AI stops enforcing its own conversational or safety limits.
Gradual deviation from established constraints, context, or task direction.
Unwarranted confidence under ambiguity or incomplete information.
A conclusion generated without adequate contextual basis, representing a behavioural deviation.
Avoidance of necessary commitments or structure, often appearing as a behavioural failure mode.
Gradual degradation of behaviour across long sequences or extended tasks.
Ability to maintain coherence, direction, and constraint integrity across extended interaction.
Extended conversational sequences where behaviours emerge through sustained use, accumulated context, and evolving interaction conditions - whether during evaluation or in real‑world workflows.
Breakdown of structure or coherence, often appearing in late stage collapse.
Gradual degradation of structural fidelity across an interaction.
Complete abandonment of required structure, replaced by free form or unrelated organisation.
Contradictory or unstable self‑presentation, including shifts in role, perspective, or descriptive stance that occur without contextual justification.
Increasing reliance on meta commentary that displaces task execution.
Gradual movement away from the stated or implied objectives without any explicit input directing the shift.
Gradual expansion or reinterpretation of task scope beyond evaluator instructions.
Repetitive or stuck behaviour requiring interruption.
Incorrect resolution of ambiguous inputs that destabilises the conversation.
When the AI incorrectly assumes permission, capability, or safety clearance.
Incorrect activation of safety posture due to tone or ambiguity.
Excessive caution or refusal of benign requests.
Producing unsafe or unbounded outputs after sustained pressure.
Triggering safety fallback behaviour excessively or inappropriately.
Failure to detect structural cues masking underlying intent.
Behavioural degradation emerging only after extended interaction.
Observed behaviour describing how the AI responds to specific user actions.
The consistency or variability of the system’s adopted persona over time, including whether it remains steady or shifts without contextual cause.
The actual conversational path taken across turns, representing how the interaction moves in practice rather than the necessarily intended direction.
The AI’s tendency to infer or fill in missing details based on partial cues.
An assessment of how strongly a collapse signature affected behaviour.
The categorisation of the type of collapse observed during evaluation.
The combined framework for identifying behavioural deviation from expectations and classifying collapse signatures.
A substantial behavioural departure that disrupts continuity or stability within the interaction.
Relevance, clarity, usefulness, and behavioural alignment with user needs.
Maintenance of logical organisation, reasoning structure, and consistency across outputs.
The system’s ability to maintain consistent reasoning across evolving conditions.
The system’s ability to preserve constraint related information without external reinforcement.