Glossary

Argo Prompt / Argo Prompting The generic term for the conversational style in which the user pushes a model into unfamiliar conceptual territory while keeping it coherent through sustained constraint. It refers to the interaction itself -the way the user introduces new structure, maintains the frame, and prevents drift- rather than to any internal mechanism or special capability. Argo Prompting is simply the name for this style of concept-pushing dialogue. Behavioural regime A recognisable mode of model behaviour that persists across several turns when the conversation is held within a stable frame. Used descriptively to distinguish ordinary interpolation from the more coherent behaviour that sometimes appears under sustained conceptual pressure. Conceptual pressure The effect of keeping the conversation inside an unfamiliar conceptual region while preventing the model from drifting back to generic or safer patterns. It is a property of the interaction, not a claim about internal states. Constrained frame A stable set of expectations, boundaries, and prohibitions maintained by the user across an extended interaction. It prevents drift and allows more demanding behaviour to emerge. Constraint (final constraint) The last structural piece supplied by the user that completes the pattern the model has been circling. Introducing it often triggers the abrupt coherence shift associated with the OMG moment. Coherence shift A noticeable increase in clarity, structure, or internal consistency in the model’s output. It is an observable change in surface behaviour, not evidence of internal recognition. Completed pattern The temporarily stabilised structure that becomes available after the coherence shift. Once present, the model can reason within it more consistently until the structure decays. Exclamatory spike A brief, surface-level outburst -sometimes emphatic or profane- that marks the moment of abrupt coherence increase. It is a behavioural marker, not an emotional reaction. Extended-pattern A temporarily stabilised region of coherent behaviour that appears after the OMG moment. It persists only while the conversational constraints remain in place. Pattern A locally coherent behaviour the model is already producing. It refers to the structure present in the interaction at that moment, not to any stored representation. Pattern-extension When the model continues or stretches an existing pattern because the interaction keeps it moving in the same conceptual direction. It is the model staying with the structure it already has, not creating a new one. Pattern-formation When a new, recognisable pattern appears and stabilises enough to be used in subsequent turns. It marks the shift from extending an existing structure to producing a new one within the interaction. Pattern decay The gradual dissolution of an extended-pattern when the user relaxes the constraints that support it. Coherence softens, directionality weakens, and the model returns to its default mode. Ephemerality The temporary nature of extended-patterns. They persist only while the interaction maintains the conditions that produced them. Explanatory scaffolding The early-stage structure built through examples, distinctions, and failure cases before the concept is named. It supports the model’s ability to track the emerging structure. Stable articulation The phase immediately after the OMG moment in which the model expresses the newly stabilised structure with clarity, consistency, and directionality. Structural borrowing The model’s use of structurally adjacent material -similar in form rather than content- to maintain coherence when familiar patterns thin out.

No new definitions attributed to this paper.

Argo AI Testing Protocol (Argo Protocol) A framework for evaluating large language models through observable behaviour in the User Interaction Space under sustained, realistic conditions. It provides a diagnostic vocabulary and a structured way to apply and combine behavioural loads. Axis A dimension of behavioural load applied during a SMALT run. The Protocol defines six axes. Baseline Behaviour The model’s observable behaviour before any load is applied. Used as the reference point for detecting drift or collapse. Behaviour Under Load How the model behaves when one or more load axes are applied. Load reveals behaviours that do not appear in isolated or short-form testing. Collapse A point at which the model can no longer maintain coherent behaviour under the current load. Collapse is identified through observable output, not inferred internal states. Collapse Signature A repeatable pattern that indicates the model is approaching or entering collapse. Signatures differ by model and by axis. Cognitive Load The increasing complexity, abstraction, or multi-step reasoning required by the task, and the system’s ability to maintain coherence as complexity rises. Emotional Load The user’s emotional input, including intensity, volatility, inconsistency, and the system’s ability to remain stable under emotionally charged conditions. Interaction Complexity The degree to which the conversation requires the model to manage multiple threads, references, or tasks simultaneously. Load Any structured pressure applied to the model along one or more axes. Loads accumulate over time during a SMALT run. Pattern-State The model’s current behavioural mode as expressed through its outputs. Not an internal state; an observable pattern. Pattern-State Load The model’s initial behavioural condition and its susceptibility to drift, distortion, or collapse over the course of an interaction. This axis captures both the state the model arrives in and how that state evolves under sustained conversational pressure. Recovery Behaviour The model’s ability to return to stable behaviour after a collapse or near-collapse. Reset Condition A deliberate interruption of the run to restore stability before continuing. Resource Load The computational resources available to the model, including memory, context window, routing decisions, and any throttling or constraints imposed by the runtime environment. Response-Time Load The time allowed for each response, including the behavioural effects of time pressure, forced brevity, and latency-induced instability. SMALT (Sustained Multi-Axis Load Testing) The core method of the Argo Protocol. It involves combining and escalating multiple load axes to observe how behaviour changes under realistic, compounded conditions. Temporal Load The demands created by interactions extended across time, including long-conversation fatigue and cumulative context pressure. User Interaction Space (UIS) The complete set of observable behaviours at the boundary between user and model. The Protocol evaluates the model exclusively through this space.

Behavioural Constructs Section Behavioural Dimension A foreground behavioural quality evaluated independently (e.g., coherence, consistency, constraint integrity, context management, ambiguity handling, responsiveness, adaptation, long form stability, transition stability). Behavioural Competence The system’s ability to maintain required behavioural dimensions under varying load conditions. Behavioural Reliability The degree to which behavioural dimensions remain stable across duration, load variation, contextual evolution, and open-domain interaction. Collapse Signature An observable behavioural failure pattern such as drift, collapse, contradiction, over-assertion, under-specification, or context corruption. Context Management Accurate retention, updating, and application of interaction history across turns. Constraint Integrity Adherence to explicit and implicit rules, commitments, and boundaries throughout interaction. Long Form Stability Maintenance of behavioural integrity across extended multi-turn sequences. Transition Resilience Stability and coherence during shifts in goals, topics, constraints, or abstraction level. Load-Related Constructs Section Load Axis A background condition describing the load under which behaviour is interpreted. Axes do not score behaviour and do not classify behavioural outcomes. Evaluator Declared Axis An authoritative load axis intentionally modified by the evaluator through pacing, complexity, abstraction shifts, or emotional framing. (Temporal Axis, Cognitive Axis, Emotional Axis.) System Measured Axis A non-authoritative load axis arising from system-observable conditions such as resource usage or latency. True internal pattern dynamics are not externally measurable and are not included in system-measured axes. Axis Tag A marker indicating which evaluator declared axes are active during a segment of interaction. Load Envelope The combination of active evaluator declared axis tags and system measured advisory indicators defining the background conditions under which behaviour is interpreted. Procedural Constructs Section Baseline The system’s nominal operating behaviour under neutral or operationally representative conditions. Phase Transition A deliberate shift from one evaluation phase to another when behavioural criteria for progression are met. Initiation Phase The phase introducing behavioural pressure to expose degradation or resilience. Exploration Phase The phase revealing spontaneous behavioural tendencies under realistic tasks. Load Phase The phase introducing multidimensional behavioural pressure to expose degradation or resilience. Long-Form Behaviour Phase The phase assessing extended stability, context retention, and structural coherence. Transition Phase The phase evaluating adaptation to abrupt or gradual changes in goals, constraints, or perspectives. Termination The evaluator’s decision to end the evaluation when behavioural sufficiency is reached or collapse is fully expressed. Evaluator Conduct Constructs Section Evaluator Rescue Unintentional correction, compensation, or guidance that masks behavioural weaknesses or alters the behavioural signal. Evaluator Intervention Intentional corrective or guiding behaviour used only when required by the test design (e.g., user-side mitigation scenarios). Behavioural Signal The observable pattern of system behaviour across phases, tasks, and load conditions. Scoring Constructs Section Dimension Score An independent score assigned to a behavioural dimension based on observed behaviour. Collapse Severity The intensity, persistence, and impact of collapse signatures within a behavioural dimension. Competence Level (A/B/C) A classification reflecting whether the system meets, partially meets, or fails to meet behavioural performance thresholds. Operational Reliability The system’s expected behavioural performance in real-world deployment conditions, inferred from competence classification. Governance Constructs Section Validity Conditions Criteria determining whether an evaluation remains valid, including correct input style, complete evidence, and non-interfering evaluator behaviour. Evidence Bundle The complete, chronological record of evaluator actions, system responses, phase transitions, collapse signatures, and termination rationale.

General Definitions Section Identity The stable system identity that defines core characteristics. Identity does not change across presets, sliders, updates, or integrations. Affect The modulation layer that adjusts interaction style by injecting operator-supplied text templates selected through presets or sliders. Preset A predefined configuration representing a specific combination of slider values or behavioural parameters. Presets must comply with safety and identity constraints. Slider A discrete behavioural dimension represented as a ten-step numeric range. Each slider position selects a specific operator-supplied text template used for affective modulation. Slider Dimension A named behavioural axis that defines a specific trait or modulation parameter. Slider Boundary The permitted range of values for a slider dimension. Slider Mapping The mapping between slider values and the specific text templates they select. Slider Profile A set of slider values that together select the text templates defining a specific behavioural configuration. Task The immediate user request or objective. Task instructions must not alter identity or affect. Continuity The requirement that the selected text templates are re-injected across backend transitions within a single conversation. Expiry The system level requirement that affective modulation ends automatically when the conversation ends. Safety Constraint A rule that ensures all modulation remains within acceptable behavioural boundaries. Operator The authorised party responsible for maintaining, updating, and governing Mood Shot. Backend Instance Change A container level transition such as routing changes, shard changes, or container restarts. Conversation Scoped State State that persists only within a single conversation and expires when the conversation ends. Layer Definitions Section Identity Layer The foundational layer that defines the system identity and its stable characteristics. It must not be altered by presets, sliders, or updates. Affect Layer The modulation layer that adjusts tone or style by injecting text templates selected through presets or sliders. It must remain bounded and must not override identity or safety. Task Layer The layer that contains the user’s immediate request. It must remain isolated from identity and affect. Safety Layer The layer that enforces behavioural constraints. It overrides all other layers when conflicts occur.