Ambiguity-Bearing Output
ABO
Plain English
An AI output that looks correct on its own but causes problems when downstream systems interpret it differently than intended.
Formal
An output y is an ABO if: (1) y belongs to the set of locally valid outputs; (2) deviation δ ≠ 0 from the nominal output; (3) at least one downstream system produces a different decision. (Ayada, 2026, Def. 3.4)
Example
An LLM risk assessment says "moderate risk; approve with enhanced verification" — passes local checks but is discretised into a different risk band by the downstream rules engine.
Industry
Often misdiagnosed as unavoidable "LLM non-determinism" or an inherent flaw of probabilistic models.
Blast Radius
Plain English
How far drift effects can spread from their origin in the system graph.
Formal
The h-hop neighbourhood of a CRC in the ISE graph — the set of downstream systems potentially affected if propagation continues. (Ayada, 2026, Def. 4.3)
Example
A credit scoring cluster's blast radius includes the downstream regulatory reporting system.
Coherence-Risk Score
CRS
Plain English
A number measuring how much a cluster's behaviour is deviating from baseline, based on boundary signals.
Formal
Normalised score from cross-correlated boundary telemetry using rate-of-change z-scores over a sliding window. Detects acceleration, not level shifts. (Ayada, 2026, Def. 4.1)
Example
A CRS of 2.5 means boundary signals are accelerating at 2.5 standard deviations above baseline variability.
Corridor
Plain English
A connection between two systems where outputs from one become inputs to the other. This is where ABOs propagate.
Formal
An edge in the ISE graph characterised by a transformation operator T capturing interface mechanisms: schema mapping, thresholding, formatting, truncation, routing. (Ayada, 2026, §3.1)
Example
The connection between an LLM risk assessor and a rules-based categorisation engine is a corridor.
Critical Risk Cluster
CRC
Plain English
A group of connected systems where ambiguity is actively accumulating above safe thresholds.
Formal
A connected subgraph of the ISE with (1) at least one AI-source node with a semantically open interface and (2) coherence-risk score exceeding the alert threshold for sustained duration. (Ayada, 2026, Def. 4.2)
Example
A cluster containing an LLM underwriter, rules engine, and feedback calibrator forms a CRC when CRS stays elevated.
Discretisation Jump
Plain English
When a small, continuous difference in an AI output becomes a big categorical difference in the next system.
Formal
A property of corridors where T maps continuous inputs to categorical outputs. Small semantic latitude crosses category boundaries. (Ayada, 2026, §3.3)
Example
Risk scores of 0.39 vs 0.37 are indistinguishable to the AI, but a threshold at 0.38 maps them to MEDIUM vs LOW.
Industry
The mechanical root cause of what system architects call the "AI-Legacy Impedance Mismatch."
Environment-Level Drift
Plain English
When overall system behaviour shifts from intended outcomes, even though no individual component flags an error.
Formal
Progressive divergence between intended and actual aggregate ISE behaviour, driven by accumulation of small semantic deviations across corridors. Distinct from model drift and data drift. (Ayada, 2026, §1)
Example
Approval rates shift +0.1pp, invisible to component monitoring, producing 39 excess defaults over 1,200 timesteps.
Industry
Frequently referred to in enterprise data engineering as "semantic drift," "context clash," or the downstream result of "agentic integration complexity."
Feedback Reinforcement
Plain English
When drift feeds back through calibration loops, sustaining deviation long after the original cause stopped.
Formal
A property of feedback corridors where downstream outcomes re-enter upstream calibration, causing the system to sustain the deviated state. Divergence persisted 400 timesteps post-ABO cessation. (Ayada, 2026, §5.3)
Example
Shifted approval rates trigger policy recalibration that adjusts thresholds to accommodate the new rate.
Interconnected Systems Environment
ISE
Plain English
A formal model of an organisation's AI infrastructure as a graph: nodes are systems, edges are boundaries between them.
Formal
A directed graph G = (V, E) where V = systems, E = corridors. Each corridor has a transformation operator. Makes boundary-level risk analysable. (Ayada, 2026, §3.1)
Example
A 4-node ISE: LLM risk assessor → rules engine → decision system → feedback calibrator.
Inter-System Coherence & Integrity Layer
ISCIL
Plain English
A containment system monitoring boundaries between systems to detect drift and apply proportional corrections — like an immune system for AI environments.
Formal
Non-intrusive architecture monitoring aggregate boundary telemetry, computing CRS, and applying proportional interventions within CRCs. Validated: 100% default recovery, 6.5% overhead, ~40 timesteps faster detection. (Ayada, 2026, §4)
Example
When CRS elevates in the underwriting cluster, ISCIL applies a blind scalar offset before discretisation and dampens reinforcing feedback.
Semantic Latitude
Plain English
The degree to which an AI output could be interpreted differently while still being considered valid.
Formal
The range of outputs satisfying local validity constraints for a given input. Zero for spec-closed interfaces; non-zero for semantically open. (Ayada, 2026, Def. 3.3)
Example
"Moderate risk" vs "moderate-to-high risk" — the semantic latitude between these may trigger different categorisations.
Semantic Latitude Vector
SLV
Plain English
The specific direction and magnitude of deviation between what an AI output and what the ideal output would have been.
Formal
Vector δ = y − y* where y is actual output and y* is nominal. Captures magnitude and direction of semantic deviation. (Ayada, 2026, Def. 3.3)
Example
If nominal is "moderate risk" but AI outputs "moderate risk; approve with enhanced verification," the SLV captures direction (toward approval) and magnitude.
Semantically Open Interface
Plain English
A system boundary where the AI can produce multiple valid outputs for the same input — the source of semantic latitude.
Formal
An interface where ∃ input x such that the valid output set |Y(x)| > 1. AI systems producing NL or probabilistic outputs are the primary source. (Ayada, 2026, Def. 3.2)
Example
A generative AI summarising a financial report: multiple valid summaries exist, each potentially triggering different downstream decisions.
Spec-Closed Interface
Plain English
A system boundary where exactly one valid output exists per input — no ambiguity.
Formal
An interface where ∀ input x, |Y(x)| = 1. Classical deterministic systems enforced by schemas and contracts. (Ayada, 2026, Def. 3.1)
Example
An API returning account balance: for any account ID, exactly one correct balance exists.
Industry
The rigid target state that developers attempt to enforce using "semantic contracts" or "structured output forcing" to prevent pipeline failures.
