The Inter-System Coherence & Integrity Layer (ISCIL) is a containment architecture that detects and dampens environment-level drift in AI-integrated enterprise systems. Introduced by Myriam Ayada (2026).
ISCIL operates at boundaries between systems. It does not inspect individual outputs, gate execution paths, or attempt to eliminate AI outputs' semantic openess.
A healthy immune system does not achieve sterility but maintains homeostasis despite foreign agents. ISCIL’s success metric is not absence of ABOs but the environment’s ability to absorb small deviations without amplification, prevent uncontrolled propagation, and restore alignment within acceptable timeframes.
This reframes AI risk from reactive fault handling to proactive coherence preservation.
ISCIL functions as an immunity layer: monitoring aggregate boundary statistics, detecting abnormality patterns, tolerating normal variation, and applying proportional containment. The below figure illustrates an example of how ISCIL can be implemented for a credit scoring pipeline (see more about Interconnected System Environment).
Step 1: Boundary Telemetry
Monitors aggregate behavioural statistics at corridor boundaries: approval rates, category proportions, escalation rates, feedback magnitudes. No individual output is inspected. Operates without proprietary model internals.
Step 2: Coherence-Risk Scoring
Uses rate-of-change z-scores relative to a rolling baseline. Detects acceleration above normal variability. Seasonal or market-driven level shifts do not trigger alerts (Ayada, 2026).
Step 3: Corridor-Level Containment
Discretisation guardrails: Blind scalar offset before discretisation boundaries, steering outputs away from jump-susceptible thresholds.
Feedback damping: Asymmetric attenuation; corrective signals boosted, amplifying signals dampened.
Step 4: Proportional Response
Containment scales with severity. Applied only within detected CRC. Blast radius expanded only if risk persists. Relaxes autonomously once stability returns.
1. Interface-level: No internal model inspection. Works with any system producing boundary outputs.
2. Non-intrusive: No synchronous gating except during active containment.
3. Graph-native: Acts on ISE corridor structure.
In a simulation that replicates a realistic AI-integrated underwriting pipeline, e.g. four interconnected systems processing loan applications over 1,200 decision cycles, we measured what happens when a single AI output carries just enough ambiguity to pass local checks but shift downstream behaviour. Below the main results:
All excess defaults eliminated (100% recovery)
78 timesteps active intervention (6.5% overhead)
Detection at ~t=50— ~40 timesteps faster than outcome monitoring
Direction-agnostic: detectedboth ABO onset and cessation
To prevent context clash and downstream errors, many MLOps teams attempt to write rigid “semantic contracts”or force strict JSON schemas. This approach attempts to achieve sterility by breaking the pipeline whenever ambiguity occurs.
ISCIL offers an alternative: instead of brittle semantic contracts that destroy the utility of Generative AI, ISCIL acts as a continuous immunity layer, dampening drift dynamically without requiring absolute determinism.
ABO Concept →/research/ambiguity-bearing-outputs
ISE Framework →/research/interconnected-systems-environment
Glossary →/research/glossary
Ayada, M. (2026). Propagation of Ambiguity-Bearing Outputs Across Interconnected Systems Environment. TechRxiv (in review).
Code: github.com/Myr-Aya/ISE_simulator
Archive: Zenodo DOI 10.5281/zenodo.18719967