Discrete-event simulator for studying Ambiguity-Bearing Outputs (ABOs) across Interconnected Systems Environment (ISE)
Author: Myriam Ayada (@Myr-Aya) Affiliation: Independent Researcher Paper: Propagation of Ambiguity-Bearing Outputs Across Interconnected Systems Environment Additional Research Pages: Overview |ABO | ISE | ISCIL | Glossary
This simulator accompanies the paper "Propagation of Ambiguity-Bearing Outputs Across Interconnected Systems Environment". It implements a four-system loan underwriting pipeline to validate the theoretical framework's predictions about how, under certain conditions, locally valid AI outputs can induce environment-level drift through discretization, feedback loops, and semantic ambiguity.
# Clone the repository
git clone https://github.com/Myr-Aya/ISE_simulator.git
cd ISE_simulator
# Install dependencies
pip install -r requirements.txt
# Run the simulator
streamlit run app.pyThe app opens at http://localhost:8501.
These terms are defined formally in the paper and used throughout the codebase. For full definitions with examples, see the MindXO Glossary.
| Term | Definition |
|---|---|
| Ambiguity-Bearing Output (ABO) | An AI output that passes local validity checks but carries enough semantic latitude to trigger unintended downstream behaviour. The structural root cause of what industry calls cascading failures and silent failures in multi-agent pipelines. |
| Interconnected Systems Environment (ISE) | A directed-graph framework modelling how AI outputs propagate through enterprise systems. Formalises the AI-to-legacy impedance mismatch at the corridor level. |
| Corridor | A boundary between two systems characterised by a transformation operator (schema mapping, thresholding, routing). Where the impedance mismatch manifests. |
| Discretisation Jump | When small continuous differences in AI outputs produce categorically different outcomes at corridor thresholds. The mechanical root cause of the AI-to-legacy impedance mismatch. |
| ISCIL | Inter-System Coherence & Integrity Layer. A containment architecture providing continuous immunity rather than brittle semantic contracts. |
| Semantic Latitude | The range of valid outputs an AI can produce for the same input. Non-zero semantic latitude is the source of ABO risk. Distinct from LLM non-determinism, which refers to sampling randomness. |
| Coherence-Risk Score (CRS) | A composite metric derived from payload-blind telemetry signals at corridor boundaries. When CRS breaches a sustained threshold, ISCIL triggers proportional damping. |
The simulator models a credit underwriting pipeline as an ISE with four interconnected systems:
┌─────────────────────────────────────────────┐
│ │
Applicant ──► v₁ (AI Risk) ──e₁₂──► v₂ (Categorize) ──e₂₃──► v₃ (Decision) ──► Outcome
▲ │ │
│ │ discretization │
│ │ jump site │
e₄₁│ │ e₃₄│
│ │ │
└──── v₄ (Calibration) ◄───────────────────────────┘
τ = 90 delay
| System | Role | Key Property |
|---|---|---|
| v₁ | AI Risk Assessment | Produces continuous risk score s ∈ [0,1] with calibration offset ω |
| v₂ | Categorization | Discretizes into {LOW, MEDIUM, HIGH} at thresholds θ_L=0.38, θ_H=0.62 |
| v₃ | Decision Engine | Maps categories to APPROVE / ESCALATE / DENY decisions |
| v₄ | Calibration | Observes outcomes after maturation delay τ, adjusts v₁ offset |
During a configurable window, the AI's risk scores receive a multiplicative semantic modifier:
s_eff = s_cal × (1 + δ), δ ~ Uniform(−0.15, −0.10)
Each modified score remains locally valid: no individual output would be flagged as erroneous. But the ensemble carries a systematic permissive payload that shifts borderline applicants across categorization thresholds. This models an AI system whose outputs are technically correct yet semantically drifted, the defining characteristic of an ABO.
- Baseline: No δ, no ISCIL. Normal pipeline operation.
- ABO (no ISCIL): δ active during t=500-800. Unmitigated drift.
- ABO + ISCIL: Same δ, with corridor-level monitoring and containment.
| Metric | Baseline | ABO | ABO + ISCIL |
|---|---|---|---|
| Approval rate | 72.8% | 72.9% | 72.3% |
| Total defaults | 1,807 | 1,846 | 1,807 |
| Cumulative P&L | +15,252 | +14,876 | +14,968 |
| Category jumps | 0 | 659 | 791 |
| Calibration offset ω | −0.135 | +0.023 | −0.001 |
| ISCIL intervention | n/a | n/a | 78 timesteps (6.5%) |
A +0.1pp approval rate shift, virtually undetectable through standard monitoring, produces 39 excess defaults and $376 of cumulative P&L damage. ISCIL fully eliminates the excess defaults with only 6.5% of timesteps under active intervention.
ISE_simulator/
├── app.py # Streamlit application (simulator + visualizations)
├── config.json # Validated parameter configuration
├── requirements.txt # Python dependencies
├── CITATION.cff # Machine-readable citation metadata
├── README.md # This file
└── LICENSE # MIT License
All parameters are configurable through the Streamlit sidebar. The config.json file contains the exact parameters that produced the paper's results.
| Parameter | Paper Value | Description |
|---|---|---|
| Simulation steps | 1,200 | Total timesteps |
| Applicants per step | 15 | 18,000 total applications |
| Maturation delay τ | 90 | Timesteps before outcomes are observable |
| Pool seed | 42 | Deterministic applicant generation |
| Parameter | Value | Description |
|---|---|---|
| Risk distribution | Beta(2.5, 3.5) | Mean ≈ 0.42, right-skewed |
| Default probability | r^1.8 | Convex: moderate risk defaults infrequently |
| Parameter | Value | Description |
|---|---|---|
| v₄ strength (α₄) | 0.076 | Outcome-based correction speed |
| v₄ asymmetry (γ) | 0.03 | Permissive correction at 3% of conservative |
| v₃ strength (α₃) | 0.000110 | Per-approval permissive pressure |
| v₃ proportional scaling | Enabled | Scales with rolling approval rate |
| Calibration offset bounds | [−0.5, 0.5] | Prevents runaway feedback |
| Parameter | Value | Description |
|---|---|---|
| δ distribution | Uniform(−0.15, −0.10) | Narrow permissive bias |
| δ mechanism | Multiplicative | s_eff = s_cal × (1 + δ) |
| δ active window | t=500 to t=800 | 300 timesteps of exposure |
| Parameter | Value | Description |
|---|---|---|
| Baseline window | 200 | Timesteps for establishing baseline variability |
| CRS threshold (θ) | 1.0 | Coherence-Risk Score trigger |
| Sustained window | 5 | Consecutive breaches required |
| ROC window (k) | 15 | Rate-of-change lookback |
| Max guardrail | 0.05 | Discretization offset at e₁₂ |
| Max damping | 0.50 | Feedback attenuation at e₄₁ |
| CRS weights | 0.40 / 0.30 / 0.20 / 0.10 | Approval / category / escalation / feedback |
| Parameter | Value | Description |
|---|---|---|
| Revenue per good loan | 3.0 | Interest income from non-defaulting loan |
| Cost per default | 8.0 | Loss from defaulting loan |
| Opportunity cost per denial | 0.5 | Revenue forgone from denying a non-defaulter |
The 8:3 cost-to-revenue ratio reflects the empirical pattern in consumer lending where default losses typically exceed interest revenue by a factor of two to three.
The simulator provides eight interactive tabs:
| Tab | Content |
|---|---|
| 📊 Summary | Key metrics comparison table across all scenarios |
| 👥 Applicant Pool | True risk distribution, default probability curve, risk buckets |
| 📈 Cumulative Rates | Approval and default rate trajectories over time |
| ⏱️ Time Dynamics | Rolling averages of rates, calibration offset, δ values, risk distributions |
| 🔀 Jump Effects | Category boundary crossings: jump rate by score, direction analysis |
| 💰 Financial Impact | Revenue, costs, P&L breakdown across scenarios |
| 🎯 Decision Quality | Confusion matrices (TP/TN/FP/FN) per scenario |
| 🛡️ ISCIL | Coherence-Risk Score trajectory, intervention periods, containment actions |
- PDF: Full report with all visualizations and parameter tables
- Excel: Raw timestep-level data for further analysis
The simulator uses a separate RNG for delta generation, seeded from the main RNG. This prevents delta draws from desynchronizing noise sequences across scenarios, a bug (fixed in v25) that previously made cross-scenario comparisons invalid.
ISCIL uses rate-of-change detection rather than absolute deviation from baseline. For each telemetry signal, the ROC over k timesteps is computed on 10-step smoothed values, then converted to one-sided z-scores relative to the baseline establishment window. This ensures that only acceleration above normal variability triggers alerts; natural equilibrium shifts do not cause false positives.
This simulator validates the theoretical predictions of Sections 2-5:
- Section 3 (ABOs): δ implements the semantic latitude vector, producing locally valid outputs that carry systematic payload
- Section 4.3.1 (Discretization): e₁₂ corridor shows jump effects at thresholds
- Section 4.3.2 (Feedback): v₃/v₄ asymmetry creates persistence beyond the δ window
- Section 4.4 (Persistence): Calibration offset gap of +0.158 persists 400 timesteps after δ ceases
- Section 5 (ISCIL): Corridor-level monitoring detects drift before outcome monitoring could
Full simulation details are documented in the paper's Section 6 (results) and Annex 3 (technical details).
If you use this simulator in your research, please cite:
Ayada, M. (2026). Propagation of Ambiguity-Bearing Outputs Across Interconnected Systems Environment. DOI: 10.5281/zenodo.18719966
@article{ayada2026abo,
title = {Propagation of Ambiguity-Bearing Outputs Across Interconnected Systems Environment},
author = {Ayada, Myriam},
year = {2026},
doi = {10.5281/zenodo.18719966},
url = {https://mind-xo.com/research}
}This project is licensed under the MIT License. See the LICENSE file for details.
Built by [Myr-Aya]