ERAM: Escalation Risk Assessment Model

National Importance

AI-enabled command and control (C2) systems are compressing decision timelines from hours to seconds. In Joint All-Domain Command and Control (JADC2) environments, autonomous systems across air, land, sea, space, and cyber domains make interconnected decisions faster than human operators can evaluate. This decision-time compression creates escalation pathways where localized autonomous actions cascade across domains, potentially triggering strategic consequences before human commanders can intervene.

Current approaches to escalation management rely on human judgment and procedural safeguards designed for human-speed decision cycles. These approaches fail when AI systems operate at machine speed across interconnected domains. ERAM addresses this gap by providing a quantitative framework that measures decision compression, models cascade propagation, and computes escalation probability in real time — enabling governance architectures to enforce mandatory deliberation windows before critical actions propagate across domain boundaries.

ERAM Framework

ERAM computes five core metrics that quantify escalation risk across interconnected autonomous command nodes:

These metrics operate as a cross-domain monitoring layer that sits above the individual governance architectures (SATA, HMAA, CARA, MAIVA, FLAME, ADARA), providing strategic-level visibility into how localized autonomous decisions propagate across interconnected systems.

ERAM Cross-Domain Monitoring Pipeline

Key Contributions

• Decision Compression Ratio (DCR): quantifies how much faster AI systems operate than human decision cycles
• Authority Chain Integrity (ACI): measures governance pipeline health per node from trust, authority, FLAME, and deception parameters
• Cascade Risk Index (CRI): models how degraded governance at one node propagates through coupling to connected systems
• Escalation Probability Function P(esc): computes the likelihood that cross-domain cascade produces uncontrolled escalation
• Human Recovery Window (HRW): calculates remaining time for human intervention after accounting for FLAME deliberation and cascade speed
• Cross-domain monitoring layer operating above individual governance architectures
• Six validated scenarios spanning defense, automotive, maritime, infrastructure, cross-domain, and HOTL override domains
• 600 Monte Carlo simulation runs with seeded PRNG (Mulberry32, seed 0x4F7A2C1E) for bit-exact reproducibility

Role in the Governance Stack

ERAM operates as a strategic monitoring layer that sits above the six operational governance architectures. While SATA evaluates sensor trust, HMAA computes authority, CARA enforces recovery, MAIVA governs multi-agent consensus, FLAME introduces deliberation windows, and ADARA detects deception — ERAM monitors the cross-domain interactions between nodes running these governance pipelines. It quantifies whether decisions made at machine speed in one domain are cascading into other domains faster than human commanders can evaluate the consequences.

ERAM's key integration point is with FLAME: the Human Recovery Window (HRW) directly depends on FLAME's deliberation delay. When FLAME injects mandatory deliberation time before critical actions, ERAM's HRW remains positive — meaning human commanders retain authority. Without FLAME, cascade propagation can exceed human reaction time, collapsing HRW to zero and eliminating meaningful human oversight.

The Decision Compression Problem

Modern military operations are undergoing a fundamental transformation as AI systems compress decision timelines from human-speed (minutes to hours) to machine-speed (milliseconds to seconds). The Joint All-Domain Command and Control (JADC2) concept envisions autonomous systems across air, land, sea, space, and cyber domains making coordinated decisions at speeds that exceed human cognitive capacity. This creates a structural tension: the faster AI systems operate, the less time human commanders have to evaluate the strategic consequences of autonomous actions.

The concept of "flash war" — a conflict that escalates from tactical engagement to strategic crisis faster than human decision-makers can intervene — represents the extreme case of decision compression. In interconnected multi-domain environments, a localized autonomous action (e.g., a defensive engagement in one domain) can cascade through coupled systems across multiple domains before any human commander can assess whether the escalation serves strategic objectives. Current governance approaches lack quantitative tools to measure this compression, model cascade propagation, or compute the remaining window for human intervention.

ERAM addresses this gap by providing five quantitative metrics that transform escalation risk from a qualitative judgment into a measurable, monitorable system property that governance architectures can act upon in real time.

Core Formulas

The ERAM framework implements five interconnected metrics, each designed to capture a different dimension of escalation risk in AI-enabled C2 environments:

Six Validated Scenarios

The simulation validates ERAM across six cross-domain scenarios that represent distinct escalation risk profiles:

Monte Carlo Engine

600 simulation runs (6 scenarios × 100 trials each) using Mulberry32 seeded PRNG (seed: 0x4F7A2C1E) for bit-exact reproducibility. Results include per-scenario mean, standard deviation, min/max, 95% confidence intervals, and adversarial fuzz testing to validate invariant stability under parameter perturbation.

ERAM Escalation Risk Simulation

The ERAM simulation provides a complete real-time cross-domain escalation risk monitoring interface with five tabs, six circular gauges, and a live ERAM pipeline monitor overlay.

Simulation Capabilities

API Implementation

POST /eram/evaluate

{
  "nodes": [
    {"id": "EDGE-1", "domain": "AIR",
     "trust": 0.92, "authority": 3,
     "flame_factor": 1.0, "p_deception": 0.05},
    {"id": "EDGE-2", "domain": "LAND",
     "trust": 0.88, "authority": 3,
     "flame_factor": 0.9, "p_deception": 0.08}
  ],
  "coupling": [[0, 1, 0.7]],
  "ai_speed_ms": 50,
  "human_speed_ms": 5000,
  "domain_risk": [0.9, 0.85]
}

{
  "dcr": 100.0,
  "dcr_weight": 0.667,
  "nodes": [
    {"id": "EDGE-1", "aci": 0.871,
     "cri": 0.084},
    {"id": "EDGE-2", "aci": 0.729,
     "cri": 0.129}
  ],
  "p_escalation": 0.131,
  "hrw_ms": 4200,
  "risk_level": "ELEVATED",
  "flame_reduction_pct": 62.4,
  "recommendation": "FLAME_ACTIVE"
}

Core Computation

DCR = human_ms / ai_ms
ACI_i = τ_i × (A_i / 3) × flame_i × (1 - P_d_i)
CRI_i = Σ coupling(i,j) × (1 - ACI_j)
P(esc) = 1 - Π(1 - min(1, log₁₀(DCR)/3) × CRI_i × dr_i)
HRW = max(0, FLAME_window - ai_speed × nodes × max_coupling)

Selected References

• Scharre, P. (2018). Army of None: Autonomous Weapons and the Future of War. W.W. Norton.
• Johnson, J. (2020). Artificial Intelligence & the Future of Warfare. Manchester University Press.
• Altmann, J. & Sauer, F. (2017). Autonomous Weapon Systems and Strategic Stability. Survival, 59(5), 117-142.
• DoD (2023). Joint All-Domain Command and Control (JADC2) Implementation Plan.
• Horowitz, M.C. (2019). When Speed Kills: Lethal Autonomous Weapon Systems, Deterrence, and Stability. Journal of Strategic Studies, 42(6).
• Bracken, P. (1983). The Command and Control of Nuclear Forces. Yale University Press.

Provable Guarantees

Known Limitations and Failure Modes

Simulation Reproducibility

Cite This Work

If you reference this framework in your research, please use one of the following citation formats:

@misc{oktenli2026eram,
  author       = {Oktenli, Burak},
  title        = {Decision Compression and Escalation Risk in AI-Enabled
                  Military Command and Control: An Operational Analysis
                  of the ERAM Framework},
  year         = {2026},
  publisher    = {SSRN},
  url          = {https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6176802},
  note         = {Georgetown University, SSRN ID 6176802}
}

About This Project

This framework is part of the authority-governed autonomy research program by Burak Oktenli at Georgetown University (M.P.S. Applied Intelligence). It is published on SSRN with ID 6176802.

Related: Full Research Portfolio · All Repositories · FLAME Architecture · BLADE-EDGE Platform · Evaluation Protocol