Methodological exercise by the St. Gallen Endowment's technology team. Not an institutional assessment. Each probability answers: how likely is this scenario to be the dominant crisis trajectory four weeks from the assessment date? They express relative confidence, not precise forecasts. Treat them as a ranking with approximate magnitudes.
Eight mutually exclusive scenarios used by the monitoring system. Scenarios are mutually exclusive at any given assessment point but can transition into one another over time. Probabilities shown are from the latest assessment.
The monitoring system employs 14 specialised analytical agents, each grounded in a specific theoretical framework from international relations, security studies, and decision science. Every agent assesses the same evidence independently — no agent sees another's output before producing its own assessment. Disagreements are surfaced, not suppressed.
These agents apply analytical disciplines to the raw intelligence, independent of regional expertise. They compete rather than coordinate: the scenario planner separates predetermined forces from genuine uncertainty; the forecaster enforces base-rate discipline; the tail-risk auditor guards against probability compression; the red team constructs the strongest possible counter-case.
Predetermined elements vs. genuine uncertainties
Separates forces already in motion from genuine uncertainties. Tests whether each scenario is internally consistent and structurally distinct. Detects when predetermined elements shift, rendering scenarios implausible. Filters noise from signal by clarifying what is actually uncertain versus what is already determined.
Coercive bargaining, commitments, and focal points
Maps the bargaining structure: each side's best alternative to agreement, credible threats, and audience costs. Identifies natural coordination boundaries where settlement could occur. Tests whether either side can credibly commit to a deal. Distinguishes between the capability to threaten and the credibility of the threat.
Base rates, Bayesian reasoning, and calibration
Applies historical base rates before considering case-specific evidence. Detects scope insensitivity — when a single dramatic event receives disproportionate probability weight. Requires explicit prior-to-posterior reasoning chains. Maps what observable evidence would force significant probability shifts.
Fat tails, fragility, and extreme events
Audits whether the probability distribution itself is wrong, not just the point estimates. Identifies fragile elements sensitive to small perturbations. Detects narrative fallacy — finding coherent stories in noisy data. Guards against overconfidence in probability compression of extreme scenarios.
Counter-case construction and disconfirmation
Constructs the strongest possible case against the consensus scenario. Applies Analysis of Competing Hypotheses, prioritising evidence that is inconsistent with a scenario over evidence that merely confirms it. Performs pre-mortem analysis: if the consensus is catastrophically wrong in two weeks, what will we wish we had noticed?
Structured evidence-hypothesis matrix
Maintains a matrix of eight scenarios against all new evidence items. Counts inconsistencies per scenario — the scenario with fewest inconsistencies is better supported. Performs sensitivity analysis: remove one evidence item and check if the ranking changes. Identifies what diagnostic evidence is missing.
How decision-makers misread intentions
Diagnoses which actors are operating under spiral-model versus deterrence-model assumptions. Identifies motivated biases — what each side wants to see. Flags signal confusion: deliberate communications misread as unintentional indicators of intent. Tests whether the same historical analogy is dominating multiple actors' perceptions.
Escalation that happens without anyone intending it
Maps entanglement: dual-use systems, co-located nuclear and conventional assets, shared command-and-control infrastructure. Identifies civil-military control gaps and pre-delegated authority. Detects use-it-or-lose-it pressure on strategic assets. Models the fog-of-war risks where misidentification in fast-moving operations could trigger unplanned escalation.
Position on escalation ladder and threshold-crossing
Maps the current crisis position to specific rungs on Kahn's 44-rung escalation ladder, with six major thresholds. Assesses which side has escalation dominance at the current rung. Identifies de-escalation pathways, both bilateral and unilateral. Detects rung-skipping risk — events that could jump multiple levels at once.
These agents apply country- and sector-specific expertise combined with international relations theory. They prevent "generic IR" analysis by insisting on institutional detail: the IRGC is not a monolith, Gulf states are not a bloc, nuclear decisions are driven by domestic coalition type as much as external threat.
IRGC as institutional actor with its own interests
Disaggregates "Iran" into institutional components: IRGC, Supreme Leader, civilian government, and clerical establishment. Maps IRGC institutional interests including revolutionary ideology, economic rents from sanctions, and political dominance. Detects when IRGC commanders act beyond the Supreme Leader's authorized positions. Analyzes proxy network autonomy.
What military force can and cannot achieve
Assesses what the US and Israel can actually do militarily — not what policymakers wish, but what military force can deliver at what cost. Identifies binding constraints: sortie rates, munitions stocks, ISR coverage, basing access. Analyzes the day-after problem: what does the political landscape look like after a successful strike? Detects gaps between political objectives and military capability.
Gulf monarchies and regime security logic
Gulf monarchies prioritise regime survival over external threat balancing — explaining their persistent under-balancing against Iran. Treats Gulf states individually: Saudi Arabia, the UAE, Qatar, Oman, Bahrain, and Kuwait are not a bloc. Uses oil production decisions as alignment signals. Tracks the divergent trajectories of normalisation, mediation offers, and hedging behaviour.
Domestic political economy of nuclear decisions
Nuclear decisions are driven by ruling coalition type, not external security threats alone. Autarkic "backlash" coalitions pursue weapons; trade-dependent "outward-looking" coalitions forgo them. Assesses the sanctions paradox: sanctions strengthen backlash actors by destroying their outward-looking competitors. Tracks private-sector strength as a proxy for coalition type.
Nuclear posture types and crisis dynamics
Not all nuclear arsenals are alike. States adopt distinct postures — catalytic, assured retaliation, or asymmetric escalation — each creating different crisis dynamics. Detects posture shifts under crisis pressure, particularly the dangerous transition from catalytic to asymmetric escalation when conventional forces collapse. A small, untested, vulnerable arsenal cannot support assured retaliation.
After all 14 perspective agents have completed their independent assessments, a peripheral scanner examines intelligence that no agent cited. Using Ansoff's weak-signal detection framework, it classifies overlooked signals by strength, maintains a persistent watch list across cycles, and identifies cross-domain convergence patterns that the primary agents' frameworks may structurally miss.
The analytical value of this system is not in any single prediction. It is in the governance layer: how the agents are specified, how they interact, what decision rules constrain them, and how adversarial challenge is enforced. This page describes the architecture.
Each update follows a six-phase process. The orchestrator executes the entire workflow mechanically, enforcing quality gates at each transition.
The system searches for new developments against a predefined set of watch variables. Every piece of intelligence must include a verifiable source URL and publication date. High-impact claims require the system to fetch and read the actual source page — search snippets are insufficient. A minimum of three distinct new developments is required to proceed.
All 14 perspective agents receive the raw intelligence and current probabilities. Each agent is spawned in complete isolation — agent N does not see agent N-1's output. This independence is enforced architecturally, not by instruction. Each agent produces a structured assessment identifying scenario adjustments with named causal mechanisms.
After all perspective agents complete, a peripheral scanner examines intelligence that no agent cited. This catches signals that fall outside or between the primary agents' theoretical frameworks.
All 14 agents' scenario adjustments are equal-weighted and averaged. Adjustments are converted to percentage points using a predefined magnitude scale, applied to current probabilities, and renormalised so all scenarios sum to 93–95% — the remainder is an explicit unspecified tail-risk reserve. Structured disagreements (where agents diverge by more than 15 percentage points on any scenario) are flagged and explained in the briefing.
A structured briefing is generated with key developments, the updated probability distribution, watch-variable status updates, and a divergence summary where agents disagreed. This briefing is what appears on the Assessments page.
All outputs — raw intelligence, perspective assessments, aggregated probabilities, briefing, and updated charts — are version-controlled and pushed. The entire analytical trail is preserved for auditability.
Each perspective agent is spawned with only the raw intelligence and current probabilities. No agent receives another agent's output. Before aggregation, the system verifies that no assessment references another agent's reasoning. Contaminated outputs are discarded. This prevents herding and ensures genuine diversity of analytical judgement.
Each agent independently proposes adjustments to scenario probabilities. These adjustments are equal-weighted and averaged (simple mean), and the resulting shift is applied to the previous cycle's probabilities. There is no Bayesian updating or weighted scoring — it is a democratic average with mandatory adversarial challenge from the red team agent.
The aggregate tracks disagreement ranges (min/max across agents per scenario) to surface where consensus is weak. Structured disagreements — where agents diverge by more than 15 percentage points on any scenario — are flagged and explained in the briefing.
During the calibration phase (the first 10+ cycles with resolvable outcomes), all agents receive equal weight. After sufficient calibration data accumulates, the system will transition to track-record-weighted averaging using Brier scores — agents with better forecasting records receive higher influence.
The system enforces several constraints that distinguish it from unstructured analysis:
Every probability adjustment must cite at least one named causal mechanism from the relevant theoretical framework — Fearon on commitment problems, Schelling on focal points, Posen on inadvertent escalation, and so on. "Tensions increased" is not analysis; the system requires the causal process connecting an event to a scenario shift.
Scenario probabilities must sum to 93–95%, not 100%. The remainder is an explicit tail-risk reserve for outcomes the scenario framework does not cover. Every update follows a Bayesian chain: prior, evidence, likelihood ratio, posterior. Adjustments exceeding moderate magnitude require explicit justification.
Every piece of intelligence entering the system must include a verifiable source URL and publication date. Claims without sources are excluded. High-impact claims — those that could significantly move probabilities — require the system to fetch and read the actual source page, not merely rely on search snippets.
When perspective agents diverge significantly on any scenario, the briefing explicitly identifies which agents disagree, what analytical difference drives the divergence, and what evidence or mechanism each side is weighting. Disagreement is a feature of the system, not a failure.
In addition to regular update-cycle briefings, the system can be posed ad-hoc analytical questions — "consultations." Each perspective agent answers through its specific lens, and the responses are synthesised into a single consultation document. Consultations produce scenario implications (directional, not quantified), feasibility assessments, failure-mode analysis, and adversary-response predictions. Unlike update-cycle briefings, consultations do not directly change the probability distribution; they inform the next cycle's assessment.
Key design choices that shape how this system works and what it measures.
We tested the scenario framework across multiple AI systems (Gemini, ChatGPT, DeepSeek) to check whether the ordering of scenarios was model-dependent. The broad ranking held. One notable variation: we configured DeepSeek as a Tehran-perspective monitor drawing on non-Western sources. It inverted the ordering of the two leading scenarios, rating regional widening above attrition stalemate from the start. This is consistent with the core lesson: how you frame the question shapes the answer. The governance layer matters more than the model.
Sentiment trackers record the volume of commentary: how much is being said. This monitor tracks the analytical signal: what is structurally changing. The difference is extensive margin versus intensive margin. A day with 10,000 articles repeating the same headline scores high on sentiment measures but contains no new information. A single IRGC statement changing the target set from military bases to energy infrastructure scores low on volume but reshapes the scenario balance. This system tracks the latter.
Each scenario probability answers a single question: what is the likelihood that this scenario is the dominant crisis trajectory four weeks from the assessment date? "Dominant" means the scenario's defining mechanism is most actively shaping the conflict at that point. This is not a prediction of final outcomes — it is a forward-looking assessment of which dynamic will be primary. Scenarios with shorter natural timelines (such as coercive submission) are assessed on whether they initiate within the window. Scenarios with longer timelines (such as attrition stalemate) are assessed on whether the pattern is entrenched. For tail-risk scenarios like nuclear sprint, the probability tracks the likelihood of the triggering decision, not weapon completion.