PipelineArchitecturePolicyExecutionCalibration
PEI methodology

How pressure becomes
an intervention map.

PEI is not a prediction feed. It is a scenario-native simulation method that turns stress signals into breakpoints, constrained actor responses, burden-transfer maps, and stabilization paths.

5

simulation phases

4

technical layers

10k+

parallel branches

Five-phase pipeline

From signal
to stability.

The pipeline is designed to answer the operational question: where does pressure go, who absorbs it, and which action changes the equilibrium?

Four-layer model

Not one model.
A system.

01

The stage

World-State Model

A directed, weighted graph of institutions, contracts, routes, buffers, capital limits, and dependencies. NetworkX keeps active runs fast; Neo4j preserves versioned state.

NetworkX / Neo4j / Redis

02

The weather

Predictive Layer

Volatility, dependency, and regime models produce distributions instead of single-point forecasts, giving each run realistic uncertainty.

PyMC / Pyro / Stan / GARCH

03

The actors

Agent Policy Layer

Institutions do not role-play. LangGraph policies choose from allowed actions and validate every move against hard constraints.

LangGraph / LiteLLM / policy YAML

04

The director

Scenario Engine

Monte Carlo branches run in parallel, mutate the graph, aggregate outcomes, and return auditable intervention maps.

Ray / Mesa / Prefect

Constrained agent policy

Agents do not
role-play.

The graph defines what exists. The constraint layer defines what is possible. The model only selects from validated actions.

Observe

Read only the local state a real institution would know: nearby nodes, contracts, buffers, and balance-sheet pressure.

Evaluate

Compute margin stress, liquidity headroom, exposure concentration, slack, and trigger conditions using deterministic business logic.

Generate

Ask the constrained model to select from pre-compiled actions, never an open-ended fantasy response.

Validate

Reject anything that violates capital, treaty, contractual, operational, or policy constraints; then write the approved action back to the graph.

policy:
  institution: Tier_1_Reinsurer
  trigger: geopolitical_tension > 85
  allowed_actions:
    - surcharge_renewals
    - activate_war_risk_exclusion
  validation:
    capital_ratio >= regulatory_minimum
    no retroactive clause voidance
> > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > >
Execution flow

Auditable from
click to memo.

FastAPI receives the scenario, Prefect orchestrates, Redis restores state, Ray and Mesa run branches, LangGraph validates actions, and the UI receives the intervention map.

01

Trigger

02

Orchestrate

03

Retrieve state

04

Distribute runs

05

Execute actors

06

Query policy

07

Validate action

08

Mutate graph

09

Aggregate

10

Persist

11

Deliver

Mechanisms and calibration

The moat is
mechanistic.

Pass-through
Margin absorption
Demand destruction
Rerouting
Subsidy intervention
Credit tightening
Insurance repricing
Buffer expansion

Backtesting flywheel

Historical crises become calibration events. PEI replays pre-crisis snapshots, compares predicted burden-transfer paths against observed outcomes, then updates the mechanism library where the model missed a real-world channel.

Methodology review

Pressure is only useful when it becomes a decision.