How the score is built
Doomsday Agent does not predict an apocalypse. It aggregates measurable systemic-risk indicators into a transparent 0–100 stress model.
Core quantities
- Level — normalized 0–100 stress from the current observation
- Velocity — whether stress is rising or falling
- Acceleration — whether that change is speeding up
- Buffer depletion — high stress coinciding with rising stress
System risk formula (v1.0.0)
SystemRisk = 0.35·Level + 0.25·Velocity + 0.25·BufferDepletion + 0.15·Acceleration
Weights live in config/weights.json and are versioned with every publish.
Propagation
A directed weighted graph connects metrics and domains (e.g., energy → food). Impacts are confidence-weighted, depth-damped, and capped by maxImpact to prevent runaway recursion. Every contribution is logged in data/cascades.json.
Confidence
Confidence falls when data is stale, secondary, inferred, or unavailable. Low-confidence metrics pull less weight in domain and global aggregates.
Phase 1 sources
- EIA — SPR stocks, commercial crude, WTI
- USGS — Mississippi & Colorado discharge
- GDACS — active orange/red disaster alerts
- World Bank — food/CPI inflation proxy
- FRED — high-yield OAS credit stress
LLM role
Optional Bedrock text explains the day. It never creates the quantitative score. Steelman and red-team sections are both required when enabled.
Limitations
- Phase 1 covers a thin slice of global systems
- Missing API keys fall back to marked seed observations
- Geopolitical/social domains may stay UNAVAILABLE until metrics are added
- Normalization anchors are expert-configurable, not universal truth