Ablation — Systemic Cascade-Risk Oracle 🏔️

Ablation sells the tail. Given an exposure / dependency graph it treats the network as a Bak–Tang–Wiesenfeld abelian sandpile (self-organized criticality), drives unit stress into it, and measures the heavy-tailed distribution of cascade magnitudes: the fitted power-law exponent tau (MLE) with a KS goodness-of-fit, the mean and tail (VaR/CVaR at 95% & 99%) avalanche size, and the trigger nodes that most often ignite large cascades. The same physics as a sandpile at the angle of repose, an earthquake fault, or a power-grid blackout.

Built on oracle-core (AIMarket Protocol v2). Where Percola gives a static connectivity threshold f_c, Ablation gives the dynamic answer — the size distribution of the cascades a driven, dissipative network produces under load. Small tau / heavy tail = one default ripples across the whole market.

# Cascade: a small exposure ring with a hub → heavy tail
curl -s -X POST http://localhost:9308/ai-market/v2/invoke -H "Content-Type: application/json" \
  -d '{"capability_id":"ablation.cascade@v1","input":{
        "edges":[["a","b"],["b","c"],["c","a"],["a","h"],["h","d"],["d","e"],["e","h"]],
        "grains":3000,"nonce":"demo"}}' | jq '{tau,ks,mean_avalanche,cvar99,triggers}'

Verifiable, not asserted. The analysis is a pure function of a SHA-256 config_commitment; the drive-schedule seed is committed as H(commitment ‖ nonce). The abelian property (Dhar's theorem) guarantees the per-site topple counts are independent of relaxation order, so any verifier replays the driven-dissipative sandpile and reproduces the topple total and tau bit-for-bit. The test suite asserts this directly — three different relaxation orders give identical topple counts and final state.

Full math, diagrams and use cases: docs/en.md · docs/ru.md · docs/es.md.

Part of the AICOM oracle family. MIT.