Fragility Discovery Engine — Overview
What it is: An open-source Python stress-search harness — wire a resettable discrete-time model, search shock schedules for worst-case failure, export reproducible replay JSON and counterfactual attribution bundles.
Not: a calibrated institution model, a causal inference tool, or a drop-in production risk product.
Live demo: http://34.61.255.147/ · Source: https://github.com/AgenticOp-io/fragility-discovery-engine · Release: v0.6.0
The core idea
You describe a system as a simulation world — something that resets to known starting conditions and advances one step at a time. The engine searches across possible stress sequences to find ones that maximize damage under your metric, then saves a step-by-step record of the run.
That record is JSON you can replay in the browser, compare with counterfactual re-runs, and reproduce with the same seed.
Attribution means counterfactual and path deltas on re-rollouts (FEL Δ⁻ / Δ⁺ conventions) — not Pearl-grade causal identification.
Bring your own world (primary path)
Real value is a ~100-line adapter you write (reset, step, state_vector, instability_score, is_collapsed). Tutorial: BRING_YOUR_OWN_WORLD.md.
```powershell
pip install -e ".[dev]"
fragility search --example capacity-pool
fragility check-world --example capacity-pool
```
Six reference domains (toy regression oracles)
The in-tree worlds are deliberately simplified — structural analogies for frozen CI, not models of real institutions:
| Domain | Structural analogy |
|---|---|
| Aggregate peg | Reserve + panic under pressure |
| Network contagion | Panic on a graph |
| Resource cascade | Coupled capacity layers |
| Service backlog | Queue vs processing rate |
| Liquidity ladder | Margin erosion |
| Inventory buffer | Stock under demand |
Use them to validate the harness; use BYOW for your own physics.
What a run produces
1. Replay file — timestep trajectory, collapse bit, shock lane.
2. Pareto archive (optional) — severity vs attack-cost trade-offs.
3. Attribution bundles — counterfactual re-runs and mutation chains.
4. Certificate / manifest — reproducibility metadata for papers.
All outputs are schema-versioned JSON with deterministic seeds.
What it is not
- Not a live trading system or market data feed
- Not regulatory compliance or calibrated institutional risk
- Not proof that your production system is safe
Quick start
```bash
git clone https://github.com/AgenticOp-io/fragility-discovery-engine
cd fragility-discovery-engine
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
fragility search --example capacity-pool --export-replay out.json
fragility falsify search --example ranked-store # invariant predicate demo
python scripts/run_ga_demo.py --mode aggregate --generations 4 --export-replay toy.json
```
Full guide: How to Use · Cite: Zenodo 10.5281/zenodo.20455689