causalnerve 1.0.3

Adaptive Structural Dependency Learning for Non-Stationary Dynamical Systems

CausalNerve

CausalNerve: online causal graph adaptation for streaming non-stationary systems.

Core Capabilities

1. Real-time causal structure revision without retraining (see Section 6 benchmarks).
2. Intervention simulation via Pearl's do-calculus (native) (demonstrated in Turbofan demo).
3. Structural isolation guarantee: do(X) only affects descendants (formalized in PAPER.md).

Installation

pip install causalnerve
pip install causalnerve causalnerve-observe  # for dashboard

Quickstart

from causalnerve import CausalNerve
from causalnerve.datasets import SyntheticStreamGenerator

nerve = CausalNerve(nodes=6, state_dim=32)
nerve.fit(SyntheticStreamGenerator.stable(n_cycles=200), epochs=20)
nerve.watch(SyntheticStreamGenerator.with_drift(drift_at=100))

Expected output:

CausalNerve Summary
────────────────────────────────────────
Cycles processed: 200
Structural alarms: 3
Edits accepted: 1 (edge 4->2, conf=0.71)
Edits rejected: 2
Final leakage: 0.031
Causal equilibrium: reached at cycle 147
────────────────────────────────────────
Run nerve.why(4) for root cause analysis.
Run nerve.what_if({2: 0.5}) to simulate interventions.

What CausalNerve Is Not

• CausalNerve is not a causal discovery algorithm (use PCMCI for that).
• CausalNerve is not a replacement for physics simulations.
• CausalNerve is not validated for safety-critical deployment without domain expert review of its structural revisions.

Scientific Benchmarks

Evaluated on NASA C-MAPSS FD001 (Engines 81-100).

Method	SHD ↓	Det. Delay ↓	Runtime ↑	Online?
CausalNerve	0.0 ± 0.0	221.7 ± 60.3	83 ms	Yes
PCMCI	105.8 ± 9.3	N/A (offline)	4613 ms	No
VAR-LiNGAM	20.0 ± 0.0	N/A (offline)	1 ms	No
Granger	158.9 ± 13.7	N/A (offline)	780 ms	No

Note: CausalNerve is structurally self-repairing (via automated dual-world validation), which incurs an ~80ms overhead compared to static architectures. It trades offline causal discovery accuracy for online adaptability; specifically, CausalNerve loses to PCMCI on Structural Hamming Distance (SHD) for stationary chain graphs.

Examples

We provide three definitive demonstrations of the autonomous capabilities:

1. Real-time Causal Self-Repair in Turbofan Engines
2. Tracking Changing Brain Connectivity During Seizure
3. Detecting Cascading Failures in Distributed Systems

Documented Limitations

CausalNerve is bound by constraints in scalability, expected calibration error (ECE) under rapid distribution shifts, and sensitivity to highly correlated sensor noise. For a rigorous audit of system boundaries and known failure modes, please read FAILURES.md.

Citation

If you use CausalNerve in your research, please cite our technical paper:

@software{causalnerve2026,
  author = {CausalNerve Core Team},
  title = {CausalNerve: Adaptive Structural Dependency Learning},
  year = {2026},
  url = {https://github.com/guru-s/CausalNerve}
}

License

MIT License