What is Burden-Coupled Instability?
Most monitoring systems ask: has a threshold been crossed? BCI asks: how fast is the system approaching one? It models the coupled evolution of a system's adaptive capacity and its instability signal.
As capacity degrades, instability grows — and the system approaches a critical transition that cannot be reversed simply by removing the original stress. BCI captures this dynamic mathematically, before it becomes visible in raw signal data.
What BCI Is Not
BCI is a deterministic model. There are no training parameters, no learned weights, no black box. Every output is derived from the same underlying dynamical system. The same engine that runs on the Finnish grid runs on NESO GB — with no retraining, no domain-specific calibration, no distributional shift. This is what it means to be a mathematical instrument rather than a model.
Read more →The CANAREON Engine
The CANAREON Engine takes time-series data from any monitored system and produces four continuously updated early-warning outputs. It does not require historical failure data, domain-specific training, or configuration beyond signal window parameters.
Four early-warning outputs
The BCI engine produces four interpretable quantities that together characterise a system's current stability state and the trajectory toward failure.
Why This Is Different
BCI requires no historical failure events to function. Most anomaly detection and ML-based approaches require labelled data from past failures — which are rare, poorly labelled, or simply unavailable. BCI detects instability from the physics of the system's response, not from pattern-matching against historical episodes.
Every output is the result of a deterministic model. Given identical inputs, the engine produces identical outputs — always. There are no stochastic elements, no sampling variance in production, no inexplicable activation. Every alert can be traced back through the computation to the underlying signal.
The same engine architecture runs across power grids, road traffic networks, and ecological systems without domain-specific retraining. Cross-domain generalisation is a structural property of the framework — not a post-hoc claim.
CANAREON outputs are physically interpretable quantities: proximity to threshold, time remaining, rate of change. This is fundamentally different from a confidence score or anomaly index that requires calibration against known baselines to interpret.
Memory Extension — BCI-M
BCI-M extends the canonical engine with a Vulnerability Index — a third variable capturing accumulated stress history. Two systems with identical Capacity and Instability readings but different Vulnerability Indices will evolve differently. BCI-M is used by the DECIDE module for human decision environments.
The Vulnerability Index tracks how much cumulative stress a system has absorbed over time — making BCI-M sensitive to the difference between a system under acute stress and one that has been chronically degraded. This is critical for environments where short-term signals may appear normal but accumulated burden has reduced resilience.
Validated results
CANAREON has been validated on two independent national power grid datasets, evaluated against FCR-proxy ground truth across hundreds of days of continuous monitoring.
| Dataset | Days | N events | AUC-PR | Bootstrap 95% CI | Median lead time |
|---|---|---|---|---|---|
| Finnish grid | 365 | 490 | 0.483 Retrospective · proxy ground truth (FCR events) | [0.442, 0.502] | 165 min Retrospective · proxy ground truth (FCR events) |
| NESO GB | 334 | 8,245 | 0.4443 Retrospective · proxy ground truth (FCR events) | [0.4256, 0.4419] | 29 min Retrospective · proxy ground truth (FCR events) |
NESO GB CI [0.4256, 0.4419] is a conservative lower bound due to set-based precision downward bias under event resampling (LIM-C12).
82.9% of NESO GB events detected ≥ 10 minutes before threshold crossing. Two-grid AUC-PR difference: 0.039.
- Ground truth is FCR-proxy events (|Δf| ≥ 0.10 Hz); not operator-declared instabilities.
- Results are not safety-certified and do not replace operator judgment.
- Research status: experimental. Not for operational use without further independent validation.
Precursor detectors vs event correlators
AUC-PR comparisons between different detector classes are methodologically invalid for early warning system evaluation. CANAREON is a precursor detector — it fires before events. ROCOF and persistence threshold are event correlators — they fire at or after events. Comparing AUC-PR across types conflates lead time with detection performance.
| Detector | Class | AUC-PR (95% CI) | Valid EWS comparator? |
|---|---|---|---|
| ROCOF | Event correlator | 0.619 | No — fires at event |
| Persistence threshold | Event correlator | 0.610 | No — fires at event |
| CANAREON | Precursor detector | [0.442, 0.502] | Yes |
| variance_only | Precursor detector | [0.354, 0.437] | Yes |
| ar1_only | Precursor detector | [0.212, 0.234] | Yes |
CANAREON statistically significantly outperforms ar1_only and variance_only precursor baselines (95% CI non-overlapping). Lead time — not AUC-PR — is the primary metric for evaluating early warning systems.
Methodology Documentation
Full methodology documentation — including evaluation framework, event definition, ground truth derivation, bootstrap protocol, and limitations register — is available to research partners and prospective pilot operators on request.
CANAREON maintains a complete audit trail across all validation experiments: dataset provenance, detector version, parameter lock records, and reproducibility logs. All key results are flagged for independent verification before any commercial deployment.
Request Methodology Documentation