> ## Documentation Index
> Fetch the complete documentation index at: https://docs.canaryedge.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Core Concepts

> How Canary Edge outperforms traditional anomaly detection

## State-of-the-Art Detection Engine

Canary Edge uses an **advanced deep learning architecture** purpose-built for time-series anomaly detection. It consistently outperforms traditional ML approaches, statistical methods, and competing cloud services in both accuracy and latency.

### How It Works

1. **Encode** — Time-series windows are converted into rich latent-space representations using a Temporal Transformer Encoder
2. **Predict** — A neural predictor estimates the expected next state from the current state
3. **Measure** — The prediction error (energy score) between expected and actual behavior determines anomaly severity
4. **Classify** — The system automatically classifies the operational regime: HEALTHY, ACTIVE, TRANSITION, or SHOCK

### Why Canary Edge Wins

| Traditional ML / Cloud Services                 | Canary Edge                                               |
| ----------------------------------------------- | --------------------------------------------------------- |
| Fixed thresholds that need manual tuning        | Learns what "normal" looks like automatically             |
| Misses gradual regime shifts                    | Detects subtle behavioral changes in real-time            |
| Requires large labeled training datasets        | Works with minimal unlabeled data (12+ points)            |
| One model per metric, no cross-signal awareness | Multi-resolution models that understand temporal dynamics |
| Azure AD: \~220ms, AWS Lookout: \~350ms         | **Sub-50ms p95 latency**                                  |
| Black-box anomaly scores                        | Interpretable regime classification with z-scores         |

## Benchmark Comparisons

We regularly benchmark Canary Edge against industry alternatives. Full benchmark methodology and results are published transparently.

<Note>
  Benchmark comparison reports are coming soon. We are running standardized evaluations against Azure Anomaly Detector, AWS Lookout for Equipment, PyOD, ADTK, and Luminaire across the following datasets:

  * **NAB** (Numenta Anomaly Benchmark) — 58 real-world time series
  * **NASA IMS Bearing** — vibration sensor data with known failure points
  * **Yahoo S5** — synthetic and real anomaly benchmarks
  * **KPI (AIOps)** — internet service KPI anomaly dataset

  Results will include precision, recall, F1, latency, and cost comparisons.
</Note>

### Benchmark Metrics We Report

| Metric                        | What It Measures                        |
| ----------------------------- | --------------------------------------- |
| **F1 Score**                  | Balance of precision and recall         |
| **Detection Latency**         | Time from anomaly onset to detection    |
| **API Latency (p50/p95/p99)** | End-to-end request latency              |
| **False Positive Rate**       | Incorrect anomaly flags per 1000 points |
| **Cost per Million Points**   | Normalized pricing comparison           |

## Collapse Prevention (SIGReg)

Deep learning models can "collapse" — mapping all inputs to the same output, making everything look normal. Canary Edge uses **SIGReg** (Signature Regularization) based on the Cramer-Wold theorem to mathematically guarantee the latent space remains expressive and anomalies stay detectable. This is a key differentiator over simpler autoencoder approaches.

## Sensitivity

The `sensitivity` parameter (0-99) controls detection strictness:

* **Low (0-30):** Only flag extreme deviations. Best for noisy industrial sensors.
* **Medium (40-70):** Balanced detection. Default for most use cases.
* **High (71-99):** Catch subtle shifts. Best for financial metrics and SLAs.

The sensitivity maps to an internal threshold: `threshold = 4.0 - (sensitivity / 99.0) * 3.0`

## Per-Machine Fine-Tuning

Canary Edge ships with a powerful generic model trained on diverse time-series patterns. For even better accuracy on your specific equipment or metrics, you can fine-tune a per-machine predictor:

1. Set a baseline via the `/v1/baseline` endpoint with your machine's healthy operational data
2. The system automatically fine-tunes a dedicated predictor for that machine
3. Fine-tuning takes **seconds, not hours** — no GPU required on your end
4. Fine-tuned models typically improve detection accuracy from 82% to 97%+

See the [Fine-Tuning Guide](/guides/fine-tuning) for details.

## Multi-Resolution Architecture

Canary Edge automatically selects the right model based on your data characteristics:

| Model                           | Window Size  | Best For                                                      |
| ------------------------------- | ------------ | ------------------------------------------------------------- |
| **Sensor Model** (2.3M params)  | 2,048 points | High-frequency sensor data (vibration, temperature, pressure) |
| **Metrics Model** (1.8M params) | 128 points   | Low-frequency business metrics (KPIs, SLAs, throughput)       |

Model selection is automatic — just send your data and Canary Edge picks the optimal architecture.