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

# Fidelity

> How strongly the input supports a reading. A number, per activation.

Fidelity is a continuous 0–1 value returned with every metamarker and every aggregate score. It reports how strongly the input supports the activation: dense, coherent, on-topic input drives it up; sparse, fragmentary, or ambiguous input drives it down.

It is granular. Each activation carries its own fidelity, so one reading can hold a `0.91` detection alongside a `0.42` one. The number is the signal; read it directly.

## Example

A 100-word reply saturated with on-topic content can return a higher fidelity than a 1000-word tangential one. Length does not set fidelity. Density and clarity do.

The bands below are a reading guide, not thresholds built into the model:

| Value   | Reading                                              |
| ------- | ---------------------------------------------------- |
| 0.0–0.4 | Input too sparse to support the activation reliably. |
| 0.4–0.7 | Activation plausible; treat aggregates with care.    |
| 0.7–1.0 | Activation well-supported by the input.              |

## Using fidelity

**In production.** Set `fidelityThreshold` on each Protocol. Activations below it are suppressed and not returned, filtering billing and downstream noise together.

**In research.** Keep the threshold low and inspect fidelity distributions across corpus subsets. Low fidelity on short inputs is expected; low fidelity on long, coherent inputs points to a Protocol misalignment.

## In the API response

```json theme={null}
{
  "protocol": "empathy",
  "score": 0.63,
  "fidelity": 0.88,
  "metamarkers": [
    { "nuance": "acknowledges-emotional-state", "fidelity": 0.91 }
  ]
}
```