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

# SNAG-Bench

> Temporal reasoning benchmark — 60 adversarial tasks, 5 scoring axes, designed to stay hard through 2030.

# SNAG-Bench

**The Quality Certifier.** Temporal reasoning benchmark for LLMs. 60 adversarial tasks, 5 scoring axes, 3 difficulty tiers. Designed to stay hard through 2030. Measures **Causal Resolution**: Coverage x Convergence.

<CardGroup cols={2}>
  <Card title="GitHub" icon="github" href="https://github.com/timepointai/timepoint-snag-bench">
    `timepointai/timepoint-snag-bench` — Apache-2.0, Python, Click CLI
  </Card>

  <Card title="Detailed Docs" icon="book" href="https://github.com/timepointai/timepoint-snag-bench">
    Full task reference, scoring methodology, and evaluation docs
  </Card>
</CardGroup>

## 5 Scoring Axes

| Axis | Name                         | Source                    | Status  |
| ---- | ---------------------------- | ------------------------- | ------- |
| 1    | **GSR** (Grounding)          | Flash API                 | Live    |
| 2    | **TCS** (Temporal Coherence) | Temporal reasoning engine | Live    |
| 3    | **WMNED** (Predictive)       | Proteus markets           | Stubbed |
| 4    | **HTP** (Human Judgment)     | OpenRouter LLM judges     | Live    |
| 5    | **GCQ** (Graph Coverage)     | Clockchain stats          | Stubbed |

## Usage

SNAG-Bench is a local CLI tool — not a deployed service.

```bash theme={null}
git clone https://github.com/timepointai/timepoint-snag-bench.git
cd timepoint-snag-bench
pip install -e .

snag-bench run --tier 1
```

## Task Tiers

| Tier | Difficulty  | Tasks |
| ---- | ----------- | ----- |
| 1    | Standard    | 20    |
| 2    | Hard        | 20    |
| 3    | Adversarial | 20    |

## Causal Resolution

The composite metric: **Coverage x Convergence**

* **Coverage** — how much of the relevant temporal space is represented
* **Convergence** — how consistent the rendered outputs are across runs