Start with a complete local example that matches your workflow. Every bundled fixture is synthetic. None demonstrates real forecasting accuracy or independently witnessed publication time.

## Download the example bundle

[comuvia-offline-examples.zip](/sdk/resources/downloads/comuvia-offline-examples.zip) contains the quickstart, three supplied SDK tutorials, their synthetic fixtures, source license and a contents manifest. No package wheels, keys, private store or portfolio data are included.

| Example | What you learn | Expected behavior |
|---|---|---|
| [Quickstart](/sdk/docs/quickstart) · `example.py` | Record → evaluate → append → reproduce | 5 records; displayed loss 0.09; reproduction match |
| [A: source evidence](/sdk/docs/evidence) · `tutorial_a_narrative.py` | Link source, assertion and interpretation | Narrative evidence is preserved without invented probabilities |
| [B: corrected forecast](/sdk/docs/evaluation) · `tutorial_b_forecast.py` | Keep original and corrected outcomes | Original 0.09 and corrected 0.49 remain reproducible |
| [C: provider replay](https://foreglass.ai/developers/reader/) · `tutorial_c_foreglass.py` | Retain source semantics through normalization | 7 questions, 8 forecasts, 0 eligible in the supplied fixture |

## A simulation publishing example

An application can retain model inputs, declared assumptions and result artifacts, then use the SDK to identify the records. A conditional estimate stays a conditional estimate: it is not silently scored as an unconditional real-world forecast.

<div class="diagram" role="img" aria-label="Simulation inputs and assumptions produce result artifacts, which are identified as conditional evidence and can be inspected without becoming unconditional predictions."><div class="flow-node"><strong>Inputs + declared assumptions</strong><span>Retain what can be disclosed; identify withheld information</span></div><div class="flow-arrow">↓ model execution outside the SDK</div><div class="flow-node"><strong>Result artifacts</strong><span>Snapshots + pinned questions + conditional estimates</span></div><div class="flow-arrow">↓ SDK validation and identity</div><div class="flow-node"><strong>Inspectable experiment</strong><span>Explicit conditions; no claim of measured predictive skill</span></div></div>

The downloadable SDK examples above are the portable starting point. They demonstrate evidence handling, not the accuracy of a forecasting model.