Use source-linked records when a research note, article or media claim needs an inspectable trail. The SDK preserves records; it does not extract articles, license source material or decide whether a claim is true.

## Preserve the chain

<div class="diagram" role="img" aria-label="Retained source snapshot links to an attributed assertion, a separate interpretation, and an optional assessment. Each step has its own identity."><div class="flow-node"><strong>Resource snapshot</strong><span>Locator · permitted retained bytes · digest · capture limitations</span></div><div class="flow-arrow">↓ reference + selectors</div><div class="flow-node"><strong>Source assertion</strong><span>Who said what, in the source's own terms</span></div><div class="flow-arrow">↓ separately attributed</div><div class="flow-node"><strong>Interpretation / assessment</strong><span>Your reading, uncertainty and review remain distinct</span></div></div>

For a paywalled article, you can retain permitted metadata and a source link without republishing the full article. This does not bypass the publisher's rights or make inaccessible source bytes independently verifiable. Record capture limitations and missing assets explicitly.

## Run the narrative example

After installing the core and extracting the [offline example bundle](/sdk/resources/downloads/comuvia-offline-examples.zip):

```powershell
./.venv/Scripts/python.exe ./tutorial_a_narrative.py --fixtures ./fixtures/synthetic --work ./narrative-run
```

Use a fresh work directory. This is the SDK's supplied source-linked statement tutorial with fictional fixtures. It demonstrates preservation and named exclusions; it does not turn prose or extraction confidence into a scoreable probability.

## Validate and verify separately

This fragment operates on your own `record.json` and permitted retained source bytes:

```python
from pathlib import Path
import comuvia

record = comuvia.load_record(Path("record.json").read_bytes())
identity = comuvia.record_digest(record)
check = comuvia.verify(record, expected_digest=identity)
print(check.record_identity)
print(check.timestamp_assurance)
```

Recomputing a digest and comparing it to itself is a local consistency demonstration. For an integrity check across a handoff, use the **previously retained expected digest** from that handoff, not a newly recomputed replacement. Pass retained `content` and referenced `records` to `verify` when those checks apply. Inspect references and assurance fields as well as issues.

## Store without overwriting

```python
from comuvia import RecordStore

RecordStore.create("./evidence-store")
with RecordStore("./evidence-store", writer=True) as writer:
    # validated_records must be ordered with referenced records first.
    for record in validated_records:
        writer.append(record)

reader = RecordStore("./evidence-store")
print(reader.verify().ok)
```

The fragment assumes you already loaded `validated_records`; the downloadable tutorial supplies the complete workflow. Keep one writer per store. The current locking design is for a single machine, not a distributed database.

## Correct the interpretation, keep the quotation

If an extractor misread a source, append a new interpretation revision naming its predecessor. Do not rewrite the source assertion to make the extraction look correct. If the publisher changes the source itself, retain a new snapshot with its own identity and explain how it relates to the previous one.

Extraction confidence remains useful for human review queues. It must never be fed into Brier scoring as the source's probability of an event.

## Application checklist

- Check your right to retain or redistribute the source bytes.
- Record who made the assertion and who made the interpretation.
- Retain missing assets, unknown timing and source caveats.
- Append corrections and preserve references to prior revisions.
- Share selected records and permitted content, rather than a raw live store.

Next: [schemas and errors](/sdk/docs/schemas-errors) or [evaluation eligibility](/sdk/docs/evaluation).

For a plain-language walkthrough of media and statistical data, read [Evidence and data quality](/sdk/docs/data-quality).