# LifeFrom example delivery v1

All data is fictional. This package demonstrates structure and engineering checks, not clinical evidence, available inventory or model performance.

Start with dataset-card.md, then observations.jsonl and source-manifest.json. The eight rows include one intentionally invalid rejected observation. Non-accepted rows are retained to demonstrate audit history and review states.

## Verify

Use Python 3.10+ in an isolated environment. Install `jsonschema[format]==4.26.0`, then run `python validate_delivery.py`. Run `python validate_example.py` for the original fixture's negative tests.

Validation checks Draft 2020-12 structure, source identity, declared participant membership, source integrity, fixture timestamp evidence, missingness, release rights, demo splits and correction graph integrity. Broad numeric sanity bounds are not clinical validation. Source timestamps are fictional assertions, not authenticated receipt times. Historical selection returns record IDs; do not feed release audit metadata directly into model inputs. The unsupported rejected record must fail the evidence check. The process writes qa-report.json from the observed results.

Files also include a data dictionary, annotation guide, split specification, rights summary, checksums and changelog. checksums.json covers payload files except itself. Re-running validation should preserve the report contents.

Nothing in this package authorizes training on real patient data or medical decision-making.