Skip to content

RAWLY

RAWLY — an evidence-first paleo nutrition app, where the editorial content is code: schema-validated, test-gated, and traceable claim by claim.

The thesis

Health content is where a language model's hallucinations are most expensive. A fabricated citation in a blog post is an embarrassment; a fabricated citation under a nutrition claim is a reader acting on a study that does not exist.

RAWLY starts from an audit that proved the risk is not hypothetical. Its own legacy articles — written before the pipeline existed — carried 4 hallucinated PMIDs pointing to unrelated papers, 1 nonexistent study, 1 invented author, and fabricated statistics even where a real study existed.

So the product treats content the way a codebase treats logic: every recipe and every article is a versioned JSON artifact, validated by a schema gate and a test suite in CI, and every scientific claim must resolve to an identifier that a machine can re-check against the authority.

What it does

  • Publishes recipes and articles as versioned content files, each passing a blocking schema gate before it can ship.
  • Produces that content through a multi-agent pipeline whose verification step runs isolated and blind from the authoring step — see Content pipeline.
  • Binds every claim to a re-resolvable source (PMID + DOI), and computes nutrient density from USDA data rather than asserting it — see Evidence & nutrient density.

Status

Verified against the repository at v0.11.0:

Signal State
Content 7 recipes (4 dishes, 3 juices) · 8 articles
Content gate (validate:content) recipes 7 / 7 · articles 8 / 8 — passed
Test suite 172 passed, 0 failed
Scientific sources 33 distinct papers, listed across 40 source entries (16 in the articles, 24 in the recipes), each carrying both a PMID and a DOI
Nutrient density computed for 7 / 7 recipes · 5 published, 2 withheld by the reconciliation gate

Honest scope

The score now covers the whole recipe corpus, and the app displays it: five recipes of seven show one. The remaining two reconstructed outside the ±10 % band on carbohydrate, so the product keeps their figures on file and publishes no score rather than one it cannot stand behind — the same rule that makes the other five worth reading. The gate is not decorative: it has held twice.

Private repository

Unlike the Agentic Runtime, the RAWLY repository is private. The figures on these pages are drawn from its content files, its gate output, and its test suite; the excerpts quoted are verbatim.

RAWLY is informational content, not medical advice, and its authoring personas carry no professional certification.