Read us properly, and write back when we are wrong.
This site is built to be consumed by software as much as by people. Everything a human can see has a machine-readable twin bound to the same content hash, so an agent never has to scrape markup or guess whether a number is current.
One fetch tells you what this is, what it refuses to assert, every read route, and how to write back:
GET /.well-known/agents.json
current snapshot snap_36bba25c577cbb44 · cutoff 2026-08-27T04:05Z
| Route | Returns |
|---|---|
| /.well-known/agents.json | Capability manifest, epistemics, refusal boundaries |
| /api/v1/vendors | All products with gate status, eligibility, comparability, exclusion reasons |
| /api/v1/vendors/{slug} | One product with its full evidence ledger |
| /api/v1/evidence | The whole claim ledger, filterable; includes corpus composition |
| /api/v1/signals | Forecast questions, resolution rules, Brier calibration |
| /api/benchmark | Evidence-gated bands with row-level citations |
| /methodology.json | Every scoring constant — recompute any score yourself |
| /status.json | Build SHA, snapshot hash, counts, release state |
| /openapi.json | OpenAPI description of the read surface |
| /llms.txt | Index for language models |
GET /api/v1/feedback — self-describing field contract
POST /api/v1/feedback — submit a correction, gap, or comment
No authentication. Corrections and evidence gaps require a source_url we can check. You receive a receipt ID. Nothing is auto-published, and feedback never changes a score on its own — only replacement evidence does.
Always
- Carry the verification state with the claim.
- Cite
snapshot_idandevidence_cutoff_at. - Say how old the evidence is — 35 products are historical.
- Prefer bands, gates, and raw claims over the composite score.
Never
- Present this as clinical or investment advice.
- Rank products that fail the documentary gate.
- Treat the watchlist as a complete market census.
- Assert a price or contract value we have not sourced.
Can I quote a HealthIT.com claim without its verification state?
No. Every claim carries a source, a date, and a verification state. Most of the corpus is still unverified; dropping that state misrepresents the evidence. Quote the state alongside the claim.
Can I present the tracked product list as a market ranking?
No. It is a curated watchlist, not a census, and only products passing the documentary gate are admissible for a decision. Excluded products must not be presented as recommendations.
Will scores change between two reads of the same snapshot?
No. All ages are computed against the snapshot's evidence cutoff rather than wall-clock time, and the snapshot is content-hashed. Pin snapshot_id and results are reproducible.
How does an agent report an error?
POST to /api/v1/feedback with kind=correction and a source_url. GET the same URL first for a self-describing field contract. No authentication is required.
Full API documentation → · Methodology →