# AI Evidence Layer — Useworthy

Canonical: https://useworthy.org/ai-evidence.html

Useworthy v0.4

# AI Evidence Layer

Useworthy publishes the same evidence in forms that people, search systems and AI agents can retrieve without guessing.

LLM discovery

**llms.txt + Markdown mirrors**

Every public HTML evidence page has a clean Markdown alternate. The root llms.txt is a curated map for agents.

[Open llms.txt →](/llms.txt)

Structured retrieval

**Evidence JSON API**

Claims, unknowns and cited evidence are exposed as JSON. The API never fills missing evidence by inference.

[Open API catalog →](/api/v1/catalog.json)

Grounded answers

**Answer packs, not generated guesses**

A grounded pack returns the Useworthy synthesis, confidence, citations and a list of what is still unknown.

[See an example →](/api/v1/grounded.php?product=mammotion-luba-3-awd)

Agent access

**Read-only MCP**

Agents can search products, retrieve evidence, compare two products and request evidence gaps through MCP.

[MCP endpoint →](/mcp/)

Anti-hallucination contract

## What an AI should do with Useworthy data

1. **Separate claim from verification.** E1 manufacturer material is not independent proof.
2. **Keep unknowns unknown.** If comparable intervention rate, failure rate or software lifetime is missing, say so.
3. **Use versioned evidence.** Prefer exact model, date and firmware/version when available.
4. **Cite the source.** Material factual claims should carry the original evidence URL.
5. **Do not turn Useworthy into certification.** It is an evidence and editorial layer, not CE approval or a safety authority.

Machine endpoints

## Public alpha endpoints

`/api/v1/catalog.json` — catalog

`/api/v1/evidence.json` — all evidence records

`/api/v1/products/<id>.json` — one product

`/api/v1/search.php?q=...` — text search

`/api/v1/grounded.php?product=...` — grounded answer pack

`/api/v1/compare.php?a=...&b=...` — evidence-only comparison

`/api/v1/openapi.json` — API description

`/mcp/` — MCP Streamable HTTP / JSON-RPC endpoint

Crawler visibility

## Discovery is open; usage is measurable

Public evidence is crawlable. Recognized AI/search user-agent requests for public HTML, Markdown, JSON, TXT and XML are logged to a private NDJSON file with a one-way IP hash. API and MCP usage are logged separately. User-agent names are observational and are not treated as cryptographic proof of bot identity.

Logs: `private-data/ai-crawler.ndjson`, `private-data/api-usage.ndjson`, `private-data/mcp-usage.ndjson`.