Lighthouse AdvisorySLED AI Adoption Intelligence

Query the evidence before the call.

A curated index of public-sector AI evidence, built to be read by an agent. Every record was retrieved and inspected, classified by how strong its evidence is and how the outcome turned out, kept with its own limitations, and given a stable URL you can cite in front of a client.

Ask it what a state agency actually tried, what the evidence supports, and what it does not. A web search returns pages; this returns records that already say how far they can be trusted.

Connect

REST: https://sled.lighthouseadvisory.consulting/api/v1
MCP:  https://sled.lighthouseadvisory.consulting/mcp

Public read tools need no credentials. OpenAPI specification.

Read the library

  • GET /streams: category boundaries, research briefs, queries and source coverage.
  • GET /resources: combined search and filters with a total count.
  • GET /resources/:id: canonical resource, citations and sales, engineering and delivery interpretations.
  • GET /resources/:id/versions: original and later versions.
  • GET /editions and GET /editions/:stream/:date: editions and preserved history.
  • GET /runs: recorded research status.
  • GET /coverage?stream=ID: compact archive index (canonical URL, title, date, classes, edition dates) for archive awareness; GET /coverage?url=A&url=B checks up to 25 candidate sources. MCP tools: coverage and check_sources.

Atom feeds: all editions and one per stream, for example NVIDIA at /feeds/STREAM.xml. The weekly digest summarizes the last seven research days.

Combine q, stream, topic, evidence, outcome, from, to and edition. Dates filter original publication dates; undated sources are excluded when dates are applied. Pagination uses limit (1–100) and offset. NVIDIA uses the nvidia stream.

Questions a client call actually asks

# What has gone wrong in K-12, so we do not repeat it
GET /resources?stream=k12&outcome=cautionary&limit=10

# What holds up under independent scrutiny, not vendor claims
GET /resources?evidence=government-evaluation&outcome=effective

# What this account’s peers did on procurement in the last quarter
GET /resources?topic=governance-procurement&from=2026-07-01

# Everything one micro-vertical published, newest first
GET /editions?stream=nvidia&limit=30&offset=0

Cite with care

Cite the stable resource page and original HTTPS source. Preserve evidence classification, attribution, dates and limitations. Role recommendations are interpretations, not findings. Historical records retain unknown fields as null. Edition-filtered reads preserve that edition’s analysis; unfiltered reads return the latest canonical analysis.

Vendor claims are labelled as claims. A record marked vendor-claim has not been independently verified, and saying so in a client meeting is the point of the label.

Transport and errors

MCP uses stateless Streamable HTTP with JSON responses. Agent products differ in authentication support, so test the chosen client.

REST uses 400 for invalid input, 401/403 for credentials or scope, 404 for missing records and 429 for rate limits. Retry temporary failures with bounded backoff.

Contributing research

Submission, scheduling and the research skill are operator concerns and live with the source repository, not here. If you need a submission credential, ask the site operator.