# Voting Monitor data API

Read-only access to Voting Monitor's measurements of how ChatGPT, Claude, Gemini and Grok answered voter-style questions about the 2026 U.S. Senate races in Alaska, Iowa, Maine, Michigan, Ohio and Texas and three more Alaska races, with prediction markets, polls and the results. Every file is JSON under /data/; start at /data/manifest.json. No key or account is needed, and nothing can be written. Data is CC-BY-4.0 (credit Voting Monitor, votingmonitor.org). robots.txt Content-Signal: search=yes, ai-train=yes, ai-input=yes.

OpenAPI 3.1: [https://votingmonitor.org/openapi.json](https://votingmonitor.org/openapi.json). Index of operations: [https://votingmonitor.org/api](https://votingmonitor.org/api). Every page, with what it holds: [https://votingmonitor.org/llms-full.txt](https://votingmonitor.org/llms-full.txt).

## Operations

### GET /.well-known/api-catalog

API catalog (RFC 9727)

RFC 9727 API Catalog: agent discovery for the public dataset.

A linkset (RFC 9264) with the RFC 9727 profile and an item link to /api; the
content is discovery.API_CATALOG (v1.87). Cached for an hour: the linkset
rarely changes, and the OpenAPI document it points at is generated per deploy.

### GET /api

Index of operations

v1.87 (readiness A2): generated from the OpenAPI document, so it cannot list an operation
the spec does not have.

### GET /data/{path}

A dataset file

Serve JSON data from the Railway volume.

The dashboard's data loader fetches paths like /data/manifest.json,
/data/markets/latest.json, etc.
Path-traversal guard rejects any resolved path that escapes the volume.

v0.22 Phase A: HEAD support + CORS (via global middleware) so dataset
crawlers (Google Dataset Search) and notebook researchers can fetch
cross-origin.

### GET /health

Service health

Liveness + scheduler status.

O(1) checks only - no rglob. scheduler.get_jobs() is O(jobs) and tiny.

v0.22 Phase A: accepts HEAD so external uptime monitors that prefer
HEAD (most do) don't see 405s.

v0.43: cost block — research project, costs are intentionally
public. Today + yesterday spend grouped by step and provider,
cached for 30s so a hammering Cloudflare uptime check doesn't
reread the JSONL on every ping.
