sourced.run — playground & hosted API

Run the engine

A public, stateless instance of @sourcedhq/core. Paste claims below and read the verdicts, or call POST /api/assess from your own code — full reference further down.

endpoint POST https://sourced.run/api/assess auth none CORS open limits 200 claims · 200 kB / call

1Playground

The prefilled example demonstrates the guarantees: four independent origins report the same event (count climbs to CONFIRMED), a repeat from reuters adds nothing (G3), and the single-origin story stays bare (G5). Edit freely — it is the live engine.

2What a call computes

Each call is batch-scoped and stateless: corroboration is computed across the claims you send in that request (titles that describe the same event merge via the dual gate; origins are counted distinct). Nothing is stored between calls — firstSeenAt equals the request time, and no data you send is retained.

That makes this endpoint ideal for evaluating the engine and for “given these N reports, what corroborates what?” batch questions. If you need persistent event memory — first-seen timelines that survive across runs — run @sourcedhq/core yourself with a store (one load()/save() pair, any KV; see the spec §3). A hosted persistent tier is on the roadmap.

Grouping hint: the engine merges events by title similarity (spec §4). If you already know which claims describe the same event — from your own clustering — pass clusters and skip the guesswork; Sourced then only counts.

3Request schema

POST /api/assess with content-type: application/json:

Field	Type	Semantics
claims required	Claim[] (1–200)	The batch to assess. Order is preserved in the response.
claims[].id required	string	Stable report id; used to look up `clusters` entries.
claims[].title required	string	The claim in words — event identity derives from it.
claims[].origin required	string	The source; the unit of independence. Normalize consistently.
claims[].publishedAt required	string, ISO 8601	Upstream publish time; drives the “breaking” window. Unparseable → that claim can’t be “breaking”, everything else works.
clusters optional	{ [claimId]: string[] }	Your own event grouping: claim id → all origins reporting that event in this batch (≤ 100 origins per entry). Duplicates collapse (G3).
config optional	object	Threshold overrides — whitelisted numeric keys only, clamped to sane ranges (§5).

4Response schema

200 OK, JSON:

Field	Type	Semantics
verdicts	(Verdict \| null)[]	One entry per input claim, in input order. `null` = unassessable (e.g. title had no meaning-bearing tokens) — the claim passes through unlabeled, per G7.
verdicts[].corroboration	number ≥ 1	Distinct independent origins for the claim’s event within this batch.
verdicts[].corroboratingSources	string[]	Receipts — the other origins (never the claim’s own), max 6 by default.
verdicts[].firstSeenAt	string, ISO 8601	Stateless endpoint → the request time. (Persistent first-seen requires running the engine with a store.)
verdicts[].signal	"confirmed" · "breaking" · "developing" · null	See spec §4.4. There is deliberately no “true” and no veracity score (G2).
engine	string	Engine identifier.
config	object, only if sent	Echo of the clamped overrides actually applied.

5Config keys

Numeric keys only; anything else is ignored. Values are clamped to the ranges below. The defaults are the honesty contract — tighten freely, loosen knowingly (spec §5).

Key	Default	Clamp	Meaning
mergeSimilarity	0.60	0.3 – 1	Jaccard floor for merging two events (gate 1).
minSharedTokens	3	1 – 10	Shared-token floor for merging (gate 2).
confirmedAt	4	2 – 20	Origins needed for “confirmed”.
corroboratedAt	2	2 – 20	Origins needed for “developing”/“breaking”.
breakingWindowMs	1 800 000	1 min – 24 h	Publish-age window for “breaking”.
receiptsCap	6	1 – 20	Max receipts per verdict (display cap; the count is never capped).
keyTokens	8	3 – 16	Tokens in the deterministic event key.

6Errors & limits

Status	Body	When
400	{ "error": "claims must be a non-empty array" }	Missing/empty `claims`.
400	{ "error": "max 200 claims per call" }	Batch too large — split it.
400	{ "error": "payload too large (200 kB max)" } / { "error": "invalid JSON" }	Body unreadable.
405	{ "error": "POST a batch of claims" }	Any method other than GET (usage doc) / POST / OPTIONS.

No authentication, CORS fully open — call it straight from a browser.
Malformed individual claims never fail the request: they come back as null verdicts (G7). Only a malformed envelope is a 400.
GET /api/assess returns a machine-readable usage summary.
Fair use: this is a free evaluation endpoint, not an SLA. High-volume or persistent workloads: run the engine yourself, or contact us.

7Examples

curl -s https://sourced.run/api/assess \
  -H "content-type: application/json" \
  -d '{
    "claims": [
      {"id":"a","title":"Major dam breach reported upstream","origin":"reuters","publishedAt":"2026-07-10T21:00:00Z"},
      {"id":"b","title":"Major dam breach reported upstream","origin":"bbc","publishedAt":"2026-07-10T21:04:00Z"}
    ]
  }'

const res = await fetch("https://sourced.run/api/assess", {
  method: "POST",
  headers: { "content-type": "application/json" },
  body: JSON.stringify({ claims }),
});
const { verdicts } = await res.json();
// verdicts[i] belongs to claims[i]

import requests

r = requests.post(
    "https://sourced.run/api/assess",
    json={"claims": [
        {"id": "a", "title": "Major dam breach reported upstream",
         "origin": "reuters", "publishedAt": "2026-07-10T21:00:00Z"},
        {"id": "b", "title": "Major dam breach reported upstream",
         "origin": "bbc", "publishedAt": "2026-07-10T21:04:00Z"},
    ]},
    timeout=10,
)
verdicts = r.json()["verdicts"]  # verdicts[i] belongs to claims[i]
# verdicts[1] -> {'corroboration': 2, 'corroboratingSources': ['reuters'], 'signal': 'breaking', ...}

8For AI agents (MCP)

Sourced ships as an MCP server — agents get the primitive as native tools: assess (with session memory: corroboration and first-seen accumulate across calls), verify_chain and run_conformance. The server identifies itself as sourced.

# Claude Code
claude mcp add sourced -- npx -y @sourcedhq/mcp

# any MCP client (config JSON)
{ "mcpServers": { "sourced": { "command": "npx", "args": ["-y", "@sourcedhq/mcp"] } } }

Machine-readable summaries of all three surfaces live at /llms.txt on each domain. This endpoint itself is agent-friendly by design: no auth, CORS open, JSON in/out, GET returns usage.