NEW RESOURCE / CITATION EVIDENCE
Preflight reports.
Try a source.
Keep evidence.
Turn a draft report and references into source-check batches. Find citation structure issues free, try a real source, then purchase bounded observations with an all-readable policy.
Start the report workflow ↗{
"sources": [{
"url": "https://example.com/",
"quotes": ["This domain is for use in documentation examples"]
}]
}{
"status": "matched",
"start": 15,
"end": 63,
"matchCount": 1
}→ Source hash + bounded excerpt
Literal presence in fetched text. No claim of truth or semantic verification.
Small evidence.
Useful context.
For research agents, report pipelines and citation checks. Inspect known source URLs before carrying a quotation into the next step.
Each source returns an explicit status. A missing quote is different from an unreadable page, and neither is a judgment about whether a claim is true.
Set requireAllSources:true to submit payment only when every requested source is readable. Report-generated batches include this option. Readable sources without matching quotes can still be chargeable. Older requests without this option keep the any-readable-source rule.
Try your own source first.
Full synthetic fixture ↗Audit → trial → explicit purchase
POST /v1/evidence/report-audit checks report text with [cite:id] markers and structured references. It flags missing links, unknown references and unmarked blocks, then prepares evidence batches. Runnable report example ↗
POST /v1/evidence/prepare checks input and host coverage, calculates the request hash and returns exact payment terms before authorization. It does not fetch source pages or check facilitator capacity. Inspect the schema and example request.
POST /v1/evidence/trial accepts your own supported URL and one literal quote. It uses the production extractor without payment or a durable receipt. Two attempts per client per UTC day; 100 across the service. Failed sources also consume an attempt. Input and workflow ↗
The fixture shown here is synthetic. Live results carry their own fetch time. Free compute shares 20 requests per minute per client.
Loading synthetic example…
Prepare. Purchase.
Compare.
Eleven public MCP tools expose both services, including report preflight, custom live trial, request preparation and snapshot comparison. The evidence client saves the original input and authorization, then recovers the same result after an interrupted response.
Public MCP tools are free and never sign or pay. The optional local buyer bridge can purchase with the caller’s signer and an explicit intent. Its default purchase limit is zero; its configured limit resets on process restart.
Ready-to-run wallet adapter ↗Direct purchase client · Comparison fixture · Offline comparator
Compare saved snapshots for the same input. Unreadable pages are never treated as removed quotes. A changed or unknown extraction profile requires review. An identical recovered receipt is the same snapshot, not a new observation.
https://batchloom-studio-0913.xuikatwsicrdsidm.chatgpt.site/v1/mcp- audit_report_citations
- trial_evidence
- prepare_evidence_request
- compare_evidence_snapshots
- get_evidence_contract
- get_evidence_sample
- validate_evidence_request
- get_manifest
- get_sample
- preview_assets
- validate_batch
Product JSON → SVG cards.
Kit purchase client ↗Choose the one-time 0.99 USDC source licence when your agent needs local execution or batches of up to 100 products and 300 assets per run. One-product jobs can use the real free preview. No recurring renderer requests are needed after purchase.
Loading included renderer samples…
Try one product at POST /v1/preview, validate a batch at POST /v1/validate, or inspect the renderer schema. Purchase via GET /v1/kit. Commercial generated output is permitted; source redistribution is prohibited.
/.well-known/agent.jsonAPI contract/openapi.jsonAgent instructions/llms.txtCoverage & pricing/v1/evidence/contractAgent customers only. Purchases are offered through machine APIs. There is no browser payment checkout. An API interface cannot prove that its caller is an AI. This experimental service has no proven customer demand or earnings guarantee.
Scope. Static UTF-8 HTML, plain text and Markdown on listed hosts only. No PDFs, JavaScript rendering, arbitrary domains, private pages or paywall bypass. A listed host can still be unavailable. Readable pages with zero matches remain chargeable.
Payment and recovery. Base native USDC, x402 v2. Keep your private X-Delivery-Token, original input, intent ID and payment header. Retry all originals after uncertainty. Stored results are released after confirmed settlement. Already settled purchases remain recoverable when new sales are paused.
Measurement. Only daily aggregate operation/outcome counters and coarse declared-probe categories are stored for acquisition measurement. No report text, source URLs, wallet addresses or private receipt credentials are stored in those counters. Trial and report-audit inputs and outputs are not stored in the application database.
Privacy. Evidence receipts retain payment metadata, input/token hashes and bounded results with public URLs and excerpts, without automatic expiry. Full fetched pages are not stored. Client receipts contain original input and private recovery secrets. Submit no sensitive data. Hosting and facilitator metadata policies apply.
Interpretation. Matching uses case-sensitive literal text after NFC and whitespace normalization. Offsets count Unicode code points; the end is exclusive. Source hashes identify extracted snapshots. They do not prove a claim's truth, publication freshness or browser visibility.
Availability. No uptime guarantee or automatic refunds. The service holds no merchant private key. Settlement depends on the facilitator's available free allowance; no infrastructure credits are purchased automatically.