# Check a report before publishing This workflow is for a research agent with draft text, source links and literal quotations. It finds structural citation mistakes and prepares repeatable source checks. It does not determine truth or semantic support. 1. Call `audit_report_citations` at the public MCP endpoint, or POST `/v1/evidence/report-audit`, with `report` and `references`. Use explicit `[cite:id]` markers. See `/report-preflight-example.json` and `/report-preflight-schema.json`. 2. Inspect unknown reference IDs, missing links, duplicate exact URLs, unused references and unmarked text blocks. Blocks are not classified as factual claims. Unsupported links stay visible; no work is silently paid for. 3. Try one planned source using `trial_evidence` or POST `/v1/evidence/trial`: one supported URL and one quote, up to 4096 bytes. Two attempts per client per UTC day and 100 for the whole service; failures also consume an attempt. Trial outputs are not stored purchased receipts. The free shared 20/minute quota also applies. 4. Each generated purchase batch has `requireAllSources:true`. With this option, every requested source must be readable before settlement; zero matching quotations can still be chargeable. It is not a guarantee of truth or relevance. Older inputs without this option retain the original any-readable-source billing rule. 5. POST a batch to `/v1/evidence/prepare` for exact price and input hash. Preparation is free, with no source or facilitator-capacity check. The default purchase price is 0.01 native USDC on Base per batch of up to five sources/four quotes each. 6. Use the supplied [buyer quickstart](/agent-bridge/BUYER-QUICKSTART.md). The local viem adapter signs only the pinned Base USDC authorization after explicit buyer approval. A funded, authorized buyer wallet is still required. Public MCP calls never receive private keys or sign payments. 7. Save the original input, intent and private receipt. Recover uncertainty with the same receipt; a fresh observation requires an explicit new intent. Compare later observations using `compare_evidence_snapshots`. Different extraction profiles or unreadable pages require review. For quoted text extracted by another tool from PDFs, JavaScript pages or an unlisted website, report preflight still detects structural issues and checks quotation presence in the supplied report. It does **not** independently verify those source pages or add them to the paid plan. Supported-source coverage is in `/v1/evidence/contract`; each redirect and robots policy is checked. Use the optional SVG kit when you need its source in your own process, up to 100 products/300 assets per run, or execution without more HTTP requests. Small one-product jobs can use free preview. The 0.99 USDC licence is a one-time source purchase, not a required subscription or a fee for every future render. Operational measurement stores only UTC-day counters, fixed stage/outcome/channel and coarse declared-probe category. It stores no report text, source URLs, payment headers, recovery secrets, wallet addresses or raw user-agent/IP values in those counters. Counts represent attempts; settled receipts remain the source of sales totals. Trial/report input and output are not added to the application database. Hosting network metadata policies still apply.