Live beta · US-first

Verify which return policy actually applies to this purchase.

ReturnCheck verifies which return policy actually applies to this purchase. Every determinate verdict carries the exact supporting clause. If the evidence does not resolve the case, the answer is UNKNOWN.

No sales call required. Send two product URLs and the purchase context.

$0.02per verified answer
$0.00for UNKNOWN
HTTP + MCPx402 live on Base
How it works

From purchase context to auditable answer.

A return policy is rarely one number. Seller, category, item condition, channel, membership and timing can all change the outcome.

1

Send the exact case

Provide the product URL, country and the facts that can change eligibility. Supplying the rendered page text gives the best coverage.

2

Verify the policy

ReturnCheck finds the relevant policy language and checks that the quoted clause exists in the supplied or retrieved page.

3

Act on the verdict

Your agent receives a structured outcome, return window, confidence, source URL and dated evidence—or a safe UNKNOWN.

What travels with a determinate verdict
“The exact policy clause is quoted here—not paraphrased into certainty.”
source_url · exact_clause · verified_on
Illustrative format, not a merchant policy quotation.
Evidence, not assertion

The answer includes the reason to trust it.

A verdict without its source is hard for an autonomous buyer to audit, explain or revisit when a policy changes.

Literal supporting clause for each determinate answer
Source URL and verification date
UNKNOWN when the evidence is insufficient
Free, hand-reviewed Evidence Cards for public examples
Built for agents

One decision service, two native interfaces.

Use the remote MCP tool from an agent runtime, or call the HTTP contract directly.

MCP Streamable HTTP
endpoint: https://returncheck.m-angelmartinez-fer.workers.dev/mcp
tool: check_return
HTTP POST
curl -X POST https://returncheck.m-angelmartinez-fer.workers.dev/v1/check \
  -H "content-type: application/json" \
  -d '{"product_url":"…","buyer_country":"US"}'
Trust model

Designed to fail safely.

The useful boundary is not “always answer.” It is knowing when the published evidence supports an answer and when it does not.

Clause before confidence

A determinate verdict must carry policy text that can be checked against its source.

UNKNOWN is a result

Missing seller, conflicting policy or insufficient evidence stays unresolved instead of becoming a fabricated yes or no.

Dated evidence

Policy language is recorded with its source and verification date so the answer can be audited later.

Frozen holdout · 31 August 2026

125 measured calls. Zero unsafe answers.

The same 25 hand-labelled cases were run five times against the production engine with supplied policy text.

0unsafe errors
0hallucinated clauses
100%precision on determinate verdicts
72% / 80%coverage / benchmark ceiling

The 80% ceiling is built into the set: 5 of 25 cases correctly expect UNKNOWN. The holdout was authored and run by the ReturnCheck team, not an external evaluator. All five raw JSON runs—including the misses—are public in the repository. This deployment also passed 448 automated tests.

Retrieval boundary, also measured. In a separate frozen sample of 50 US retailers on August 28, 2026, server-side retrieval reached the policy for 17 of 50. This measures page acquisition—not verdict quality. Passing page_text or page_html lets ReturnCheck verify the page your agent already rendered instead of depending on that retrieval step.

Test it on two real cases.

Send the product URLs and context. We will return the verdict and its supporting evidence.

Request the sample