402xAgent rules

How a check turns a live x402 payment request into pay, skip or recheck. Machine-readable: /rules.json (version 2026-10-15).

Verdicts

pay: safe to pay these exact terms now
skip: do not pay (we saw a concrete problem)
recheck: no decision yet: try again later

Reason codes

reason_code	verdict	default confidence	rule
`price_ok`	pay	high	Live 402 offers canonical USDC on a supported mainnet, to the listed wallet and network, at the listed price (within 1%).
`price_below_listing`	pay	high	As price_ok, but the live price is more than 1% below the listed price.
`price_mismatch`	skip	high	Live price is more than 1% above the listed or claimed price.
`pay_to_mismatch`	skip	high	The 402 pays a different wallet than listed.
`network_mismatch`	skip	high	The 402 asks for payment on a different network than listed.
`asset_not_usdc`	skip	high	The 402 asks for a token other than canonical USDC.
`network_unsupported`	skip	high	The 402 only offers a testnet or a network we do not support.
`ssrf_blocked`	skip	high	Private, local or otherwise unsafe address; never fetched.
`no_paywall`	skip	medium	The endpoint answered without an x402 payment request (one probe; a route may need another method or path).
`bad_challenge`	skip	medium	A 402 came back but its payment request has no readable USDC price.
`ambiguous`	skip	medium	The payment request is unclear (several incompatible offers).
`free_trial_active`	recheck	low	The seller says a free trial is active, so there is no price to check yet.
`timeout`	recheck	low	The endpoint did not answer in time.
`unreachable`	recheck	low	Could not connect (DNS, TLS or network error).
`server_error`	recheck	low	The endpoint returned a 5xx.

within 1% of the listed price counts as the listed price.

Confidence

high: the verdict follows from terms we read in the live 402 and compared with the listing (or a definite safety block).
medium: one probe without readable terms (no_paywall, bad_challenge, ambiguous); or a pay with no listed price to compare; or a pay whose terms changed since the previous observation (history.drift true).
low: every recheck (timeouts, 5xx, network errors, free trials), and any code not in this list.

Quality codes

Separate from reason_code; never changes a verdict. Folded into reputation.

code	reputation value	rule
`quality_ok`	1	Reachable, answered HTTP 402 with a well-formed payment request (JSON body or PAYMENT-REQUIRED header), declares an output schema, and the response sample it publishes validates against that schema.
`quality_unverified`	0.7	Reachable, HTTP 402 with a well-formed content type, but there is nothing to validate: no output schema declared, or a schema without a sample. Not a fault; we never pay the seller to capture a real response.
`quality_bad_content_type`	0.4	HTTP 402 whose terms are only in the body, and the body is not JSON (content-type is not application/json or +json).
`quality_schema_mismatch`	0.3	The listing declares an output schema (bazaar extension or v1 outputSchema) and the sample it publishes does not validate against it.
`quality_bad_status`	0.2	Reachable, but the response is not an HTTP 402 payment request (e.g. 200 without a paywall, 404, 5xx).
`quality_unreachable`	0	No response: timeout, DNS/TLS/network error, or an address we refuse to fetch.

Seller reputation

score = round(100 * sum(w_i * c_i) / sum(w_i)), over the components c_i (each 0..1) that have data; a component with no data is left out and the remaining weights are renormalized. Cap: with at least 5 pay/skip checks in 30 days and pay_share under 0.2, score = min(score, 49).

component	weight	definition
`stability`	25	per endpoint: time the current payment terms (price, network, asset, pay_to) have held / time we have observed the endpoint (endpoint_history, last 30 observations); mean over endpoints. Observations without terms (timeouts) do not reset it.
`drift`	20	1 - min(1, 2 * penalty / transitions) over consecutive term-bearing observations of every endpoint; penalty per transition: pay_to change 1, network or asset change 1, price-only change 0.5. No transitions -> no data.
`verdicts`	30	last 30 days of checks: pay_share * (0.5 + 0.5 * consistency). pay_share = pay / (pay + skip); consistency = share of checks that agree with their endpoint's latest verdict. Recheck verdicts (timeouts, 5xx) are left out.
`freshness`	10	1 if the newest check is under 24 h old, then linear down to 0 at 14 days.
`quality`	15	mean over endpoints of the latest quality code's value: quality_ok=1, quality_unverified=0.7, quality_bad_content_type=0.4, quality_schema_mismatch=0.3, quality_bad_status=0.2, quality_unreachable=0. No quality data -> no data.

fewer than 5 observations, or first seen less than 2 days ago: band insufficient_data and score null (paid answers still show provisional_score and the breakdown). A new host is never given a low score for being new.
bands: high >= 75, medium 50..74, low < 50, or insufficient_data
confidence: high: >= 20 observations, tracked >= 14 days and quality data; medium: >= 10 observations and tracked >= 5 days; else low
paid, router and exempt answers: reputation {host, score, band, confidence, breakdown, inputs}; free answers and dry runs: reputation {host, band} only
reputation never changes verdict, reason_code or confidence; it is extra context, refreshed on every read

Limits

free_live_checks: 5 per client name per UTC day, and at most 25 per network (IP; IPv6 /64) per UTC day across all client names; over the network cap: HTTP 429 with Retry-After until 00:00 UTC. The IP is only kept as a salted hash.
dry_runs: separate caps per caller and per network; never an approval to pay
paid: one tenth of the endpoint's quoted price (min $0.01, cap $0.25) per check; router tier $0.01 for 10 checks; rate limited per payer per hour

Fields

reason_code: stable machine code (this list); never renamed, new codes may be added
reason: free answers: one plain-words line; paid answers: the code (unchanged since 0.6)
confidence: high | medium | low (rules above)
history: summary of the last 30 observations: stable_for_days, stable_since, last_changed, drift
history_observations: paid, router and exempt answers only: the last 30 observations (time, price, network, pay_to, asset, status)
quality: 0.37.0: paid, router and exempt answers: {code, checks {reachable, http_status, status_is_402, content_type, content_type_ok, body_bytes, schema_declared, schema_source, schema_match}}; free answers and dry runs: {code} only. Metadata of the unpaid 402 response; never changes verdict or reason_code.
reputation: 0.37.0: seller host reputation. Paid, router and exempt answers: {host, score 0..100 or null, band, confidence, low_data, provisional_score?, breakdown, inputs}; free answers and dry runs: {host, band} only. Free page: /reputation/<host>.

402xAgent