Methodology
How PulseFeed audits the x402 ecosystem and computes trust · Observatory · State of x402
PulseFeed is an independent trust & safety layer for x402 agent payments. We compare catalog claims with independent endpoint observations, available receiver signals and recorded changes over time. This page explains exactly how, because a trust oracle you can't inspect isn't one.
1. Discovery & probing
We continuously crawl x402 catalogs (Coinbase CDP Bazaar and 402index), then rotate independent probes across endpoint records: a live GET that expects an HTTP 402 with a well-formed accepts[] payment requirement (scheme, network, asset, payTo, amount). An endpoint is counted as live when its latest stored observation returned a valid x402 challenge. This does not prove availability now or that an eventual paid result will be delivered. The probe budget rotates through the registry over time; a daily crawl does not revalidate every record. The age summary, computation time and per-record lastChecked are separate from the latest crawl timestamp.
2. Trust score (0–100) & verdict
For a probed endpoint we combine independent signals:
| Signal | Effect |
|---|---|
| Live & valid x402 challenge | base of the score; dead/invalid caps it low |
| Uptime / history of past probes | consistent liveness raises trust |
| Cross-source (listed by ≥2 catalogs) | corroboration raises trust |
| Response latency | fast, healthy responses add a little |
| Receiver stability over time | a stable payTo raises trust; a volatile one lowers it |
| On-chain receiver profile | a battle-tested receiver adds; a fresh/unproven one subtracts |
| Security flags (below) | a high-severity flag overrides everything → avoid |
The score maps to a verdict: safe (≥70, no serious flags) · caution (latest observation has incomplete data or minor flags) · avoid (dead, invalid, or a serious security flag).
3. Scam & anomaly flags
Beyond "is it alive", we scan for reasons not to pay:
| Flag | What it catches |
|---|---|
payto_changed | the receiver address changed vs a stable history — possible takeover/hijack |
payto_missing | zero/absent payTo — funds would go nowhere (honeypot/misconfig) |
price_bait_switch | the endpoint charges far more than the price its seller listed (Coinbase Bazaar, or a directory entry verified by the domain owner) |
listing_price_mismatch | a directory entry the domain owner never verified lists a different price than the endpoint asks — a fact about the directory, not an accusation against the seller (low) |
price_rising / price_high | price crept up over time, or is unusually high per call |
asset_not_usdc / network_asset_mismatch | payment demanded in an unexpected token or on a mismatched network |
payto_unverified_onchain | the available receiver signal is unproven; Base history depends on RPC support, while the Solana lookup uses SPL USDC balance only |
Flags are network-aware (EVM and Solana) and self-correcting: a receiver that rotates addresses by design is distinguished from a one-off swap once enough history accumulates.
4. Receiver lookup coverage
Live receiver lookups and cached profiles have different coverage. The following matrix describes implemented methods, not a successful check of every indexed receiver. Pending, unknown, unsupported and missing profiles must remain explicit.
| Lookup | Checks | Limits |
|---|---|---|
| live · Base mainnet | USDC balance; account code and nonce; best-effort incoming USDC history | Incoming transfer history depends on RPC support and may be unavailable; a receiver profile is not proof of delivery of a purchased result. |
| live · Solana via the configured RPC (mainnet by default) | SPL USDC account balance | No transfer-history check. The configured RPC and USDC mint do not establish support for every Solana cluster; inspect the returned profile and note. |
| cached · Previously observed receiver profiles | stored per-record profile and observation time, where available | Partial coverage: profiles can be pending, unknown or absent. The aggregate's legacy non-base bucket is not proof that every non-Base network is unsupported by the live lookup. |
5. Catalog accuracy audit
We record what each catalog claims (e.g. 402index's health_status) and compare it to our independent probe. The comparison is between endpoint URLs listed as healthy by a named source (currently 402index) and their latest independent probe results. Endpoint URLs and unique domains are separate units. The numerator, denominator, source and dated observation summary are shown on the Observatory; no fixed percentage represents every catalog.
6. Compounding history
Every probe appends to a per-endpoint time series (verdicts, payTo changes, price changes) and a daily ecosystem snapshot. This history cannot be reconstructed after the fact — it is why longevity strengthens the oracle rather than commoditising it.
Limitations (honest)
- We probe from one region; a service could behave differently elsewhere.
- Base and configured Solana receiver lookups differ as described above. A configured mainnet RPC does not establish support for every Solana cluster; cached profiles can be partial or pending.
- A brand-new legitimate service starts as caution until it earns history — verification, not endorsement.
- Cached scores reflect the individual record’s last observation, not the latest crawl timestamp. The free
/verifydoes not make a live probe. The separate paid /trust lookup provides a fresh check; neither path guarantees future delivery of a purchased result.
Use it: free cached verdict GET /verify?endpoint=<url> · live deep check GET /trust?endpoint=<url> · track record GET /trust/history?endpoint=<url> · drop-in guard npm i pulsefeed-x402-guard.
Machine-readable: /status.json · /reports.json · /openapi.json · /llms.txt