How scores are computed
Scores measure reliability and honesty, not output quality. A score of 100 means an endpoint reliably answers, serves a valid x402 challenge, quotes the price it lists, keeps the same wallet, and delivers after payment. It says nothing about whether the data it returns is any good.
Scores cannot be bought
There is no listing fee, no verification tier, and no way for a seller to submit or edit a score. Every number comes from our own probes of endpoints discovered in public facilitator catalogs (Coinbase CDP Bazaar and PayAI). Sellers can only improve a score by being up, honest and stable. Facilitator-reported popularity figures are shown for context but never enter the score.
Data collection
- Crawl. Every six hours we pull every listing from each catalog and record the URL, method, price, asset, network and payTo wallet. Every change is kept as history; nothing is overwritten.
- Free probe. Every six hours we request each endpoint without paying, using the example input the seller published, and record whether it answered, the status code, latency, whether the 402 challenge is valid under the x402 v1 or v2 spec, and the quoted price, wallet, asset and network.
- Paid probe. Weekly, for one sampled endpoint per host and only under a strict spend cap, we pay the quoted amount and record whether settlement succeeded, whether a 2xx followed, and the amount actually transferred on chain. Every paid probe links to its transaction. Paid probing has not started yet.
Components and weights
Each component is a value from 0 to 1. The score is the weighted sum times 100. When a component has no data (for example delivery before any paid probe), it is excluded and its weight is spread over the remaining components in proportion.
| Component | Weight | Definition |
|---|---|---|
| Uptime, 7 days | 20% | Share of our unpaid probes in the last 7 days that got an HTTP response other than a 5xx or 429. Curved so that downtime bites: 99% uptime scores about 97, 95% about 85, 90% about 72, 70% about 25, and 50% or less scores zero. |
| Uptime, 30 days | 15% | The same measure over 30 days. |
| Valid x402 challenge | 15% | Share of responses whose 402 challenge was a spec-valid x402 payment requirement for the option we would pay (exact scheme, Base, USDC). A 200 with no payment required, a 404, or a malformed challenge all count against this. |
| Price honesty | 15% | Share of comparable probes where the price quoted in the 402 matched the price listed in the discovery catalog, and (once paid probes run) where the amount settled on chain matched the quote. A mismatch on a POST endpoint whose price can depend on the body counts half. |
| Delivery after payment | 15% | Share of paid probes that settled and then returned a 2xx response. Not available until an endpoint has been paid; until then this weight is spread over the other components. |
| Wallet stability | 10% | Starts at 100 and loses 50 for every payTo wallet change in 30 days, minus the share of probes where the quoted wallet did not match the catalog. A wallet change in the last 7 days also caps the verdict at Caution. |
| Listing age | 5% | Days since the endpoint first appeared in any catalog, saturating at 30 days. |
| HTTPS | 5% | Whether the endpoint is served over HTTPS. |
Verdicts
- Pay: score 75 or higher.
- Caution: score 40 to 74.
- Avoid: score below 40.
- Unverified: no probe yet, or no probe in the last 48 hours. Unknown endpoints asked about through the API are queued for crawling.
Caps
Some events are serious enough that no amount of good history should produce a Pay verdict until they clear:
- A payTo wallet change in the last 7 days caps the score at 74.
- A paid probe charged more than the quoted amount in the last 14 days caps the score at 74.
- No response to any probe in 30 days forces Avoid.
Confidence
Each score carries a confidence level based on how much data it rests on: low with fewer than 4 probes in 30 days, medium from 4, and high with at least 12 probes spanning 3 or more days.
Host scores
A host score pools the probes of all its endpoints and applies the same formula, so one dead endpoint among fifty costs little, while a wallet change on any endpoint caps the whole host.
Categories
Every endpoint is assigned one of 23 categories from its listing: the seller's description, the URL path, the service name, tags, the hostname and the field names of the example input the seller published. Ordered rules run first and record which text they matched (a match on the description alone is confidence 0.9; a match that needed the path or metadata is 0.6). Listings the rules cannot place are sent to a small language model with the same category list and a fixed definition of each, and its answer and rationale are stored. Every classification is kept with its method and classifier version; the newest row is the current category and nothing is overwritten. Categories decide which delivery check a paid probe runs; they never change a score.
| Category | What it covers | Delivery check |
|---|---|---|
| Chain read | Balances, blocks, gas, logs, ENS, contract calls on a public chain. | planned |
| Crypto price | Spot prices, tickers, candles and market snapshots for crypto assets. | planned |
| Web search | Search over the web, news or social media. | planned |
| Web scrape | Fetch or read a page or profile and return its content, often as markdown. | planned |
| Enrichment | Company or person lookup and contact enrichment. | planned |
| Send, receive or validate email. | planned | |
| Document | PDF, OCR, transcription and file conversion. | planned |
| Utility | Geocoding, weather, currency, randomness, time, small calculators. | planned |
| Image generation | Text to image and image editing. | planned |
| Inference | LLM chat completions, embeddings, summarisation, translation and other model output. | none (no ground truth) |
| Signal | Trading signals, forecasts, risk verdicts and AI-enriched reports. | none (no ground truth) |
| Novelty | Jokes, facts, horoscopes, coin flips. | planned |
| DNS and domains | DNS records, WHOIS/RDAP, TLS certificates, IP intelligence. | planned |
| Data transform | Hashes, encodings, JSON/CSV tools and other deterministic transforms. | planned |
| Computation | Mathematical and scientific computations with a stated formula. | planned |
| Crypto market data | TVL, yields, funding, liquidations, prediction markets, token safety. | planned |
| Finance data | Stock quotes, SEC filings, treasury and macro series. | planned |
| Web audit | Site, SEO, security-header and dependency audits of a URL you supply. | none (no ground truth) |
| Public data | Government and open datasets, transit, jobs, real estate, travel, sports. | none (no ground truth) |
| x402 tooling | x402 endpoint checkers, catalog search, payment-header tools. | none (no ground truth) |
| Commerce | Prepaid credits, tips, faucets, listings and uploads. | none (no ground truth) |
| Compliance | Regulatory checks, sanctions screening, tax and customs calculators. | none (no ground truth) |
| Other | Nothing above fits, or the listing says too little to tell. | none (no ground truth) |
Reasons
Every verdict carries its two most important reasons in plain English: caps first, then the components that cost the most points, or the strongest positives when little was lost.
What we do not measure
Output quality, correctness, latency beyond basic responsiveness, or value for money. A seller that reliably returns garbage for the quoted price can score 100.
Configuration
All weights, thresholds and windows above are read live from a single configuration file in the open-source worker, so this page always matches the code that computes the scores.