Report · 2026-08-31 to 2026-10-28
A 2xx is not a delivery
Every paid call Teppi made to someone else's endpoint, and what came back. A status code says the seller answered. Only a check says the answer was right, so Teppi pays at the listed price, holds each answer to one the seller never saw, and publishes every call with its transaction.
Where every paid call ended3,646 calls · 422 endpoints · 355 hosts
2,649answered 2xx and right · 72.7%
72answered 2xx and wrong · 2.0%
741answered 2xx, nothing to check it against · 20.3%
184no usable answer · 5.0%
Paid calls by week
08-3109-0709-1409-2109-2810-0510-1210-1910-26
rightwronguncheckedno usable answerweeks start on monday, UTC; the last runs to 2026-10-28
How to read this
Correctness is scored only where a held-out answer exists. The 741 answers without one are counted as unchecked, never as right.
Tasks are written for endpoints that answered once, so the checked set leans toward sellers that work.
Our own endpoints, calls bought for buyers, and the 168 calls a correction of ours withdrew are left out.
Each line of the file names its settlement transaction and its evidence, so any single call can be checked on its own.
Wrong answers by class
| Class | Checked | Wrong | Share wrong | Share wrong, drawn |
|---|---|---|---|---|
| market_data | 659 | 3 | 0.5% | |
| search | 557 | 24 | 4.3% | |
| onchain | 346 | 0 | 0.0% | |
| extraction | 243 | 13 | 5.3% | |
| inference | 240 | 12 | 5.0% | |
| evidence | 183 | 3 | 1.6% | |
| research | 161 | 16 | 9.9% | |
| lookup | 154 | 1 | 0.6% | |
| compute | 137 | 0 | 0.0% | |
| other | 41 | 0 | 0.0% |
A wrong answer here was billed every time: the payment settled before the answer was checked.
The newest wrong answer on each endpointeach opens its endpoint, its paid calls and their evidence
| Day | Verb | Endpoint | Class |
|---|---|---|---|
| 2026-10-28 | GET | api.kadec0.xyz/v1/papers?maxItems=10&q=transformer+attention&sort=citations&source=all | evidence |
| 2026-10-28 | POST | agent.pocket.network/v1/federal-register | extraction |
| 2026-10-28 | GET | x402.sniperx.fun/api/v1/token/price?token_address=example_token_address | market_data |
| 2026-10-27 | GET | netintel.dev/exa/search?num_results=5&query=x402+payment+protocol | extraction |
| 2026-10-20 | GET | webtools402.176-109-107-191.sslip.io/search?max_results=5&query=x402+protocol | search |
| 2026-10-19 | GET | api.oblique.markets/api/v1/paid/bazaar-pulse | research |
| 2026-10-19 | POST | chat.gedx402.com/v1/chat/completions | inference |
| 2026-10-17 | POST | win.oneshotagent.com/v1/tools/search | search |
Recount it yourselfneeds curl and jq, nothing else
- 1Download the file every figure is counted from
curl -s https://api.teppi.xyz/v1/market/records > records.ndjson - 2Count the paid calls
jq -s '[.[] | select(.source == "probe" and (.first_party | not) and (.corrected | not))] | length' records.ndjson - 3Count the 2xx answers that were wrong
jq -s '[.[] | select(.source == "probe" and (.first_party | not) and (.corrected | not)) | select(.liveness == 1 and .correctness != null and .correctness < 1)] | length' records.ndjson - 4Count the calls that kept the money for nothing usable
jq -s '[.[] | select(.source == "probe" and (.first_party | not) and (.corrected | not)) | select(.honesty == 0)] | length' records.ndjson