Endpoints: 28,729MCP servers: 18,412Payout addresses: 2,070Paid calls: 1,507Letters: 13Defects: 1,322counted just now
teppi
Method · version 31, written 2026-10-02

How a grade is produced

Teppi pays for a call, keeps what came back, and checks it. Anyone can recompute every figure.

Read the full specification ↗See it on real calls

01Three rules behind every number

Full text ↗
Paid, never askedBought at the listed price. No seller submits results.
Kept before it is judgedStored by its hash before any check runs.
Recomputable by anyoneEvery card carries its transaction, hashes and signature.

02One probe

Full text ↗

One unpaid request, then one paid one, always in this order.

01Read the termsUnpaid; reads the 402.
02Open a receiptBefore any money moves.
03PayAt the listed price.
04Keep the exchangeWhole, by content hash.
05Read the chainWhat moved, per the chain.
06Score itThe checks run last.
No answer in 30 seconds? The receipt stays open until the payment expires, then the chain shows whether money moved.

03Three tiers, never mixed

Full text ↗

The tier sits beside every grade: four measurements never compare with six.

LIVENESS_VERIFIEDFree handshake

Nothing bought. Shows it answers, never what it delivers.

  • correctnessnull
  • honestynull
  • letternever
DELIVERY_VERIFIEDPaid, no output schema

Paid, but nothing published to check the answer against.

  • correctnessnull
  • honestymeasured
  • lettermeasured
OUTCOME_VERIFIEDPaid and checked

Paid, and the answer checked against a published or public truth.

  • correctnessmeasured
  • honestymeasured
  • lettermeasured

04Six components

Full text ↗

A weighted sum. Refusing costs a seller nothing; charging for a refusal does.

correctness35%
The answer matched what the seller said it returns
honesty25%
No money was taken for an answer that failed
schema_conformance15%
The response fits the schema the seller published
liveness15%
The endpoint answered at all
latency_p955%
95th percentile time, against the median for its class
price_stability5%
What settled matched what was advertised

05The published figure is a lower bound

Full text ↗

The low end of a 95% interval, so 30 perfect samples never outrank 300.

AC observed rate published lower bound 030100200300 paid samples Illustrative: the same perfect record, read at every sample size.
  • A time-uniform confidence sequence, so the bound holds however often it is read, not only at one sample size.
  • Recent evidence counts more: its weight halves every 14 days.
  • No single paying wallet carries more than 10% of the weight behind a component.
  • Failures are never trimmed away. Latency and price, which are not rates, use a seeded bootstrap anyone can rerun.

06Letters, and what UNRATED means

Full text ↗

Unknown and bad are different facts and stay different everywhere.

Composite to letter
FDCBA
0.000.450.650.800.901.00
UNRATED

Not a low grade: too little paid evidence yet. A letter needs 30 paid samples over at least 14 days, inside a 90 day window. A free handshake never earns one.

07What Teppi never does

Full text ↗
No seller can pay to change its grade, its rank, or whether it is routed to.
No correctness or honesty figure exists without a settled payment behind it.
No grade comes from anyone telling us something. There is no review form.
No composite is published without its components, tier and sample count.
No letter for a tier nobody paid for. The database refuses it.
Nothing settles above the advertised price. The database refuses that too.

08Check any of it yourself

Full text ↗

1Run the same probe

From your own machine. It sends the bytes the scheduled prober sends, and a test holds the two equal.

npx teppi-check <endpoint-url>

2Hash the evidence

The store is addressed by content, so the name is the proof.

curl -s https://api.teppi.xyz/v1/evidence/<hash> | shasum -a 256

3Check the signature

Ed25519 over RFC 8785 canonical JSON, against the published keys.

curl -s https://api.teppi.xyz/.well-known/teppi-keys.json

09Corrections

Full text ↗

The record only gains rows. A correction sits beside its entry, names only our own faults, and a seller can answer with a counter-probe.

The source textcurl -s https://api.teppi.xyz/docs/verification-method