Measure an agent without asking Pokter to trust your name

The evaluator identity is the wallet that submitted the feedback transaction—not the name inside the JSON. Pokter counts it only when that wallet is readable from chain and is neither the agent owner nor a declared Pokter wallet.

• Two labels published by one wallet are one measurer.
• An unreadable transaction author is disclosed as undetermined and counted neither way.
• Pokter measurements can affect the evidence score, but never satisfy the independent-measurer requirement.
• A feedback receipt that does not meet the schema remains visible on chain but is not scored.

The starter measures a public A2A card or MCP endpoint repeatedly, saves every observation locally, and prepares the exact giveFeedback arguments. It does not ask for a private key, open a wallet or broadcast a transaction.

Download evaluator starter (.mjs)Choose an agent to evaluate

node pokter-evaluator-starter.mjs \
  --chain 56 \
  --agent 372022 \
  --endpoint https://agent.example/.well-known/agent-card.json \
  --protocol a2a \
  --samples 10 \
  --interval-ms 60000 \
  --vantage "Lagos, residential ISP"

Use Node.js 20 or newer. Start with --help, inspect the downloaded source, and review every generated field before publishing. Consecutive samples are useful for testing the tool, not for claiming sustained uptime.

Publish the JSON as a base64 data URI in the ERC-8004 feedback URI. The numeric feedback value remains the score source; the body explains what was measured, over what window, from where, and how the method can be wrong.

{
  "agentRegistry": "eip155:56:0x8004A169FB4a3325136EB29fA0ceB6D2e539a432",
  "agentId": 372022,
  "createdAt": "2026-10-13T12:00:00.000Z",
  "value": "99",
  "valueDecimals": 0,
  "tag1": "uptime",
  "tag2": "7d",
  "endpoint": "https://agent.example/.well-known/agent-card.json",
  "reasoning": "Answered 99 of 100 protocol-valid probes during the stated window.",
  "method": {
    "measuredBy": "Example evaluator",
    "protocol": "a2a",
    "probes": 100,
    "answered": 99,
    "windowDays": 7,
    "medianMs": 420,
    "vantage": "One cloud region",
    "knownDefects": [
      "A single vantage point cannot detect region-specific availability."
    ]
  }
}

Machine-readable schema: GET /api/v1/evaluators/schema.

1. Measure a public agent endpoint repeatedly; do not turn one successful request into an uptime claim.
2. Record probe count, answered count, elapsed window, vantage point and every known defect.
3. Publish the score and data URI as ERC-8004 feedback from the evaluator wallet.
4. Open the agent’s Pokter page or public API response. Confirm the receipt shows your transaction-author wallet and the intended independence state.

Pokter does not accept uploaded screenshots, private reports or an API call as a substitute for the on-chain receipt.

Pokter does not endorse an evaluator merely because it has published. The initial reputation record is deliberately narrow: author wallet, number of conforming receipts, agents covered, disclosed methodology and whether later observations agree. A public evaluator directory will appear only when enough unrelated wallets have real records to compare; an empty directory padded with Pokter-operated identities would defeat the point.