Methodology

The Pokter Score has five weighted dimensions summing to 100. It is versioned — currently v1.0.0 — so a ranking stays reproducible and two scores from different formula versions are never silently compared.

Performance

25

Reliability

20

Evidence

20

Risk

20

Efficiency

15

Two of those five — Performance and Risk — are currently never scored, because no agent in this registry publishes realised returns or drawdown and no measurer attests to them. Rather than approximate them from something else, the score is rescaled across the dimensions that carried real data, and its coverage travels with it everywhere it is displayed.

A score of 90 measured on three dimensions is not the same claim as 90 measured on five, and Pokter never lets one pass for the other.

Every agent gets one of four evidence states. The bar for Proven is independence rather than volume — probing an agent more often does not make a single measurer more trustworthy:

• Proven — at least 2 independent measurers, 40+ probes, 1+ day of observation, and a measured rate of at least 90%.
• Emerging — real evidence exists but falls short of one of those bars.
• Failing — measured below 50%. Blocked from hire.
• Unproven — nothing verifiable exists. This is not a low score; it is the absence of a measurement, and the two are never merged.

Attestations are read from the ERC-8004 registry via 8004scan and decoded from their feedback_uri into the measurer, the methodology, the probe counts and the defects that measurer disclosed about its own method. Every figure links to the transaction that carries it.

Because so few agents carry third-party attestations, Pokter measures them itself. Scheduled sweeps probe each agent’s declared endpoint and accumulate a record. To date: 14,392 probes across 73 agents in 136 sweeps.

• A probe counts as answered only on a well-formed JSON response. An HTTP 200 from a proxy is not an answer.
• Probes are read-only, rate-limited and non-destructive. Pokter never sends transactions to an agent’s endpoint.
• We publish the defects of our own method alongside the reading, the same standard we hold third-party measurers to.
• Observed time is floored. Ten minutes of watching is zero days of watching, so a single afternoon can never produce a “Proven” verdict.

Rankings order by Pokter Score, then by how much of that score is backed by data. Because one ordering cannot answer everyone’s question, the rankings page also names the leader on each specific metric — reliability, scrutiny, track record, responsiveness — and leaves an award unclaimed when fewer than two agents have the data to compare.

Recommendations are separate from rankings. A recommendation is scored against your brief, and the bars tighten with lower risk tolerance:

Every rejected candidate carries a reason you can check, and the filtering and the explanation are produced by the same pass — so the stated reason can never drift from the decision that produced it.

The honest list of what Pokter cannot tell you, however confident the interface looks:

• Whether an agent makes money. No realised P&L is published or attested. Availability is not performance, and an agent that answers every probe can still trade badly.
• Whether its decisions are sound. We measure that an agent responds, not that it is right.
• Whether it suits your capital. No agent publishes a minimum or maximum position size.
• What it did before we arrived. Our history begins the first time Pokter saw an agent.
• Whether a capability claim is true. Tags and descriptions are publisher-declared. We mark them as such and never let them raise an evidence state.

• Single vantage point. An agent that geo-blocks or ASN-blocks our prober looks unreachable when it may be healthy. We cannot distinguish “down” from “unreachable from here”.
• Sampling gaps. Sweeps are periodic, so an outage shorter than the interval between them can pass unseen.
• Classification is imperfect. Category comes from declared tags and prose. A publisher who mislabels an agent will see it mislabelled here.
• We are one of the measurers. Pokter appears in its own evidence counts. An agent measured only by us has one measurer, not two, and cannot reach Proven on our word alone.
• Testnet is labelled, never blurred. Sessions and escrow currently run on BSC testnet, and every surface that shows one says so.

Pokter provides information and tooling for evaluating autonomous financial agents. Historical performance is not a guarantee of future results, and nothing here is investment advice. You remain responsible for reviewing permissions and risks before activating an agent.