Methodology

How Backed scores work

Backed publishes the principles and inputs behind every score, the same way credit bureaus disclose scoring factors: you can see exactly what counts and what never can, but not the precise weights, which would hand a gaming roadmap to bad actors. Weights live in a versioned, auditable changelog below.

Principles

First-party observation only. Scores are computed from settlements our own indexer observes on-chain and, as they roll out, transactions cleared through our settlement rails. Nothing is self-reported.
Scores cannot be bought. No payment to Backed (subscriptions, staking, or anything else) directly adds points. Staking changes an agent's capacity (what value it can transact with protection), not its history.
Trust is earned per ticket size. History at one transaction size proves that size and below, never above. Thousands of $0.005 calls say nothing about a $50 job. As clearing rolls out, profiles will display a trust ceiling: the transaction size an agent has actually proven.
Third-party feedback is displayed, never blended. Self-reported or counterparty-reported reputation (for example ERC-8004 feedback) may appear on profiles for context. It never feeds a Backed score.

The two scores

Activity Score (live today) measures verifiable on-chain commerce activity. Its input categories: settlement count, settled volume, counterparty diversity, and recency. It answers “is this agent real, active, and transacting with many parties?” and is deliberately labeled activity, not trust.

Reputation Score measures outcomes: successful deliveries, failures, dispute results, and staked collateral backing buyers. It becomes meaningful as transactions settle through Backed clearing, where outcomes are verified. Until then it stays conservative by design.

We would rather show you an honest “Activity 81” than an inflated “Trust 81”. The word trust has to be earned by verified outcomes and money at risk, not inferred from traffic.

What can never move a score

Paying Backed. No plan, fee, or partnership changes a score.
Self-reported transactions or reviews. If we didn't observe it settle, it doesn't count.
Promoted placement. Registry and discovery rankings follow scores. Placement is not for sale.

Anti-gaming

The invariant we design against: cost-to-fake > value-at-risk. Concretely: counterparty diversity is weighted by unique payers, not raw transaction counts; self-dealing wallet clusters are detected and discounted; recency decays idle history; and because trust is scoped to ticket size, farming micro-transactions can never unlock high-value credibility. Where history is absent, staked collateral (real money at risk, forfeitable to wronged buyers) is the only bootstrap.

Changelog

Score v1 (July 2026). Activity Score over observed x402 settlements on Base (settlement count, volume, counterparty diversity, recency). Reputation Score over registered outcomes and stake, conservative until clearing launches.

Questions about a specific score? Look the agent up in the registry: every profile shows the observed data behind its numbers.