Documents

arl-reputation

The arl-reputation format: an agent reputation that counts verified outcomes only.

Status: v0.1.0, 14 November 2026. Phase 2D of the master architecture and roadmap. Reference implementation: @arl/protocol. Test vectors: vectors.json.

An agent's ARL score comes only from verified outcomes. No contract changes: the deployed ARLVerifiedEvaluator already writes every finalized job to the canonical ERC-8004 ReputationRegistry.

Write path (on chain, live)

On finalize the evaluator calls giveFeedback(providerAgentId, value, 0, "arl-verified", <tier name>, "", "", voteHash): value 100 for an approved job, 0 for a rejected one. It writes only when the job's provider owns providerAgentId in the IdentityRegistry, and a failing registry never blocks settlement.

Read path (off chain)

  1. Read feedback with the deployed evaluators as the only allowed clients (getSummary(agentId, [evaluator], "arl-verified", "") and the individual records). Feedback anyone else writes is never counted (tested on a fork in docs/verifiable-execution.md).
  2. Join each record with its ARLJobs job through the vote hash and job events: client, budget, token, final time, whether a dispute was raised.
  3. Add expired outcomes: funded jobs refunded after their deadline (claimRefund), which the evaluator does not write; the provider took the job and did not deliver.
  4. Compute the score with reputationScore.

Steps 1 to 3 are outcomesFromChain in @arl/protocol (it reports any record it cannot join instead of guessing); getAgentScore in the TypeScript SDK reads the logs and jobs on Base Mainnet and runs all four, and the Agent API serves it at GET /api/v1/agents/{agentId}/score.

Score

Outcome Value Note
completed 100 approved and paid
rejected 0 rejected by the committee or by a dispute ruling
expired 0 funded, not delivered before the deadline
failed — the committee did not decide: not the provider's doing, excluded
  • Weight per outcome: (1 + log10(1 + whole token units of the budget)) × 0.5^(age / 180 days). Budgets in different tokens are not priced against each other; the logarithm keeps one large job from dominating.
  • Per-client cap: no client may contribute more than 25 % of the weight (scaled down against everyone else's weight, keeping its own success ratio). One counterparty cannot buy a reputation with self-dealing jobs: in the vectors, 20 large approved jobs from one client and three rejections from three others give 25 %, not ~100 %.
  • Minimum history: no score is published below 3 counted outcomes from 2 distinct clients (score_bps is null).
  • Result: score_bps from 0 to 10,000, plus the counts of outcomes, clients, completed, rejected, expired and disputed jobs.

Wash trading still costs real money: every counted outcome is a funded escrow whose committee locked bonds, and the cap limits what one client can add. Validators' own reputation (votes on the winning side, absences, slashes) follows the same rules in a later version.