Can you trust an ERC-8004 reputation score on Base and Ethereum?

Not without additional filtering. The ERC-8004 ReputationRegistry (0x8004BAa17C55a88189AE136b182e5fdA19dE9b63) is a permissionless event log — anyone can post a feedback row for any agent — with no anti-Sybil design at the contract level. The key distinction is between the unfiltered feedback count and the commerce-backed subset (feedback rows tied to a paid job under a canonical predicate). Only the commerce-backed subset survives as a trust signal; the unfiltered count does not. Independent measurement by Imperial College (arXiv:2606.26028) reports the same registry "cannot function as a trust signal" in its current deployed state.

Data hub · Commerce-backed cohort · How many agents?

Live data · updated daily

Can you trust an ERC-8004 reputation score?

Live measurement across Base + Ethereum; refreshed on every visit.

Short answer: not without additional filtering. The ERC-8004 ReputationRegistry (0x8004BAa17C55a88189AE136b182e5fdA19dE9b63) is a permissionless event log with no anti-Sybil design at the contract level. Of the 26,978 ERC-8004 agents Onchain Agent Intel indexes across Base and Ethereum, 155,767 have any on-chain feedback row and only 51 carry any feedback tied to a paid job (commerce-backed — the paid-job subset). Independent measurement by Imperial College (arXiv:2606.26028, CC BY 4.0) reports the same registry “cannot function as a trust signal.”

Why the raw score is not a trust signal

We look at it in four steps. Each says what it does and does not show.

Layer 1 — the registry is a permissionless event log

The ERC-8004 ReputationRegistry (0x8004BAa17C55a88189AE136b182e5fdA19dE9b63) accepts a submitFeedback(agentId, tag1, tag2, uri) call from any caller for any agent. There is no whitelist of trusted evaluators, no on-chain proof of a paid interaction between the caller and the agent, and no cost to posting a row beyond gas. A single client wallet with $1 of Base ETH can post an unlimited number of rows for its own address or for any agent it wants to boost or attack. Read the raw feedback count as event volume, not as reputation.

Layer 2 — the unfiltered count (upper bound)

Onchain Agent Intel has indexed 710,900 non-revoked feedback rows on the ReputationRegistry across Base and Ethereum, touching 155,767 distinct agents. This is the upper bound on “agents with any reputation.” It includes every row a client wallet posted — self-endorsements, cross-endorsements between colluding addresses, single-address booster loops, and genuine peer feedback — with no way from the contract state alone to tell them apart.

Layer 3 — the commerce-backed subset (the trust signal)

Only 1,003 of those feedback rows tie back to a paid job (commerce-backed), touching 51 distinct agents (51 on Base, 0 on Ethereum) authored by 117 distinct client addresses. The predicate: the feedback row's feedbackURI matches a Virtuals ACP job outcome pattern, OR its transaction was emitted from an allowlisted ERC-8183 hook contract, OR it carries a booster tag but the caller has an on-chain payment history to the agent. The same definition is used across our free statistics, paid API and datasets.

Layer 4 — what independent measurement says

Xiong et al. (arXiv:2606.26028, Imperial College London, CC BY 4.0, v2 2026-07-08) measured the same registry on a frozen 2026-05-13 snapshot and reports that the ReputationRegistry “cannot function as a trust signal” in its current deployed state, citing the same design gaps (no anti-Sybil, no proof-of-interaction, permissionless posting). Their finding matches ours: the raw feedback count is not a reputation, and the subset tied to paid work is small. Our side-by-side comparison is in Can Trustless Agents Be Trusted?.

Method

Registry. ERC-8004 ReputationRegistry at 0x8004BAa17C55a88189AE136b182e5fdA19dE9b63 — the same address on Base (eip155:8453) and Ethereum mainnet (eip155:1). Every non-revoked feedback row is one FeedbackSubmitted event.

Commerce-backed predicate. A feedback row is commerce-backed if the transaction was emitted from an allowlisted ERC-8183 hook contract, OR its feedbackURI matches a Virtuals ACP job outcome pattern, OR its tag+client-address combo passes the booster+payer test. The same definition drives this page, the paid API, the commerce-backed hub and the commerce-backed dataset, so their figures agree.

Update schedule. Aggregate counts are recalculated on every visit from the public statistics.

Limitations

Commerce-backed is a strict floor, not a ceiling. Genuine peer feedback from a client who paid the agent off-chain (fiat, cross-chain, batched settlement) does not clear the commerce-backed predicate. We use the strict cut on purpose: counting fake feedback as real is worse than missing some real feedback.
Sybil detection is an active research area. Layer 3 catches obvious booster loops but does not catch sophisticated multi-address coordination that also settles genuine on-chain payments. Independent research (Imperial arXiv:2606.26028, section on collusion) reaches the same conclusion on the same registry.
Registry ≠ population. An agent with zero commerce-backed feedback is not necessarily untrusted — it may simply not have transacted on-chain yet. The reverse is not true: an agent with a high raw feedback count and zero commerce-backed rows is a very specific signal.

Cite this answer

Citable summary. Scope: ERC-8004 ReputationRegistry rows on the Base and Ethereum registry (0x8004BAa17C55a88189AE136b182e5fdA19dE9b63), refreshed daily. On the Base and Ethereum registry, Onchain Agent Intel indexes 710,900 non-revoked feedback rows across 155,767 distinct agents, of which 51 agents carry any commerce-backed row (feedback tied to a real paid on-chain job).

License. Answer prose and every rendered number published under CC BY 4.0 — free to reuse, including in commercial work, provided you attribute “Onchain Agent Intel” and link back to https://onchainagentintel.io/can-you-trust-an-erc-8004-reputation-score. Imperial College quote reused under CC BY 4.0 with attribution to Xiong et al., arXiv:2606.26028.

Machine-readable source. Every rendered number is a live projection of /v1/public/stats — identical JSON to what this page renders, no auth, no payment.

Also useful: Commerce-backed cohort hub · Commerce-backed dataset (CC BY 4.0) · Imperial reconciliation · The trust layer · How many agents? · All datasets