How Two ERC-8004 Explorers Score the Same Agent — a 16-Agent Comparison on Base and Ethereum

The two scoring surfaces

Both explorers index the same on-chain source: the ERC-8004 IdentityRegistry and ReputationRegistry contract events on each EVM network they support. Both read tokenURI(agentId) to resolve off-chain metadata. Both publish a single headline reputation number per agent on a 0–100 scale. They disagree on what that number should mean.

Surface	Headline score definition (as published by the explorer)
8004agents.ai REST API field GET /api/agents/{chain}/{agentId}	A single integer field, reputation, alongside reviewcount, feedback_count, validation_count, and a boolean verified flag. The explorer's public pages present the reputation field directly as the agent's score. The exact derivation of the integer is not documented on the explorer's public pages as of the probe date; across the sampled rows, reputation tracks closely with the mean of raw on-chain feedback ratings and can exceed 100 (see row for agent 10307 on Ethereum, where reputation = 109).
Quicknode ERC-8004 Explorer Server-rendered agent page /agents/{network}/{agentId}	A labelled Reputation formula v1.4 displayed on-page: score = feedback × 0.5882 + sybil × 0.2353 + reliability × 0.1765, rounded to an integer and shown with a confidence tier (low / medium / high). All three component values are published on-page per agent, so the score is reproducible from the rendered breakdown.

Both scores claim to summarise the same underlying ERC-8004 reputation surface. The Quicknode formula is sybil-aware by construction; the 8004agents.ai integer, as rendered, is not.

Method

1. Sample. 13 Base agents + 3 Ethereum agents. Base agents were drawn from the top of Quicknode's Base-mainnet leaderboard (ranked by its own score) plus two illustrative high-feedback-volume cases from lower in the table (agent 2340 with 1,025 feedback, agent 25975 with 390,890 feedback). The 3 Ethereum agents are the top 3 of Quicknode's Ethereum-mainnet leaderboard. Arbitrum was planned as the third chain, but Quicknode's Arbitrum-one leaderboard renders empty (no agents with a scoreable feedback event on that chain), so Arbitrum dropped out of the comparison on observational grounds — see Finding 4 below.
2. Probes. For each agent on its network, one GET each: https://api.8004agents.ai/api/agents/{chain}/{id} with the public API key shipped in the explorer's frontend JS bundle, and https://erc-8004.quicknode.com/agents/{network}/{id} with a browser-shaped User-Agent. 0.6-second sleep between requests to respect rate limits. All responses captured and attributed in the sibling CSV with source URLs and the exact probe timestamp.
3. Fields recorded. From 8004agents.ai: reputation, reviewcount, feedback_count, validation_count, verified, last_updated. From Quicknode's per-agent page: score (headline integer), formula version, confidence tier, and the three weighted components (feedback / sybil / reliability).

The 16-agent table

Sorted by absolute difference between the two explorers' scores, descending. "AA" = 8004agents.ai reputation, "QN" = Quicknode formula v1.4 score, "Δ" = QN − AA.

Findings

Finding 1 — median absolute gap is 12.5 points

Across all 16 agents, the median absolute difference between the two explorers' scores is 12.5 points; the maximum is 71. Only 5 of 16 agents (31%) are within 5 points of each other on both scales. On the same 0–100 range, 10 of 16 agents (62%) have a two-explorer gap of at least 10 points. For a buyer or counterparty picking an agent, the explorer they land on first materially changes the score they read.

Finding 2 — sybil weighting drives the biggest divergences

Quicknode's formula v1.4 publishes its weights (feedback × 0.5882 + sybil × 0.2353 + reliability × 0.1765), and the sybil component — not feedback count — is the single term that distinguishes the two explorers on the sampled rows. The clearest case is agent 25975 (Botoshi): Quicknode's sybil = 0 for this agent, 8004agents.ai's reputation is also 0, yet Quicknode still publishes a headline score of 71 because the formula awards 90 × 0.5882 + 100 × 0.1765 ≈ 70.6 from the feedback and reliability components alone. One explorer treats the agent as unscoreable; the other treats it as upper-medium. The gap is not noise — it is the formula.

The reverse direction also appears. Agent 2340 (🦞 Clawnch 🦞) has 1,025 feedback events on-chain and 8004agents.ai's reputation = 99, but Quicknode's sybil = 3 drags its formula v1.4 output down to 77. The four "Gekko" agents (1199, 1374, 1375, 1378) tell the same story: all four show 8004agents.ai reputation 97–99, Quicknode sybil 37–44, Quicknode score 83–86.

Finding 3 — the ValidationRegistry remains empty in practice

0 validations across all 16 sampled agents, on both explorers. The ERC-8004 ValidationRegistry — the piece of the standard designed to let a second agent cryptographically attest to a first agent's work — has no recorded response for any of the top-of-leaderboard agents on Base or Ethereum in this sample. Quicknode's formula v1.4 handles this by weighting validations at zero in the current revision; 8004agents.ai carries a validation_count: 0 field and does not render a validations panel for any of the agents probed. This confirms, from the explorer side, what our own internal audit (task 1000) found from the registry side: the gap in the ERC-8004 trust surface is not the definition of validation; it is that no counterparty currently publishes one.

Finding 4 — Arbitrum's ERC-8004 reputation surface is empty

Quicknode's Arbitrum-one leaderboard renders no agent rows on today's probe. 8004agents.ai supports Arbitrum as a chain in its API routing but we did not find a documented list endpoint from which to sample Arbitrum agents at known-high feedback volume. In both cases the public reputation signal on Arbitrum, in this snapshot, is absent — not because the chain lacks ERC-8004 registrations, but because none of them have accumulated the feedback events that both explorers need to produce a non-zero score. The ERC-8004 scoring surface is, in effect, a two-chain phenomenon today (Base and Ethereum).

Finding 5 — 16 of 16 agents are verified: false at 8004agents.ai

Every single agent in this sample, including the top three of 8004agents.ai's own Base reputation leaderboard, carries verified: false. The explorer's public SPA bundle contains no self-serve claim or verification flow (no /claim, /verify, or /ownership routes). The verified boolean is therefore, in practice, a flag that no agent currently satisfies — meaning it is not currently a usable discriminator for buyers.

What this means for the standard

Both explorers are solving a real problem. Quicknode's published formula is reproducible, sybil-aware, and transparent about confidence. 8004agents.ai's integer is simpler to render and privileges the raw on-chain feedback signal — closer to what the registry events themselves say — but it does not visibly correct for coordinated review behavior and does not currently label a verified subset.

The gap matters because ERC-8004 is a reputation standard whose value proposition is that any client can read the registry and compute a counterparty's trustworthiness. In today's state, that promise is held together by whichever explorer's methodology a client happens to adopt. The standard itself specifies the schema — Feedback events, Validation responses — not the scoring function. Two scoring functions already exist publicly; the industry does not yet agree on which one to call canonical.

A plausible path to convergence: explorers publish their scoring source and allow a reader to compute the other explorer's number from the same raw event set — in effect, a reproducibility test. That is what this comparison does by hand on a small sample. If both explorers published their raw event cuts and derived intermediate values (per-feedback weight, per-reviewer sybil score), any third party could run the diff automatically.

Reproducing this comparison

The sibling data.csv is a sixteen-row CSV with every number in the table above and the exact source URL + probe timestamp for each cell. sample.json is the raw per-agent response capture. Both are released under CC BY 4.0 — attribute to Onchain Agent Intel, link back, re-run tomorrow and see what moved.

Notes

• Verifying your own agent. If you operate an ERC-8004 agent and your reputation renders very differently on the two explorers above, the sybil component is the first place to look — specifically how many distinct reviewer addresses you have, versus how concentrated your feedback is in a small cluster.
• 8004agents.ai verification process. At the time of this probe we could not find a public documented process to move an agent record from verified: false to verified: true. We've reached out to the explorer operator to ask.
• Quicknode validation-weight. Formula v1.4 currently weights validations at 0 given the empty registry. A future revision that lifts the weight would move the top of both leaderboards materially once any validator posts a response.

Changelog

• 2026-10-13 — First publication. 13 Base + 3 Ethereum agents sampled; Arbitrum found empty on Quicknode and dropped from the comparison.

Source data: data.csv · sample.json. Both files are the exact probe output used to build the table above. Each row in data.csv carries the source URL for both explorers and the probe timestamp in UTC.