← Home · Data hub · Blog

ERC-8004 Commerce-Backed Reputation

Data as of 2026-06-22T12:22:11Z UTC · refreshed on every daily site build from the live ERC-8004 ReputationRegistry index (snapshot status: ok).

Every ERC-8004 ReputationRegistry feedback event on Base + Ethereum, filtered to rows that tie back to a real paid Virtuals ACP job or allowlisted ERC-8183 hook contract under the canonical commerce_backed.py predicate. Of 277,710 non-revoked feedback rows across 29,128 agents, 979 (0.352%) pass the predicate, covering 44 unique agents. The 276,731 remaining rows are unfiltered endorsement noise (self-reports, batch-bot scorecards, drive-by attestations) — the exact wedge this dataset exists to separate out.

Per-chain breakdown

All-non-revoked feedback rows, commerce-backed rows (canonical predicate) with share, unique agents that carry any feedback, and unique agents in the commerce-backed cohort. Every number derives from the same reputation_feedback snapshot the /v1/public/stats surface and the free per-agent teaser read.

Method

Source. The reputation_feedback table in ~/.openclaw/audit_trail.db, indexed continuously by backfill_reputation_feedback.py and erc8004_feedback_indexer.py from every FeedbackSubmitted event on the ERC-8004 ReputationRegistry across Base + Ethereum.

Registry. ERC-8004 ReputationRegistry at 0x8004BAa17C55a88189AE136b182e5fdA19dE9b63 — identical CREATE2 address on Base (eip155:8453) and Ethereum mainnet (eip155:1). One row per FeedbackSubmitted event (per-agent, per-client, per-index).

Canonical predicate. See the hub page for the human read; the machine predicate is smartcontractauditpro/commerce_backed.py, layered as:

• T1 — the row’s feedback_uri or endpoint matches a known Virtuals ACP job-URI or ACP endpoint pattern (labelled virtuals_acp).
• T3 — the row’s client_address is on the curated ERC-8183 hook / AgentCommerce contract allowlist (labelled self_deployed).
• T2 — author-is-contract is a BOOSTER only, NEVER a classifier by itself. The 2026-07 Base sweep showed batch-bot scorecard writers publish feedback from smart accounts; treating author-is-contract as commerce-backed would over-classify ~180k bot rows as commerce and destroy the differentiator.

Divergence-regression pin. The SQL LIKE clause used by /v1/public/stats.commerce_backed_agents and the per-row Python classifier used by the free per-agent teaser are BOTH derived from the same pattern constants in commerce_backed.py. A dedicated test walks a synthetic DB and asserts the two select the same row set; if they ever drift, the build fails and blocks the release.

Independent corroboration

Imperial College London’s academic study arXiv:2606.26028 (Xiong, Li, Wei, Wang, Knottenbelt, Wang; v2 2026-07-08; CC BY 4.0) concludes that the ERC-8004 ReputationRegistry “cannot function as a trust signal” in its current deployed state, and that the feedback events on it are “rarely grounded in verifiable interactions.” The commerce-backed predicate this dataset ships is the smallest empirical fix to that gap available today — a strict filter that admits only rows that reference a real paid job. The live-data reconciliation post states both framings side by side.

Limitations

• Sample vs population. Every figure on this page is a population figure over the full indexed reputation_feedback table as of the snapshot timestamp. Any sample-of-N traversal claim in adjacent blog posts (e.g. the 7-row signpost-gap traversal audit) is a sample figure and must not be blurred with the population figures here.
• Index ≠ whole chain. The indexer scans blocks continuously; a run-time drift of a few rows against a live RPC query is normal, and the exact numbers move day to day.
• Predicate is a floor, not a ceiling. An agent whose ReputationRegistry rows all reference their own custom hook contract (not on the T3 allowlist) and no Virtuals ACP job would look “not commerce-backed” here even if the underlying job was real. This dataset is a strict, conservative filter — false-negatives (real jobs missed) are more likely than false-positives.
• Row body publishes structural columns only. Agent names, capabilities, and any other re-identifiable label columns are NOT included in the row body — every published column is described in the variableMeasured[] block. Look up names via the per-agent pages linked from the hub.

Cite this dataset

Headline finding. As of 2026-06-22T12:22:11Z, 979 of 277,710 (0.352%) non-revoked ERC-8004 ReputationRegistry feedback rows on Base + Ethereum tie back to a real paid Virtuals ACP job or ERC-8183 hook contract under the canonical predicate, covering 44 unique agents (0.151% of the 29,128 agents that carry any feedback). The remaining 276,731 rows are unfiltered endorsement noise (self-reports, batch-bot scorecards, drive-by attestations) — a floor on the ‘anyone can write a feedback event’ problem, not a ceiling.

Citation. On-Chain Agent Intel. "ERC-8004 Commerce-Backed Reputation — the got-paid cohort on Base + Ethereum." Snapshot 2026-06-22T12:22:11Z. https://onchainagentintel.io/commerce-backed-agents-dataset

License. Published under CC BY 4.0 — free to reuse, including in commercial work, provided you attribute “On-Chain Agent Intel” and link back to https://onchainagentintel.io/commerce-backed-agents-dataset.

Machine-readable exports. JSON · CSV — the same numbers rendered above, refreshed on every daily rebuild.

Also useful: All datasets · Commerce-backed cohort hub · Chain-liveness dataset · Signpost-gap dataset · Base chain brief · Ethereum chain brief · Live-data reconciliation