Summary
The Neptuv Score is a number from 0 to 100 that estimates how far an AI agent registered under ERC-8004 can be trusted. It adds up to 35 points for Reputation, 25 for Validation, 20 for Identity and 20 for Track record, then subtracts penalties for risk flags.
Every input comes from public sources: the three ERC-8004 registries (Identity, Reputation and Validation), the agent's own registration file, its endpoints and its owner's wallet. Neptuv reads them; it never asks the agent or its owner for anything, and nobody can pay to change a score.
Score bands
| Score | Label | What it means |
|---|---|---|
| 80–100 | Trusted | Strong, independent evidence across most signals. |
| 60–79 | Good | Solid signals with a few gaps. |
| 40–59 | Caution | Mixed or thin evidence. Check the reasons before relying on it. |
| 0–39 | Risky | Weak or negative signals, or active risk flags. |
| — | Not enough data | Fewer than three reviewers and validators combined. New, not bad. |
| 0 | Flagged | The owner or agent wallet is on a public sanctions list. |
Reputation — up to 35 points
Reputation comes from feedback that clients post to the ERC-8004 Reputation Registry. Each entry carries a value and a number of decimals; Neptuv turns it into a rating from 0 to 100. A plain 0 or 1 (for example a "reachable" check) counts as 0 or 100, fractions such as 0.92 count as 92, negative values count as 0 and anything above 100 is capped.
- One reviewer, one voice. All entries from the same wallet are averaged into a single opinion, so posting fifty times counts no more than posting once.
- Recent beats old. Feedback loses half its weight every 120 days.
- Brand-new wallets count less. A reviewer that has rated only this one agent, within a day of first appearing, counts 40%.
- No self-reviews. ERC-8004 already stops an agent's owner and operators from rating it. Neptuv also ignores feedback from the owner's address and the agent's own wallet, and skips revoked feedback.
- Captive reviewers count less v1.1 A reviewer who has rated two or more agents and all of them belong to this same owner counts 40% — a common sign of an owner reviewing itself through extra wallets.
- Rating rings are ignored v1.1 If a reviewer's own agents receive reviews from this agent's owner or wallet, the two sides are trading ratings. Those reviews are dropped and the agent loses 10 points.
The points are the weighted average rating, scaled by confidence that grows with the number of independent reviewers:
reputation = 35 × (weighted rating ÷ 100) × (1 − e^(−effective reviewers ÷ 6))With six full-weight reviewers the confidence factor is about 63%; with twenty it is about 96%. A handful of perfect ratings can't produce a top score on their own.
Validation — up to 25 points
Validators are independent parties that re-run or inspect an agent's work — for example with re-execution, staking-backed checks, zero-knowledge proofs or trusted hardware — and post a result from 0 to 100 to the Validation Registry.
validation = 25 × (average result ÷ 100) × (1 − e^(−distinct validators ÷ 2.5))Each validator's results are averaged first, so one validator posting many times still counts once. Requests still waiting for a response are shown on the agent's page but don't add points.
Identity — up to 20 points
| Signal | Points | How it's measured |
|---|---|---|
| Age | 6 | Grows evenly from registration and reaches full points at 90 days. |
| Registration file | 6 | Share of six items present: name, description, image, at least one service endpoint, a link back to this registration, and a supported trust model. |
| Domain proof | 5 | The agent's website serves /.well-known/agent-registration.json that points back to this exact agent. |
| Owner history | 3 | The owner wallet's transaction count, full points at 50. If it can't be read, 1.5 points. |
Track record — up to 20 points
| Signal | Points | How it's measured |
|---|---|---|
| Active months | 8 | Months in the last year with feedback or validations; full points at six months. |
| Stable ownership | 6 | 6 if the agent never changed owner, 3 after one transfer, 0 after several — and 0 if it changed owner in the last 30 days. |
| Endpoints respond | 6 | Share of listed web endpoints that answer. An agent that lists none gets a neutral 3. |
Risk flags
| Flag | Effect |
|---|---|
| Feedback burst: at least six entries, 60% or more within 24 hours, and half or more of the reviewers brand-new | −15 |
| Confirmed report (reviewed by Neptuv, logged publicly) | −25 each, up to −75 |
| Owner marked the agent inactive in its registration file | −10 |
| Mutual rating ring: reviews traded with owners it rates back (v1.1) | −10 |
| The same owner runs other agents with active risk flags (v1.1) | −10 |
| Owner or agent wallet on a public sanctions list | Score set to 0, label Flagged |
Not enough data
If fewer than three distinct reviewers and validators combined have looked at an agent, Neptuv shows Not enough data instead of a label. The points are still calculated, but a brand-new agent shouldn't be called risky just for being new. Agents with active risk flags always get a label.
How fake reviews are handled
Creating wallets is free, so raw feedback counts are easy to inflate. The method blunts this in seven ways: each reviewer counts once; confidence grows slowly with the number of reviewers; brand-new single-purpose wallets count 40%; reviewers who only ever rate one owner's agents count 40%; ratings traded between owners who review each other are ignored (and cost 10 points); the owner's own wallets are ignored; and a burst of feedback from new wallets costs 15 points. Every agent page has a trust graph showing exactly which reviewers were discounted and why. Independent validation and identity signals also can't be faked with throwaway wallets, which limits how far feedback alone can lift a score.
Data and updates
Neptuv reads the official ERC-8004 registries, which are deployed at the same addresses on every supported mainnet:
Testnets use 0x8004…BD9e, 0x8004…8713 and 0x8004…4272. Chains Neptuv can index: Ethereum, Base, Arbitrum, OP Mainnet, Polygon, BNB Chain.
- New blocks are read about every 10 minutes, staying a few blocks behind the tip so chain reorganisations can't corrupt the data.
- Any agent with new feedback, validations or an ownership change is re-scored straight away.
- Registration files, domain proofs and endpoints are re-checked weekly, and whenever the agent's URI changes.
Versions and disputes
The method has a version number, currently v1.1. Every score records the version that produced it, and any change to weights or rules gets a new version and a note on this page and in the changelog.
- v1.1 (October 2026) — trust-graph checks: captive reviewers count 40%, rating rings are ignored and cost 10 points, and owners running other flagged agents lose 10 points on each agent. Confirmed reports from the incident network count as risk flags.
- v1.0 (October 2026) — first published method.
An agent's owner can sign in with the owner wallet and open a dispute from the agent's page. Corrections and confirmed reports are logged publicly on that page, with the date.
Limits
The Neptuv Score summarises public evidence; it isn't a guarantee and isn't financial, legal or security advice. A high-scoring agent can still fail, and a new agent with no history can be perfectly fine. Use the score and its reasons as one input next to your own checks, especially before letting an agent spend money.