How scoring works
The short version
- Every page has a trust score from 0 to 100. Red (0–25) means critical risk. Green (76–100) means low risk.
- From October 2026, scores are calculated by a published formula from the page’s sourced findings. What matters most is who is responsible for each problem.
- Until now, most scores were an AI’s judgment. We said otherwise on one of our pages, and we were wrong to.
- Every score change is saved with a written reason that you can see on the page’s version history.
- A page not yet rescored by the formula is marked Provisional.
How a score is calculated
Every page starts at 65. Each problem the page documents subtracts points: more for serious problems, and much more when the subject itself is responsible.
| The subject… | Example | Share of the points that count |
|---|---|---|
| did it | ran a Ponzi, took customer funds, lied to users | 100% |
| has a real concern, not fraud | centralisation, disputes, compliance problems | 40% |
| was the victim | was hacked, had keys stolen | 40%, capped at 20 points in total |
| was near it | scams run by others on a chain, impersonators | 10% |
A problem is worth 25 points if critical, 15 if high, 8 if medium and 3 if low, before the share above. Limited credit is then added for legitimacy: being well known and widely used, long-running, regulated, or audited.
Then these rules apply, in this order:
- No red label without evidence of fraud. A page goes below 26 only if it records an established finding that the subject did harm: a conviction, a guilty plea, a court or regulator's finding, or plain on-chain facts. Accusations alone do not.
- A scheme goes to 0. If the page establishes that the subject is a scam, Ponzi, rug pull or laundering service, the score is 0 and no credit applies. “Established” also covers a law-enforcement seizure, a final regulatory order, or charges plus a documented collapse.
- Regulatory wrongdoing stays in WARNING. A guilty plea for anti-money-laundering, sanctions or licensing failures, with no harm to customers, holds a company between 26 and 50.
- People who harmed customers or victims are red. Where it is established, for example by a guilty plea for fraud, a person's page is 25 or below.
- Defunct projects are capped at 50 and labelled.
- A recent, unresolved hack lowers the score for about six months, fading as it ages.
Fact-checks are separate. Our reviewers check each page’s claims against its sources. The result is shown as its own label, for example “5 of 9 claims confirmed, 2 need better sources”. It does not change the trust score. A claim that fails a fact-check stops counting against the subject until it is resolved.
Who does the sorting. AI agents read each page and decide, for each problem, who is responsible. A second agent checks their work, and doubtful cases go to a person. The formula itself is fixed and public, and the same for every page. Each scored page has a “Why this score” box with its own numbers.
How our scoring has changed, and what we got wrong
We think the history matters, so here it is.
The AI that researched each page also gave it a score, as a judgment.
What was wrong: No fixed rules, so similar cases could score differently.
What we did: Wrote a formula (May).
Accepted community evidence started adjusting scores by a fixed rule.
What was wrong: None found in the rule itself.
What we did: Kept.
Wrote a scoring formula.
What was wrong: It was only connected to a little-used part of the system. About 98% of pages kept the AI's own score. Meanwhile our walkthrough page described the formula as how scores were calculated. That was not true, and we did not notice.
What we did: Corrected the walkthrough page; every page now uses the formula (v2).
Added fact-checking, with an AI judge whose verdict could lower a score.
What was wrong: It lowered the subject's score for our page's sourcing mistakes.
What we did: Fact-checks moved to a separate label (v2).
Found Aave, the largest lending protocol, scored 0, the same as BitConnect.
What was wrong: Both the AI's judgments and the early formula counted hacks a project suffered as if it had done them, and gave no credit for legitimacy.
What we did: Investigated, then designed v2.
Removed a paid part of the system.
What was wrong: That also removed the little-used formula path, so every score was AI judgment until v2.
What we did: v2 replaces it.
Scoring v2, described above. Tested on every published page before anything changed. About 886 pages change band.
What was wrong: We will find more.
What we did: This page will record each change.
What changed in October 2026
- About 886 of 1,862 pages change band.
- The red band shrinks from 1,147 to about 646 pages.
- Known frauds stay at 0: FTX, BitConnect, OneCoin, Celsius, Terra.
- Some pages go down: companies with regulatory guilty pleas, people convicted of fraud, and projects that no longer operate.
- 137 pages were held for individual review before changing.
A note on Solana. We are presenting avoid.net to the Solana ecosystem, and Solana’s score rises from 62 to about 89. The rise comes from the same rule applied to every page: Solana’s page had counted scams run by other teams on its network, a hack of a separate bridge, and an investor’s collapse as if Solana had done them. Solana’s own problems still count against it: network outages, a quietly patched bug, and validator concentration. We found this flaw while checking our own page, and we are fixing it everywhere at once.
How to check us
- Every page shows when its score last changed and why. Its version history keeps every earlier version.
- Decisions are recorded on the Solana blockchain, so the history cannot be quietly rewritten (audit log).
- We will publish the full list of October 2026 changes, with each page’s old and new score.
- If you think a score is wrong, submit evidence on the page. Accepted evidence moves the score.
What we have not solved
- The “who is responsible” sorting is done by AI and will sometimes be wrong. Every change is reversible, and doubtful cases go to a person.
- Some real scams are described only as “alleged” because their sources are thin. Those pages stay above red until the sourcing improves. We would rather under-label than wrongly label.
- The weights (65 to start, the 100/40/10 shares, the caps) are judgment calls. We will publish any change to them here, with the date.