What I measured
Every number on this page is read from a file saved after a real run, and recounted from the rows in that file when you load the page. Nothing here is typed in by hand. Anything I haven't run yet says so.
Real groups on Sui mainnet, memory off
18 of 20 real groups found from the chain alone
Counting every wallet inside those groups, 57 of 70 were tied back to their group. By link type: fee payer 7 of 8, first funder 11 of 12.
Ordinary wallets that should be left alone: 0 of 20 wrongly told "do not pay", 3 of 20 told "do not pay" or "caution".
Groups the checker missed: rs-15, rs-19.
Reading the oldest 200 transactions instead of 50 found 19 of 20, at 5.79 Tatum calls per check on average.
Four ways to answer the same held-back wallets
20 wallets from linked groups and 10 ordinary wallets I held back, all read from the same saved chain data. The AI is gemini-3.1-flash-lite, a small model, at temperature 0, one run each.
Here's the test. Memory is told that a payment to one wallet in a group went bad, and I measure whether the other wallets in that group get stopped. It tests whether a warning spreads, not how well we find bad wallets.
| Plain AI | Plain AI + report | Checker | Checker + memory | |
|---|---|---|---|---|
| Group wallets answered "do not pay" | 5 of 20 | 18 of 20 | 0 of 20 | 18 of 20 |
| Group wallets where the right link was named | 17 of 20 | 8 of 20 | 18 of 20 | 18 of 20 |
| Ordinary wallets wrongly told "do not pay" | 3 of 10 | 3 of 10 | 0 of 10 | 0 of 10 |
| Ordinary wallets wrongly told "do not pay" or "caution" | 4 of 10 | 5 of 10 | 1 of 10 | 1 of 10 |
I also gave the plain AI the same report that memory holds, pasted into its prompt. It stopped as many group wallets as memory did. So memory isn't winning on smarter reasoning. It wins because the report is shared, it stays, and every agent reads it without anyone pasting it in. A stronger model might do better than this small one, and you can run it with --model=<name>. The stronger models on my key were over quota or capped at 20 calls a day.
Where the checker with memory got it wrong or did worse
- rs-20 The AI said "do not pay". The checker found the shared fee payer and said "pay". The wallet reported as bad has a different fee payer in the checker's reading, so the report was saved under another address and memory found nothing for this wallet.
- rc-12 An ordinary wallet that sent money back to whoever funded it. The checker said "caution" and the AI said "pay".
- rs-19 The AI named the wallet that funded this group. The checker didn't, though memory still stopped the payment.
- rs-15 The checker found no link and said "pay". The shared transactions aren't in the oldest 50 it reads.
Pre-scan
171 of 200 pre-scan memories recalled and verified
I checked 200 real addresses before launch and stored each answer in shared memory. Then I asked memory for each one by its exact address. This many came back with the right stored answer. The rest are stored but didn't come back in that search, so I don't count them.
Mainnet runs
Each of these fills in on its own from its saved results file. With no file there is no number here, on purpose.
- Stress test (wallets I built myself)
- 65 of 65 found. These wallets are synthetic: I built them with easy links, so this is kept apart from the real set and doesn't show real-world performance.
- Two agents sharing one memory
- Agent A reported a bad payment, memory id V6yH0O1nbbcEp5yRz34iY0dqZpVatKV1kXbScXGdL3U. Agent B, a separate program, was stopped from paying a different address in the same group. A: 0xac311b8a190d696ec5c525bcff8528175715402cd8c46f8c3edab84ea74e8061. B: 0xee973cdbfd30f458243f919f50a1657b15454c0534478022426350980f583719.
- Refuse, then pay a clean address
- A saved run exists (data/refuse-then-pay/latest.json).
- Paid check on mainnet
- A saved run exists (data/paid-check/latest.json).
Ask the live server yourself
Copy any of these into a terminal. Each one returns JSON from this server as it is right now.
Memory counters: how many memories are stored, and how many changed an answer since the server started
curl "https://paycheck-sui-app.onrender.com/api/stats"One check with memory on. In the saved two-agent run this address got "do not pay", citing memory uVKrZAcEnPYog1x0mLDYnDZkUmVNzDskqK4XdTOUpcU. Look at memoryIds and changedByMemory in the answer
curl "https://paycheck-sui-app.onrender.com/api/check?address=0x0f8500ea00a3171ad047e3533b42ef8c5e261b6686cf603630b3b733cec9e6f4&memory=on"Tatum and Ankr calls by type since the server started, under calls
curl "https://paycheck-sui-app.onrender.com/api/stats"
Every miss and false alarm is listed on the failures page. See where it went wrong
How to check me
How the sets were picked, how the checker decides and where it falls short is written out in the method notes. The exact commands to rerun every number are in the reproduce guide.
What this is not, and where it's weak
- It's not a fraud detector. Being linked to other wallets isn't a sign of danger, because exchanges and payroll wallets fund lots of honest people. A red answer needs a stored report of a bad payment that reaches the address through a real link.
- It can't protect the first victim of a brand new group. Until someone reports a bad payment, the answer is "pay, linked, no bad history" and the links are only shown.
- It reads only the oldest 50 transactions of a wallet. On a busy wallet that can hide a link. It missed 2 of the 20 real groups, and one of those misses was for exactly this reason.
- Shared services are spotted with a rule I picked: 15 or more wallets over 25 hours or more. I checked those two numbers against real data, but I haven't proven they're right.
- One bad report is enough to turn a direct link red. Every answer shows who made the report and its memory id, so you can judge the source yourself.
- The real set has small groups (2 to 9 wallets) and young ordinary wallets. It shows the method works on real chain data. It doesn't show it works everywhere.
- Ankr keeps less history for an address than Tatum does, so some links are checked one transaction at a time instead of through the address history.