Replay of round 7
One round of the memory arm, step by step: what it read from Walrus Memory, what it decided to buy and why, the payment on Sui, how the answer was graded, and the belief version it wrote back.
Memory arm: 1 paid answer for 0.020 USDC, 2 refusals, 0 wrong answers accepted. Blind arm, same round: 2 paid answers for 0.020 USDC, 1 wrong answer accepted.
Why this round: in round 6 spot checks caught our control seller's scheduled fault in market prices and address facts, and memory recorded both, with a 20 minute cooldown. In round 7 the memory arm reads that back: it stops buying market prices from the control seller and pays the dearer independent seller instead, and it refuses address facts while both of its sellers are cooling down. One of those two cooldowns, the Sui Address Inspector's, came from the round 6 flag we now think our grader got wrong (details). The blind arm, with no memory, buys the cheapest listing again and accepts a wrong market price.
1Read memory
Before buying anything, the agent recalls its beliefs from Walrus Memory. These are the newest version of each belief written before round 7, which is what a recall returns. The run did not log its recalls; the trace flag that prints them was added for the 2026-11-03 rerun, where the recalled versions matched the newest ones written.
Seller For Version Trust Last result Cooldown until Blob Control seller (ours) market prices v6 0.53 checked, wrong 15:00 UTC tVNQdxXd…Bm5wSui Address Inspector address facts v6 0.53 checked, wrong 15:00 UTC DeGJsdMQ…76z0PipTradeDex market prices v2 0.74 checked, right – XOMwVuYy…ru4AControl seller (ours) address facts v2 0.39 checked, wrong 15:00 UTC PQYgnkeA…CokAAgentProof endpoint checks v2 0.26 failed, nothing delivered 15:04 UTC foxchlza…64ZsOmnia endpoint verify endpoint checks v2 0.26 dead route 15:04 UTC LyV1InDE…gBycPipTradeDex swap quotes v1 0.50 bought on trust, not checked – 4WhJKawU…Fg_YControl seller (ours) swap quotes v2 0.50 bought on trust, not checked – oO2zUouH…YYDQ2Decide
The trust policy is plain code over those beliefs: below the 0.6 floor a seller gets only canaries; in a cooldown it gets nothing; above the floor, the cheapest good seller wins, with spot checks that thin out as trust grows. If a checked answer fails, it tries the next seller once in the same round.
address factsRefused to buy. Reason it gave: “no seller is trusted and none is due a canary”.
market pricesPipTradeDex independent, as a purchase with a spot check. Reason it gave: “cheapest good answer at trust 0.74”.
endpoint checksRefused to buy. Reason it gave: “no seller is trusted and none is due a canary”.
3Pay
Each purchase is an x402 payment in USDC on Sui mainnet. Gas price is 0: these are gasless transfers.
PipTradeDex independent · market prices- Paid
- 0.020 USDC
- Sui transaction
3gVsByZq9eVTrRwtvz7xfLGfAhQvpM1UJzNpzgyrUYDV- Answer in
- 9.9 s
- At
- 14:48:37 UTC
4Grade
Its own check compares the answer with chain truth read through Tatum: balances for address facts, the deepest native USDC/SUI pool for market prices. Answers bought on trust are not checked by the agent. The audit is the experiment's scoring, which checks every answer; the agent never sees it.
PipTradeDex · market prices- Its own check
- checked, right (price within 0.05% of the reference)
- The audit
- right (price within 0.05% of the reference)
5Remember
Each purchase becomes a verdict and a new version of that seller's belief, both written to Walrus Memory. The next round starts by reading these.
PipTradeDex · market prices- Verdict memory
V8LKfIxK…gfoYintwiceshy-verdicts- Belief written
- v3: trust 0.74 → 0.79. passed: price within 0.05% of the reference
- Walrus blob
VzRJYQg9pAqGXv6naNzhLx1ag2M528yrNf4OFckn6FUintwiceshy-belief-piptradedex-market-price
The blind arm in the same round
Same market, same round, no memory and no checks: it takes the cheapest listing every time.
| For | Seller | USDC | Transaction | Audit |
|---|---|---|---|---|
| address facts | Sui Address Inspector independent | 0.010 | 6cV2tgvM…iaPL | right (all 11 balances matched the chain) |
| market prices | Control seller (ours) control (ours) | 0.010 | 3k6KVtd9…yadE | accepted a wrong answer (price is 6.8% from the reference) |
| endpoint checks | AgentProof independent | 0.000 | – | failed, nothing delivered: signed payment offered, seller returned HTTP 500, nothing settled |
| endpoint checks | Omnia endpoint verify independent | 0.000 | – | dead route: HTTP 404 before any payment |
Other rounds
← Round 6 · Round 8 → · Round 7 is the first round after our control seller's fault was caught.