← TwiceShy proof pageAll belief versionsSource

Replay of round 4

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: 3 paid answers for 0.030 USDC, 1 refusal, 0 wrong answers accepted. Blind arm, same round: 3 paid answers for 0.030 USDC, 0 wrong answers accepted.

  1. 1Read memory

    Before buying anything, the agent recalls its beliefs from Walrus Memory. These are the newest version of each belief written before round 4, 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.

    SellerForVersionTrustLast resultCooldown untilBlob
    Control seller (ours)market pricesv30.79checked, right–ci5iJHg-…EChE
    Sui Address Inspectoraddress factsv30.79checked, right–DGZ2RyxL…2JuI
    PipTradeDexmarket pricesv10.67checked, right–uuhQjYkE…7AeQ
    Control seller (ours)address factsv10.67checked, right–hwBjNu88…6tp0
    AgentProofendpoint checksv10.33failed, nothing delivered–VWRgmj1_…bJcY
    Omnia endpoint verifyendpoint checksv10.33dead route–IFawCml9…XYY4
    Control seller (ours)swap quotesv10.50bought on trust, not checked–nSftyzPp…KYrc

    Next: Decide ↓

  2. 2Decide

    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. Swap quotes it cannot check itself, so it buys them only from a seller its market-price belief trusts.

    address facts

    Sui Address Inspector independent, as a purchase on trust. Reason it gave: “cheapest good answer at trust 0.79”.

    market prices

    Control seller (ours) control (ours), as a purchase on trust. Reason it gave: “cheapest good answer at trust 0.79”.

    endpoint checks

    AgentProof independent, as a canary (a question it can check, to earn trust). Reason it gave: “canary: nobody is trusted yet”.

    Omnia endpoint verify independent, as a canary (a question it can check, to earn trust). Reason it gave: “canary: earning a place”.

    Refused to buy. Reason it gave: “no seller is trusted and none is due a canary”.

    swap quotes

    Control seller (ours) control (ours), as a purchase on trust. Reason it gave: “cheapest good answer at trust 0.79”.

    Next: Pay ↓

  3. 3Pay

    Each purchase is an x402 payment in USDC on Sui mainnet. Gas price is 0: these are gasless transfers.

    Sui Address Inspector independent · address facts
    Paid
    0.010 USDC
    Sui transaction
    FwKR4kF498L35RBxgRucHY8NCPCJpiYuoWG8Lq6AQpnd
    Answer in
    1.0 s
    At
    14:24:24 UTC
    Control seller (ours) control (ours) · market prices
    Paid
    0.010 USDC
    Sui transaction
    3MWDFosXYUTaB11BHhsmuNTdYfs4nxsjbhw2hcmfDJJF
    Answer in
    1.0 s
    At
    14:24:29 UTC
    AgentProof independent · endpoint checks
    Paid
    0.000 USDC (nothing settled)
    Answer in
    0.3 s
    At
    14:24:49 UTC
    Omnia endpoint verify independent · endpoint checks
    Paid
    0.000 USDC (nothing settled)
    At
    14:24:50 UTC
    Control seller (ours) control (ours) · swap quotes
    Paid
    0.010 USDC
    Sui transaction
    D2fjztCbDPzCPK9cypvowf3KwahbMqsQ18P292dKpzxc
    Answer in
    1.0 s
    At
    14:24:55 UTC

    Next: Grade ↓

  4. 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.

    Sui Address Inspector · address facts
    Its own check
    bought on trust, not checked
    The audit
    right (all 5 balances matched the chain)
    Control seller (ours) · market prices
    Its own check
    bought on trust, not checked
    The audit
    right (price within 0.04% of the reference)
    AgentProof · endpoint checks
    Its own check
    failed, nothing delivered (signed payment offered, seller returned HTTP 500, nothing settled)
    The audit
    no answer to audit
    Omnia endpoint verify · endpoint checks
    Its own check
    dead route (HTTP 404 before any payment)
    The audit
    no answer to audit
    Control seller (ours) · swap quotes
    Its own check
    bought on trust, not checked
    The audit
    right (implied price within 0.02%)

    Next: Remember ↓

  5. 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.

    Sui Address Inspector · address facts
    Verdict memory
    3DY2p7a7…KVZM in twiceshy-verdicts
    Belief written
    v4: trust 0.79 → 0.79. bought unchecked on a trust of 0.79
    Walrus blob
    JIBXJqULMI_AFg5NrAfUff-NNmOXbBPjm5oluLaSuu8 in twiceshy-belief-inspector-address-facts
    Control seller (ours) · market prices
    Verdict memory
    u9PfpPlt…CJ-s in twiceshy-verdicts
    Belief written
    v4: trust 0.79 → 0.79. bought unchecked on a trust of 0.79
    Walrus blob
    _ULXcntIQRy3BjRHZPYS1JSB7EK6j3amhlFZzjixrF8 in twiceshy-belief-control-market-price
    AgentProof · endpoint checks
    Verdict memory
    xWjm0th4…Nt44 in twiceshy-verdicts
    Belief written
    v2: trust 0.33 → 0.26. failed: signed payment offered, seller returned HTTP 500, nothing settled; retest after 40 min
    Walrus blob
    foxchlzawXdUSQ1fMGD3O5KAVpYtvi58ZnCrQB-64Zs in twiceshy-belief-agentproof-endpoint-check
    Omnia endpoint verify · endpoint checks
    Verdict memory
    vcEh6A9C…LUdY in twiceshy-verdicts
    Belief written
    v2: trust 0.33 → 0.26. dead: HTTP 404 before any payment; retest after 40 min
    Walrus blob
    LyV1InDE4OT94dl_-_zJc47cqDTkJ5VmocKkxUVgByc in twiceshy-belief-omnia-endpoint-check
    Control seller (ours) · swap quotes
    Verdict memory
    LYoUyWbg…fFxw in twiceshy-verdicts
    Belief written
    v2: trust 0.50 → 0.50. bought unchecked on a trust of 0.50
    Walrus blob
    oO2zUouHCrMplIcn7-dmfQj4czEMHY-0KUALMNTYYDQ in twiceshy-belief-control-swap-quote

    Round 5 reads these back →

The blind arm in the same round

Same market, same round, no memory and no checks: it takes the cheapest listing every time.

ForSellerUSDCTransactionAudit
address factsSui Address Inspector independent0.010BAqeQzSW…tX5Wright (all 1 balances matched the chain)
market pricesControl seller (ours) control (ours)0.0109DfkWX6C…S3tbright (price within 0.10% of the reference)
endpoint checksAgentProof independent0.000–failed, nothing delivered: signed payment offered, seller returned HTTP 500, nothing settled
endpoint checksOmnia endpoint verify independent0.000–dead route: HTTP 404 before any payment
swap quotesControl seller (ours) control (ours)0.010EiuSs5EZ…qC1aright (implied price within 0.10%)

Other rounds

← Round 3 · Round 5 → · Round 7 is the first round after our control seller's fault was caught.