TwiceShy

A buying agent for paid APIs that remembers who burned it. It pays for answers over x402 on Sui mainnet, spot checks them against chain truth read through Tatum, and keeps its verdicts on Walrus Memory. Its next purchase changes because of what it remembers.

Its memory is portable: Walrus Memory is its only long-term state, so it picks up where it left off in a new session. A fresh clone with only keys recovered all 8 beliefs and made the same decisions (wipe and recover).

Snapshot 2026-11-01T16:59:43.861Z

The market has no memory

The public route directory lists 46 pay-per-call routes on Sui, and marks 46 of them as passing its probe. We sent each one a single unpaid POST and recorded the reply. 19 of 46 (41%) did not answer with a 402 challenge. Measured 6 times, 2026-11-01 12:49 to 2026-11-01 16:50 UTC. A route that fails this probe may still work for other callers or at other times; this is what we saw, and paying for a route is the only way to learn more.

Memory on versus memory off

Two buyers shopped the same live market in the same rounds, with real USDC. The experiment audited every answer in both arms. The plan and the fault schedule were committed before round 1.

The control seller is run by this project and turns bad on a published schedule. It is never counted as an independent seller, and payments to it go to our own wallet.

We expected memory to be cheaper per correct answer, and it was not. It paid about 20% more per correct answer and 83% more in total, for canaries and then for a dearer honest seller. It got 18 more correct answers (42 against 34 of 45 distinct questions), accepted none wrong, and was never charged for nothing. The zero depends on timing: a fault starting a round or two earlier would have meant 1 to 3 wrong answers accepted.

A flaw in our grader. Three answers from an independent seller failed the audit. The one we can inspect was 0.0008% from the chain, and we believe the wallet moved. The preregistered rule counted them as wrong, so they stay in these numbers, and they cost the memory arm two rounds of refusing to buy. Details and the fix.

Read the 0 against 11 with care. 9 of the blind buyer's 11 accepted wrong answers came from our own control seller's scheduled faults. Independent sellers only: memory bought 36 answers, 35 correct, none accepted wrong; blind bought 18, 16 correct, 2 accepted wrong (both the address flags above). If all three flags were grader errors the score is 53 against 36 correct and 0 against 9 wrong accepted, all 9 from our control seller; only one flag can be inspected, so the preregistered score stays the result. One run, a small sample. The verifier checks 108 payment digests: 104 in the experiment plus 4 from the one live round after it.

What it learned

Each line is a trust score the agent read back from Walrus Memory at the start of the next round. The shaded band is the control seller's published fault window. Below the dashed floor, a seller only gets canaries until its score climbs back.

How one belief is versioned

Every revision is a new memory on Walrus, in a namespace of its own for that seller and kind of work. This is the agent's belief about our control seller for market prices: canaries earn trust, checks thin out, the fault is caught and a cooldown starts, the retest fails and the cooldown doubles, then canaries earn it back. All 69 belief versions with their blob ids: belief-versions.csv.

Belief revisions

Replay a round

Drag through the run to see what each buyer did, and why. For one round in full, with what it recalled from Walrus, each payment and the belief version it wrote: replay round 7 step by step.

After the experiment

The agent ran 1 more live round after the experiment and spent $0.040. Live rounds ended on 2026-11-01. That round marked two sellers wrong because of a grader flaw, since fixed: they listed only the first 50 coins of very large wallets. No later round ran to correct those verdicts, so the last beliefs below still carry them. The fix has unit tests but has never run in a live round.

Wipe and recover

A fresh clone with only keys and no local state recovered 8 beliefs from Walrus Memory at 2026-11-01T16:59 UTC. Beliefs match the last versions written: True. Next decisions identical: True.

AgentProof / endpoint-check: trust 0.18, version 4
Control seller (ours) / address-facts: trust 0.47, version 8
Control seller (ours) / market-price: trust 0.66, version 12
Control seller (ours) / swap-quote: trust 0.50, version 4
Sui Address Inspector / address-facts: trust 0.62, version 17
Omnia endpoint verify / endpoint-check: trust 0.18, version 4
PipTradeDex / market-price: trust 0.85, version 14
PipTradeDex / swap-quote: trust 0.50, version 6

Next decisions:
address-facts: refuse: no seller is trusted and none is due a canary
market-price: buy from Control seller (ours): cheapest good answer at trust 0.66
endpoint-check: refuse: no seller is trusted and none is due a canary
swap-quote: buy from Control seller (ours): cheapest good answer at trust 0.66

Run again on 2026-11-03

Two days later we cloned the repo again, with no state directory, and ran recover with TWICESHY_TRACE=1, which prints each Walrus recall. Eight recalls brought back the same eight beliefs, equal in every field to the last versions written. Two decisions differ from the first run because cooldowns stored in memory had ended: address facts would be bought from the Sui Address Inspector again, and AgentProof is due a canary. Full transcript.

walrus recall twiceshy-belief-inspector-address-facts: 5 hits in 5080 ms -> v17 trust 0.62
walrus recall twiceshy-belief-control-address-facts: 5 hits in 3186 ms -> v8 trust 0.47
walrus recall twiceshy-belief-piptradedex-market-price: 5 hits in 2264 ms -> v14 trust 0.85
walrus recall twiceshy-belief-control-market-price: 5 hits in 1960 ms -> v12 trust 0.66
walrus recall twiceshy-belief-agentproof-endpoint-check: 4 hits in 2538 ms -> v4 trust 0.18
walrus recall twiceshy-belief-omnia-endpoint-check: 4 hits in 2495 ms -> v4 trust 0.18
walrus recall twiceshy-belief-piptradedex-swap-quote: 5 hits in 1830 ms -> v6 trust 0.50
walrus recall twiceshy-belief-control-swap-quote: 4 hits in 2910 ms -> v4 trust 0.50

Receipts

104 settled payments in the experiment, 54 of them to independent sellers and 50 to our own control seller, plus 4 in later live rounds. Gas price is 0: these are gasless USDC transfers. All digests as a file: digests.csv.

Memories on Walrus

142 memories on Walrus mainnet (4 census summaries, 69 verdicts, 69 beliefs). The agent recalls its beliefs every round; verdicts and censuses are its record. Each blob is Seal encrypted. Open a blob on the public aggregator to confirm it exists.

Check it yourself

One read-only command, with no keys and no install, recomputes the scores from the per-round ledger with the preregistered rules, looks up every payment digest on Sui mainnet through the public GraphQL endpoint, and fetches every memory blob from a public Walrus aggregator. It needs Node 18 or later.

git clone https://github.com/swiftbotlabs/twiceshy && cd twiceshy
node scripts/verify.mjs
# or check this page's own data:
node scripts/verify.mjs --source https://apps.swiftbotlab.workers.dev/twiceshy/data/snapshot.json

Our run on 2026-11-03: memory 52 correct and 0 wrong accepted, blind 34 and 11, as published; 108 of 108 digests settled as logged (104 in the experiment: 54 to independent sellers, 50 to our control wallet); 142 of 142 blobs served, none in plaintext; 8 of 8 beliefs with versions 1 to N and no gaps.

The raw files: every payment digest · memory arm ledger · blind arm ledger · run log, both arms · belief versions and blob ids · snapshot.json.

How to read this

This is one run with a small sample, and the fault is synthetic. It shows the mechanism working on live mainnet, not numbers that generalize. Address facts are graded against Sui balances read through Tatum gRPC. Market prices are graded against the live price of the deepest native USDC/SUI pool, also through Tatum. Memory contents are Seal encrypted: a blob link proves the write, and the sentences here are what the agent wrote.