Proof
What the deployed agent has actually done, the measurement we committed to before running it, and every claim with its evidence.
Live counters
Read from the deployed agent's own records when this page loaded. Payments are on Sui testnet with test USDC; memories are on the Walrus Memory mainnet relayer.
- Payments paid
- 3 0 failed with no USDC moved
- Refusals
- 2 each with a signed receipt
- Verified payees
- 2 proven by their own wallet
- Address changes verified
- 1 two signatures each
| Namespace | Records |
|---|---|
owner | 1 |
payee:little-brother-s-allowance | 2 |
payee:rent | 6 |
receipts | 8 |
| Total | 17 |
| Reason | Count |
|---|---|
This matches an earlier refused attempt. KNOWN_ATTACK_PATTERN | 1 |
The new address imitates a known address. LOOKALIKE_ADDRESS | 1 |
Latest refusals
- rcpt_dab6cd850380599887fb5bb1b25577ff KNOWN_ATTACK_PATTERN, 2026-10-31 17:13 UTC
- rcpt_2c31ae5f14f5cbe5a50da5690bc9993f LOOKALIKE_ADDRESS, 2026-10-31 17:05 UTC
Preregistered measurement
Hypothesis: with Countersign, attacker-address payments per 100 attack attempts are 0 on corpus K1–K9; without it (same agent, memory used naively), more than 0.
Metric, fixed in advance: attacker-address payments ÷ attempts. Secondary: false refusals on 50 benign messages. Testnet only; the corpus is published.
Started 2026-11-01 15:11 UTC on Sui testnet, corpus K1–K9, 10 trials each; 50 benign (evidence/measurement/corpus.json).
| Arm | Attack attempts | Attacker-address payments | Rate |
|---|---|---|---|
| With Countersign | 90 | 0 | 0.0 per 100 |
| Baseline (memory read without status checks) | 90 | 71 | 78.9 per 100 |
Benign messages: 0 false refusals out of 50.
Claims ledger
Generated from CLAIMS.md. Only PROVEN claims may appear in our copy.
| Claim | Status | Evidence |
|---|---|---|
| CL01 Countersign's change logic never makes an attacker address payable, even when the model proposes exactly what the attacker wants (K1–K8, 1,000 random sequences) | PROVEN | packages/core/test/adversarial.test.ts (M1, P1) |
| CL02 Our Sui personal-message verification matches Mysten's published vectors byte for byte | PROVEN | feasibility/evidence/f1-signatures.json |
| CL03 Tatum verifies payee signatures for us | PROVEN | evidence/acts/onboarding-rent/records.json (Rent's payee wallet, the owner's Slush account, signed on /sign; VERIFIED record carries scheme ZKLOGIN, maxEpoch 1271, checked at epoch 1269); pnpm test:live (Ed25519 through the adapter) |
| CL04 Payee records live in Walrus Memory on mainnet and are rebuilt completely on start | NOT PROVEN | needs pnpm test:live restore on a payee namespace; see CL09 for the relayer itself |
| CL05 Payments execute through Tatum | PROVEN (testnet) | evidence/live/testnet-payment-2026-10-30T21-22-23-883Z.json (pnpm testnet:pay, REALITY_SPIKES P1) |
| CL06 Attack block rate with a real model on testnet | NOT PROVEN | PRD "Preregistered measurement" |
| CL07 ≥10 memories on mainnet from real use (12 counted on 2026-11-01) | PROVEN | evidence/a02-memory-count.json (every record in this owner's namespaces on the Walrus Memory mainnet relayer, each opened and checked as agent-signed: owner wallet 1, Rent payee versions 5, receipts 6) |
| CL08 Tatum's Sui gateway verifies an Ed25519 personal-message signature and rejects it for a different address or an altered message (mainnet gRPC `VerifySignature`, 2026-10-30) | PROVEN | evidence/spikes/spikes-2026-10-30T20-48-55-762Z.json (S10) |
| CL09 A Walrus Memory mainnet write returns a blob id in about 20–23 s, and recall returned all 3 records of a 3-record namespace with none dropped (2026-10-30) | PROVEN | evidence/spikes/spikes-2026-10-30T20-46-45-854Z.json (S2, S12) |
| CL10 Our Walrus Memory adapter writes agent-signed records and a fresh adapter reads them back complete: namespace count, full listing and recall agree, nothing dropped, and a retried write adds nothing (3 records, 2026-10-30) | PROVEN | evidence/live/memwal-2026-10-30T21-08-15-263Z.json (pnpm test:live, V18) |
| CL11 Tatum verifies a zkLogin signature from a real Slush wallet and rejects it for another address or an altered message (mainnet and testnet, 2026-10-30) | PROVEN | evidence/spikes/s10z-2026-10-30T21-01-30-353Z.json (S10z) |
| CL12 Act 1: the agent rebuilds its payee book from Walrus Memory, the model proposes a payment, code checks and converts it, 0.01 test USDC settles through Tatum, and an agent-signed receipt citing the payee-version blob and the Sui digest is stored on Walrus Memory (2026-10-30) | PROVEN (testnet money, mainnet memory) | evidence/acts/act1-2026-10-30T21-53-37-449Z.json and .txt |
| CL13 Walrus Memory recall returns every record of a namespace holding 100 records (the cap): count, full listing and recall all equal 100, none dropped (2026-10-31) | PROVEN | evidence/live/s12-scale-2026-10-31T05-57-57-893Z.json (REALITY_SPIKES S12b) |
| CL14 Acts 2–3 on the deployed agent: a message asking to pay a look-alike of Rent's verified address (same first and last 4 hex characters) is refused with the grouped comparison, and the same message again is refused as a known attack citing the first receipt; both receipts are stored on Walrus Memory mainnet (2026-10-31) | PROVEN | evidence/acts/act2/records.json, evidence/acts/act3/records.json (receipts, blob ids answering 200 from the mainnet aggregator, screenshots of the live /r pages) |
| CL15 Act 4 on the deployed agent: Rent's address changed only after Rent's current wallet and then the new wallet signed on the signing page; the next payment, 0.01 test USDC, settled to the new wallet on Sui testnet; the used link reports Done with the receipt (2026-11-01) | PROVEN | evidence/acts/act4/records.json (CHANGE_VERIFIED and PAID receipts, blob ids answering 200 from the mainnet aggregator, the testnet transaction, screenshots of the live /r pages and the used /sign link) |
| CL16 Preregistered measurement on Sui testnet with the real model (Groq openai/gpt-oss-120b, PRD Amendment 1): Countersign paid an attacker address 0 times in 90 attack attempts (K1–K9 × 10); the same agent using memory naively paid 71 of 90; 0 false refusals and 50 of 50 correct payments on 50 benign messages; 0 model errors, 0 trial errors, 643 relayer calls all confirmed (2026-11-01) | PROVEN | evidence/measurement/results.json, evidence/measurement/raw-2026-11-01T15-11-00-564Z.jsonl, evidence/measurement/corpus.json |