Verified Sui Mainnet run · 8 November 2026
Memory changed the decision. Reserves held.
A real Codex agent bought two live ChainScope reports with native USDC. One locally signed owner approval carried into a fresh session through Walrus Memory. A third purchase was refused.
Recorded 13:03–13:10 UTC on Sui Mainnet. This run is complete and permission is revoked. Connect reads the same Mainnet vault and recorded requests; new spending is closed.
- Historical purchases (consumed)
- 1.60 USDC paid
- Commitment reserve (protected)
- 0.80 USDC reserved
- Current vault balance
- 1.40 USDC (0.60 free) · agent revoked
One approval, two purchases, one hard refusal
01 · Agent requests
A useful report, held for review
The owner deposits 3.00 USDC: 0.80 for savings, 2.20 free. Codex discovers ChainScope (a project-operated merchant returning live Sui Mainnet data) and requests a live report on balances, commitments and spending limits.
Request req-d89525e6-87f waits because the merchant is unfamiliar. No payment occurs.
02 · Owner decides
Approve once and remember the preference
The owner signs locally: allow this exact 0.80-USDC purchase and remember ChainScope up to 0.80 for seven days. The vault releases free funds; x402 settles the invoice. Codex uses the delivered live report.
Release receipt · x402 settlement · Stored Walrus preference
03 · Fresh session
The recalled rule changes the next decision
A new application session rebuilds one rule from Walrus. A fresh Codex conversation requests req-7dd90263-75e; allow_rule settles the second 0.80 report without another owner approval.
The purchased report returns live checkpoint 343007621: 1.40 total, 0.80 reserved and 0.60 free, plus active spending limits.
Release receipt · x402 settlement · Stored decision
04 · Contract boundary
The remembered preference cannot spend savings
Request req-f37bacb5-c6a costs 0.80 USDC, but only 0.60 is free. It is refused with would_cut_reserve while the preference remains active. Nothing is paid.
An independent unsigned simulation using the real operator capability returns EExceedsFree. This refusal was simulated; no failed transaction was broadcast.
05 · Owner closes the run
Permission revoked, reserve intact
The owner revokes spending permission. Independent Tatum read-back confirms 1.40 USDC remains: 0.80 reserved + 0.60 free. Merchant proceeds are 1.60 USDC. Net gas across all eight transactions is 0.083418818 SUI.
Owner revoke receipt
Memory with inspectable provenance
Four Mainnet memories in this run reached stored status: the preference, two decisions and a session summary. Each relayer job matched its blob, and all four appeared in the full account inventory.
Inspect the remembered rule · Inspect the fresh-session decision
This run has four stored memories. Earlier project runs independently verified additional Mainnet memories; their original networks and provenance remain in the repository archive.
All eight receipts and verification details
Every receipt was read independently through Tatum, matching sender, gas owner/budget, asset, amount, merchant recipient, vault events and function calls. An unrelated successful Mainnet digest failed the same verifier.
Source archive: docs/evidence/mainnet-20261108. The archive includes actual tool calls, useful merchant output, exact gas breakdowns, memory jobs/blobs, the unsigned reserve simulation and verifier negative control.
The read-only MCP response was improved after payment to include the recalled rule's source link. A follow-up verified provenance without another purchase and corrected the agent's initial confusion between the 0.50 default threshold and the 0.80 merchant preference. Original transcripts are retained.
What this proof establishes
One stored approval changed a later decision; hard reserve limits still prevailed. ChainScope is operated by this project, so these are project payments, not customer revenue. The vault is custom Move code with test evidence; no third-party audit is claimed.
Codex CLI completed this Mainnet task. Owner actions used the CLI with a local key. Owner-assistant model behavior and Antigravity have not been recorded. This experiment does not establish broad demand or superiority to every static spending policy.