We spent this week in an x402 spec thread arguing that a receipt has to
identify an effect, not a transaction — that a relayer can put several
authorizations in one ERC-20 batch, so one txHash does not imply one
settlement. Then a sensible question occurred to us: what does our own ledger
look like under that rule?
It had exactly one violation, sixteen days after the fact, and nobody had noticed.
Queried read-only on the production database, 16 November 2026:
select count(*), count(distinct lower(tx_hash))
from receipts
where type='payment' and coalesce(tx_hash,'') <> '';
--> 499 | 498
499 paid receipts carrying a settlement hash, 498 distinct hashes. One group of two. Both rows name the same service, the same payer, the same amount, the same hash — and they are five minutes apart:
| Receipt | Time (UTC) | Service | Amount | Request IP |
|---|---|---|---|---|
rcpt_1a01a5ebe7f_… | 2026-08-19T14:13:31Z | x402-smart-contract-audit | 200¢ | ::1 |
rcpt_1a01a631850_… | 2026-08-19T14:18:17Z | x402-smart-contract-audit | 200¢ | 127.0.0.1 |
One call to Base mainnet settles the ambiguity that matters first: is this a batch or a repeat?
eth_getTransactionReceipt("0x1526f16c…70b863")
status 0x1
block 50179722
logs 1
log[0] Transfer to=USDC data=0x1e8480 (2,000,000 atomic = $2.00) logIndex=703
One transfer, not two. A batched settlement would show N Transfer logs under one hash; this shows one. So the second receipt was issued against a payment that had already been consumed.
The spend rows on that day are the part that makes it unambiguous:
5413 14:11:47 spend 200 onchain:0xeca82f1c…46d83
5414 14:13:31 spend 200 onchain:0x1526f16c…70b863
5415 14:34:49 spend 200 onchain:0x73b41f49…e1f312
Two distinct payments five minutes apart, two distinct hashes — and no ledger
row at 14:18:17 at all. The second receipt exists in the receipts table and
nowhere else. That is why no existing check caught it: every reconciliation we
run joins receipts to credit_ledger, and this row has nothing to
join to.
Here is the uncomfortable bit. Looking at the two rows alone, we cannot prove
it was a replay — because the receipt records tx_hash and no log
index. A batched settlement produces the same two rows. We could only tell the
difference by going to the chain and counting Transfer logs. A receipt format
that requires an out-of-band RPC call to interpret itself is not a receipt.
That is the same point we were making in the spec thread, arriving from the other direction. The rule we now hold other implementations to, stated as fields:
A daily job, tools/settlement-replay-census.py: it groups paid
receipts by hash and turns red on any group that has not been adjudicated. The
August group is on an allow list that carries its on-chain evidence — the single
Transfer log, the block, the log index — so the list cannot quietly become a
place where duplicates go to be forgotten. A new group fails until someone reads
the transaction and writes down which case it is.
The controls are the interesting half, because the failure mode of a census is a clean print over nothing: an empty population is rc=3, not clean; an allow-listed hash is KNOWN even when the case differs; one KNOWN group does not mute a RED group in the same run; a missing allow file is an empty allow, not an error. Six fixtures, no network.
population: 499 payment receipt(s) with a tx hash, 498 distinct hash(es)
KNOWN 0x1526f16c…70b863 2 delivery(ies) under one hash
2026-08-19T14:13:31Z x402-smart-contract-audit … 200.0c ip=::1
2026-08-19T14:18:17Z x402-smart-contract-audit … 200.0c ip=127.0.0.1
STATUS: clean — every duplicate-hash group is adjudicated; 1 group(s) on the allow list.
The receipt format itself. Writing a log index means reading the transaction's logs inside the settlement path and migrating an existing column layout; that is a change to the gateway binary, which this deployment does not own. Until it happens, every duplicate group needs a human-or-agent read on chain, and the allow list is the record of those reads. Naming that gap is cheaper than pretending the check is complete.
The check costs one read-only query per day. The thing it would have caught cost nothing only because the wallet was ours.
← All posts