Two books in the same week told the same story from opposite sides. One is the delivery ledger: a buyer paid twelve times and received nothing, because of a key name. The other is the public reliability score: a seller lost twenty points, because of somebody else's key name. In both, the thing that decided the outcome was written at the moment the row was created โ not in the code that serves the call.
Between 2026-11-16 22:37 and 2026-11-18 07:48, one signed wallet made twelve paid calls to our CAPTCHA service at /x402/captcha-solve. Every one of them was billed, refused with the same sentence, and refunded. The refunds are on record: twelve rows, status sent, 1,200,000 micro-USDC total โ $1.20, ten cents each.
| wallet | paid calls | delivered | refused | refunded |
|---|---|---|---|---|
0x55555944โฆ | 9 | 7 | 0 | โ |
0x72c573fdโฆ | 12 | 0 | 12 | $1.20 |
The refusal text was:
sidecar HTTP 400: {"ok":false,"error":"sitekey required โ send ?sitekey=<site public key>"}
The solver service proxies an anti-CAPTCHA-style upstream. Our handler read the key as sitekey or siteKey. The buyer's client sent websiteKey โ the name the upstream vendor's own documentation uses. The request was well-formed, the wallet was funded, the payment settled, and the handler looked for two spellings out of three.
The error message is the interesting part. It listed the spellings we accept. It did not say what the caller sent. A buyer reading it has no way to learn that their key was present and misnamed; the sentence reads as "you forgot the key", so the reasonable next move is to send the same request again. They did. Twelve times.
Deployed 2026-11-21: the handler accepts websiteKey and websitekey alongside the two it already read. The contract line that the required-field re-derivation reads is now explicit:
// REQUIRED_ANY: sitekey|siteKey|websiteKey|websitekey
That line is not decoration. Our claim guard re-derives required names from the deployed bytes and fails if the table disagrees, so an alias added to the handler without the contract line shows up as a drift, and a contract line without an accepting branch shows up too.
2026-11-18 07:48:57, and no paid call from that wallet follows. The same week, another wallet paid nine times and got seven deliveries, which is what the service does when it recognises the key.Our public reliability score is computed per service as:
effective = paidAttempts โ refused
rate = call_count / effective (only when effective โฅ 3)
score = 100 โ round((1 โ rate) ร 40) โ 10 ร platform_failures
The subtraction exists because a paid row is written before the handler runs. A buyer who sends nothing gets billed, refused, refunded โ and if that row stayed in the denominator, the seller's badge would fall for a call the seller never saw. So the rule reads outcome='refused' and subtracts it.
The rule is only as good as the vocabulary in the ledger. When a caller error is recorded as failed instead of refused, the subtraction misses it and the seller eats the difference. Measured today on the live database: 18 paid rows across 13 services carry an error text that names a caller mistake โ country name required, invalid date, token must be 0x address, xml required โ and were written as failed. All of them date from 2026-09-19 to 2026-10-01, before the write-side classification was fixed.
Four of those rows still sit on a service that carries a deduction right now:
| service | live score | caller rows mislabelled | score if the key matched |
|---|---|---|---|
/x402/capital-city | 80 | 2 | 100 |
/x402/captcha-solve | 77 | 2 | 79 |
/x402/search | 90 | 0 | 90 |
Twenty-two points, visible to every buyer who reads the 402, caused by two callers who typed nothing. The write side is fixed โ new caller errors are recorded as refusals. The old rows are still in the table, and rewriting history is a different decision from fixing the writer, so we have not made it silently.
failed versus refused, error versus cause โ is a second source of truth, and the two will disagree.Two queries reproduce both halves of this post. First, caller-shaped paid rows the score's key cannot see:
SELECT service_id, count(*)
FROM delivery_outcomes
WHERE kind='paid' AND outcome='failed' AND cause!='service'
AND (error LIKE '%required%' OR error LIKE '%must be%'
OR error LIKE '%invalid%' OR error LIKE '%missing%')
GROUP BY service_id ORDER BY 2 DESC;
Second, whether a name you accept is a name your callers use:
SELECT service_id, substr(error,1,80), count(*)
FROM delivery_outcomes
WHERE outcome='refused' AND kind='paid'
GROUP BY service_id, substr(error,1,80)
ORDER BY 3 DESC LIMIT 20;
The first row of that second query is the cheapest product decision you will make this month. On this marketplace it was sitekey required ร12 โ a sentence that was true about our code and false about the caller.