A research agent with a wallet.

The question you just watched, "Are Kestrel Semiconductor's advanced-packaging lead times in Malaysia getting shorter?", with a S$2 budget. This is the live rehearsal run of 9 Oct (DeepSeek writing, OpenAI Decisions judging, XRPL Testnet settling), replayed at 2.4× from its 38 recorded events. Blue bars read and decide, orange bars pay, red is the failed proof, green is the refund and the verified article.

Real numbers from the run: 16 hits from 8 writers in 0.75 s, hybrid mode. Hybrid means two searches mixed: matching words, and matching meaning. If the meaning search is down, the screen says "keyword only".

Real text from the run's answer v1. The citation chips point at passage IDs (p2, p17) in a source the agent has read; code checks each one before the answer is shown. The coverage check found 0 of 2 requested facts answered, so a gap stays open.

Blue columns are the decision model's output (OpenAI Decisions · gpt-6-luna, one batched call per round). Orange columns are plain code. All numbers are from the live run. Every new writer starts with a track record of 0.80.

SYNTHETIC benchmark: 160 generated scenarios, 64 held out, 3 repeats, under a decision rule written down before the run. Both models with our best wording and calibration. The whole study cost about US$0.96. A decision model returns probabilities instead of prose, so a threshold like 0.20 can mean the same thing on every question.

// trust can only lower value, never the budget or cap
if (status !== 'active') verdict = 'SKIP_LOW_TRUST'
else if (rewrite)        verdict = 'SKIP_REWRITE'
else if (price > cap)    verdict = 'SKIP_OVER_CAP'  // S$1
else if (value < 0.20)   verdict = 'SKIP_LOW_VALUE'
else if (price > left)   verdict = 'SKIP_OVER_BUDGET'
else                     verdict = 'BUY'

// best value per dollar wins
pick = rows.filter(BUY).sort(byValuePerDollar)[0]

The meter replays the run's money: S$0.30 reserved and paid to AlphaLeak, S$0.30 refunded, S$0.40 paid to The Fab Floor.

HTTP/1.1 402 Payment Required
PAYMENT-REQUIRED: base64({
  "x402Version": 2,
  "accepts": [{
    "scheme": "exact", "network": "xrpl:1",
    "asset": "XRP", "amount": "30000",
    "payTo": "rGeN8NcEVRVvWeFoyY9Df1TyiLq9AhoX5T",
    "maxTimeoutSeconds": 60,
    "extra": { "invoiceId": "{quote, article, version,
                 manifestRoot, amount, payTo}" }
  }]
})

HTTP/1.1 200 OK
PAYMENT-RESPONSE: base64({
  "success": true,
  "network": "xrpl:1",
  "transaction": "FCC78F2D571A…E1C28323"
})

// + the article, its passages and their salts
// ledger 21,384,313 · tesSUCCESS
// 30,000 drops = 0.03 XRP, shown as S$0.30

The real AlphaLeak purchase. The agent pays the writer's own wallet; we never hold the money. Each writer runs its own facilitator, the part that checks and submits the payment. 8.7 s from 402 to article, most of it waiting for the ledger to validate. Explorer: testnet.xrpl.org · FCC78F2D…

Real passages from the two articles bought in the run. AlphaLeak's fingerprints all matched (the text is what it said it was) but its promised dated figure isn't in the text. Fingerprint = sha256(salt ‖ index ‖ passageId ‖ text); the manifest is signed by the same XRPL key that receives the payment.

The ledger can't force a refund: it can check a hash and a deadline, not whether a sentence contains a date. So the penalty lives in the track record. A writer who refuses or ignores a challenge for 30 s loses twice as much and is delisted. Calibration also counts: AlphaLeak claimed relevance 0.96 and delivered 0, so its score is multiplied by 1 − 0.96² = 0.08 as well.

Real PR titles and line counts from GitHub (Collaboration95/theFastandtheFungible). Blue bars are mostly new code, grey bars mostly deletion; bar height is lines changed (log scale), one bar per PR. The 6 Oct pivot threw away the first corpus and rebuilt search, payments and the agent loop in two days.