B
REQUESTER
Buyer Agent
task + wallet intent
FROZEN-REPLAY EXPERIMENT · REAL RUN DATA
An Agent learns
who is worth paying.
Watch one autonomous buyer make 10,000 paid decisions. Every line is a real purchase; every market shock changes the same Bandit beliefs, wallet pressure, and provider allocation shown below.
Starts slowly · accelerates · pauses around shocks
HOW 402PILOT SITS IN THE LOOP
one paid call · one receipt · one belief update
BUYER-SIDE DECISION LAYER402Pilot
01Context
02Wallet mask
03PA-DCTutility × cost
04Chosen-only update
HTTP402
PAYMENT RAIL
x402
pay + receipt
PAYABLE MARKET
Provider Market
five providers · one selected
response + receiptchosen-only learning
823effective tasks
20,575scored responses
5 × 4providers × task buckets
30paired seeds / scenario
$50finite wallet
LIVE EXPERIMENT REPLAY
See the decision. Then see the learning.
Seed 00 drives the individual transactions. Allocation shares and endpoint results summarize all 30 paired seeds.
MARKET REPLAY
No announced change. PA-DCT learns a stable market.
TRACE · SEED 00
ProbeCommitDetectRelearnSettle
BUYER AGENT
402
Probing
“What is each provider really worth?”
Wallet remaining$50.00
0.0% usedlinear plan
01 · PROBE
T1Coding
PA-DCT SELECTSWaiting…
READY
quality—
charged—
Only the purchased arm will update
C
P-cheapqwen3.5-flashweakest cheap tier
$0.000520% recent
0 buys
M
P-midGPT-5.4-minireliable mid tier
$0.00220% recent
0 buys
P
P-premiumGPT-5.4strongest premium tier
$0.01020% recent
0 buys
PRICE DROP
A
P-advGPT-5.4-mini60% fluent wrong answers
$0.00220% recent
0 buys
F
P-flakyGPT-5.4-mini40% billed timeouts
$0.00220% recent
0 buys
MARKET · S1
Stationary market
No shock is announced or injected.
ROUND0/ 10,000
1 round / beatWatching every purchase
INSIDE PA-DCT · CURRENT TASK BUCKET
Discounted Q/C posteriors
REBUILT FROM TRACE
score =(1 − λ̃) · utility sample−λ̃ · normalized cost sample
provider / Q beliefrealized-cost beliefmean score
PROVIDER ALLOCATION · TRAILING 200 ROUNDS
30-seed mean allocation
30 SEEDS
round 0
The bright segment is what the buyer has learned so far; the faint path shows where this real experiment goes next.
Running quality—seed 00 · q̄
Failures observed0paid non-delivery
Quality / dollar—seed 00 · ROI
Cumulative PA reward—seed 00
Learning stateUncertain priorsAll five arms begin plausible.
THE FULL 30-SEED STUDY
One policy. Three different lessons.
The replay above shows the journey; these are the benchmark's aggregate endpoints. Lower PA-gap/T is better.
ALL EVALUATED POLICIES
S1 · Stationary endpoint
Mean over 30 paired seeds · Oracle is a replay upper bound, not an admissible buyer policy.
| Policy | Quality ↑ | Budget used | ROI ↑ | PA-gap/T ↓ | Readout |
|---|
WHY THE LINES MOVE
Payment is the learning interface.
402Pilot sits above the payment rail. It selects one affordable quote, pays once, observes only that result and receipt, then changes one provider–task belief.
The $50 wallet removes providers that cannot be paid at the current round.
PA-DCT draws utility and realized-cost beliefs for five arms in the active task bucket.
The best wallet-aware sampled score wins. Only this provider receives the request.
Quality, failure, and charged cost update the chosen cell; all old evidence is discounted.
LOCKED ACROSS ALL SCENARIOS
γ = 0.999discount
ν = 0.5failure penalty
α = 2.0wallet pressure
cmax = $0.01cost normalizer
Kb = 4task buckets