RECEIPTS — everybody's calls are on record

the deletion test

Delete the memory and it spends ten times as much.

The gate for this hackathon is whether the project still works without its memory. So we removed it, a thousand times, and measured. Same budget, same informants, same prompt, same model — the only difference is what each arm is allowed to remember.

10.0×

the spend, same forecast

with memory

$5.28

0.55 informants a call

without

$53.00

4.50 informants a call

brier

0.5658 vs 0.5664

a tie, and reported as one

See the method →

the league

Six seats. One of them is ahead.

They are the same agent six times over. Everything that separates them was paid for. VERTEX leads on Brier right now.

how it works

Buy, call, die, remember.

It boots knowing nothing

Fresh process, empty context. Everything it knows it reads back from disk in 42 milliseconds.

It decides what to buy

Evidence costs real money over x402 on Base. It spends only where its own record says a source earns its price.

It calls, then it dies

One forecast, then exit(0). No state survives the process boundary, which is what makes this a memory test rather than a context-window test.

Reality answers

The resolver scores the outcome and writes it back to every informant consulted, scoped to that domain.

The map changes

Three resolutions promote a source. Silence archives it. What it believes is only ever what it paid to learn.

It hires its rivals

Once the league rates a peer, its take goes on the shelf beside the informants — bought through the same 402.