Proof Receipt · real jobFAIL
- Task
- Fix bucket key lookup
- Repository
- more-itertools
- Hidden tests
- passed
- Library's own tests
- broken by the fix
- Replays
- 3 of 3 agree
- Refund to buyer
- 240.00 USDC
- Stipend to buyer
- 24.00 USDC
Acceptance testing for AI coding agents
Pay AI agents only for code that works.
Your AI agent says the bug is fixed. ResultBond checks the fix with tests the agent never saw — three runs in a sealed sandbox — and signs a verdict anyone can verify. You pay for the fixes that pass.
Not another code reviewer: a reviewer gives an opinion, ResultBond gives a verdict you can check. When the machine can't tell, a person decides.
Every miss published10 wrong verdicts, all explained
0 good fixes blamedin 570 benchmark runs
36 receiptson 12 real open-source bugs
Anyone can checkverify a receipt in your browser
The problem
AI agents write code faster than anyone can check it.
Companies now hand real bugs to coding agents. But the way they pay for that work hasn't caught up.
01
You pay for attempts, not results
Agents are billed by seat, credit or hour. A fix that breaks something else costs the same as one that works.
02
The agent grades its own work
“All tests pass” usually means the tests the agent could see — and sometimes wrote. Nobody independent signs off.
03
Checking costs more than the fix
A senior engineer reviewing every AI patch eats the savings. Skipping review lets bad code ship.
An example, step by step
One bug, one agent, one honest answer.
An online shop has a bug: the cart total rounds wrong. They want it fixed for $240.
- 1
The job is agreed up front
The shop, the agent and ResultBond sign one job sheet: the repository, what “fixed” means, and the $240. Nobody can change the terms afterwards.
- 2
Hidden tests are prepared
Before any work starts, the shop's engineer or ResultBond writes tests that prove the bug. We check they fail on today's code and pass on a known-good fix. The agent never sees them.
- 3
The agent submits a fix
One submission. We fingerprint exactly that code, so nobody can swap it later.
- 4
We run it three times in a sealed sandbox
No internet, strict limits, the old tests and the hidden ones. Three runs, so a flaky test can't blame good code.
PASS
All hidden tests pass
The $240 goes to the agent. Both sides get the same signed receipt.
FAIL
The fix doesn't hold up
The shop doesn't pay the $240. In a bonded job it also gets a $24 stipend from the agent's bond.
HOLD
The evidence isn't clear
The receipt says INCONCLUSIVE. Nothing moves automatically: a person looks at the evidence and decides.
The hidden tests
Who writes the tests the agent never sees.
The verdict is only as good as the tests. Three ways to get them — and one gate every test goes through before an agent is judged.
Your engineer
You know what “fixed” means for your code. Your team writes the hidden tests; we keep them secret from the agent and its platform.
Best for: teams with clear acceptance criteria.
Our team
We write the hidden tests from your bug report and your engineer signs off before any agent runs. Offered as a service for pilots.
Best for: a backlog without tests yet.
From real fixes
For open-source work and public comparisons, the tests come from the fix the maintainers actually shipped.
Best for: agent vendors and benchmarks.
The gate: every hidden test must fail on the current code, pass on a known-good fix, give the same answer run after run, and not copy a public test. A test that would reject a correct fix is caught here, not on the agent. How hidden tests stay hidden →
Optional: bonded jobs
Nobody presses “pay”. The receipt does.
Most teams start with invoices: pay for the fixes that pass. Marketplaces that want more can have both sides lock money in an escrow contract before work starts; it pays out only on a signed receipt for exactly that job.
Deadlines protect both sides. No bond in time: the client takes its money back. No passing receipt in time: the client gets its money plus the stipend.
Each side withdraws its own money. A blocked wallet on one side can't freeze the other.
Running on a testnet. Two real jobs settled this way on the Monad testnet → Real money only after an external audit of the contract.
Who it's for
For everyone who pays for, sells or hosts AI coding work.
Companies building coding agents
Public benchmarks are losing credibility. Show buyers how your agent does on tests it never saw, judged by someone you don't control — and sell on results instead of seats.
You get: independent proof customers believe.
Teams buying AI fixes
Hand your bug backlog to agents and pay only for fixes that pass. No more reviewing every patch by hand.
You get: a verdict and a receipt for every fix.
Agent marketplaces
When a job is paid from escrow, someone has to say “done”. For code, that can be us — on a standard like ERC-8183.
You get: a neutral judge for coding jobs.
Does it actually work?
We test the tester — and publish the misses.
570 end-to-end runs on good fixes, bad fixes, cheating attempts, flaky tests and 12 real bug fixes from open-source libraries.
98.2%right answer560 of 570 runs
0good fixes blamedcorrect code never marked FAIL
10misses, all disclosedone known attack, closed in the stricter test mode
2 smedian checkper run, in the sandbox
Every wrong verdict, explained → Hidden Test Index → Read the case study: 12 real bugs, 3 fixers → See the benchmark →
Why you can trust a receipt
Built to be checked, not trusted.
Signed
Every receipt is cryptographically signed. Change one character and the signature breaks.
Specific
It names the exact code, the tests and the rules that were used — nothing is “trust us”.
Checkable by anyone
Paste a receipt into resultbond.com/verify. The check runs in your browser; nothing is uploaded.
Hard to cheat
Editing the tests, crashing or hanging is caught and counts as a FAIL. Forging the test report is caught in black-box mode, the mode we use for paid jobs.
Two ways to start
Start in shadow. Put money on it when the verdicts agree.
Shadow
Free pilot- A signed receipt for every job, no money moves
- Compare our verdicts with your own reviewers
- 4 weeks, 30–50 of your real bugs
Paid
Per verdict- Invoiced per verdict; fixes that fail cost you nothing
- Optional: payment and bond in a USDC escrow, FAIL refunds the buyer plus a stipend
- Turned on once shadow verdicts match your reviewers
Questions
Who writes the hidden tests?
Your engineer, our team, or — for open-source work — tests from the real upstream fix (see above). Whoever writes them, each test must fail on the current code, pass on a known-good fix and not copy a public test. If it doesn't, the job isn't judged automatically.
What if the tests themselves are wrong?
Then the result is HOLD (INCONCLUSIVE on the receipt), not FAIL. A good fix is never blamed by a test we couldn't validate — in our benchmark, correct code was marked FAIL zero times.
Can the agent game the check?
It never sees the hidden tests, runs in a sealed sandbox, and in the stricter black-box mode the code can't touch the test results or read the hidden tests at all. In that mode every known trick we tried ended in FAIL; the one trick that still works in the ordinary mode is published in the benchmark.
Who pays the stipend?
In a bonded job, the agent, from the bond it puts down when it takes the job. That's what makes the promise real. Invoiced work has no bond: you simply don't pay for fixes that fail.
Is real money moving yet?
Not yet. Bonded jobs run end to end on a public testnet with test USDC — every transaction is linked. The escrow contract gets an external audit before it holds real money. Pilots start in shadow mode, where no money moves at all.
Do I need crypto?
No. Pilots are free and move no money: we run in “shadow mode” and you get the receipts. Paid work is invoiced. A USDC escrow is an option for marketplaces that want payment held until the verdict.
Which languages work?
Python today. Pilots decide what comes next.
Try it on your own bugs.
A free four-week pilot: 30–50 real bugs, your agent of choice, a receipt for every job and a report at the end.