Verification benchmark
Check finished work in 3.7 ms instead of doing it again.
When one agent hands work to another, the next one either trusts it blindly or pays to redo it. With Proofable, the work comes with a receipt the next agent checks in milliseconds, for free.
3.7 ms
To check a finished result. Redoing it: 2.79 s.
150/150
Forged records caught, offline, no account
3.3×
Cheaper per correct handoff than redoing the work
Every system tested.
- Check the receipt3.7 ms
- Redo the work2.79 s
| System | Result | Run by | Source |
|---|---|---|---|
| Check the receiptProofable, median of 10 | 3.7 ms | Proofable, median of 10 | Run it yourself |
| Redo the workSame job, same model | 2.79 s | Same job, same model | Method |
Against the alternatives.
Time to trust a finished result
Check the receipt3.7 msRedo the work2.79 sTrust the textInstantAI tokens used
Check the receipt0Redo the work2,426Trust the text0Cost per correct handoff
Check the receipt$0.024Redo the work$0.078Trust the text$0.009Tampered handoffs caught
Check the receipt3 of 3Redo the work–Trust the text0 of 3Forged records caught
Check the receipt150 of 150Redo the work–Trust the text0 of 150format check
Handoff dollars price the measured tokens at the Claude Sonnet 5.5 list rate. Trusting the text is cheapest until the text was changed in transit.
One identity check, reused by 100 apps.
Stripe Identity, $1.50 a check
Prove once with Proofable$1.75100 apps each re-running it$150.00Saving85× lessSumsub Basic, $1.35 a check
Prove once with Proofable$1.60100 apps each re-running it$135.00Saving84× lessVeriff Essential, $0.80 a check
Prove once with Proofable$1.05100 apps each re-running it$80.00Saving76× lessFootprint KYC, $0.50 a check
Prove once with Proofable$0.75100 apps each re-running it$50.00Saving67× less
Priced from each vendor’s published list price and Proofable’s $0.0025 per check. A pricing model, not a live run.
Check the results yourself.
Questions
What is a receipt?
A tamper-proof record of what an agent was asked, the rules it followed and what it produced. Anyone can check it on their own computer, without trusting Proofable. Change one character and the check fails.
Why not have another model grade the work?
A grading model costs tokens on every handoff and can be fooled by the same text that fooled the first agent. Checking a receipt takes 3.7 ms, costs nothing, and fails the moment anything was changed.
How were the forgeries made?
We took 15 real receipts and changed each one ten ways, from editing the content to swapping who owns it to flipping a single bit. All 150 fakes failed the check. All 15 originals passed.
Verify this run yourself
git clone --depth 1 https://github.com/proofable/docs cd docs/public/evidence/checks npm install node tamper-check.mjs node reuse-check.mjs

Put the same checks on your agents.
Connect an agent. Every action it takes is checked against your rules and leaves a receipt you can show anyone.