Adapt the task family
Benchmark prompts, proxy tasks, and official fixtures become a repo-local runner plan with harness versioning, model routes, budget limits, and restartable ledgers.
Managed proof runs for agentic apps
ProofLoop Live
Give us a live URL or a codebase. We adapt benchmark-style tasks to your product path, run the app under proof gates, and return receipts that separate product proof from official benchmark scores.
2 lanes
Capability checks plus live browser certification.
4 receipts
Trace, scorecard, cost ledger, verifier receipt.
0 fake done
Claims require deterministic proof or an explicit scorer contract.
How it runs
Separate capability checks from user-interface proof.
Run capability first
The agent harness checks the task path before every item is multiplied through a browser. Free or low-cost OpenRouter model sweeps belong in this lane.
Certify the live path
Browser runs prove the real user path: load the public URL, interact through stable selectors, capture traces, and verify visible outcomes on the deployed app.
Publish receipts
Final reports label each result as product-path proof, proxy benchmark proof, or official scorer output. No runner report can blur those categories.
Evidence
Evidence a reviewer can inspect.
- Live browser trace Viewport, interaction log, screenshots, and selector contract failures.
- Agent harness ledger Restartable task state, model policy, estimated spend, and failure history.
- Scorecard and cost chart Pass rate by task family and model, including cost per completed proof.
- Official scorer receipt Only when the upstream scorer or an explicitly recorded judge contract ran.
Lane
Verifier
Status
Product path
Playwright + selector contract
passed
Proxy benchmark
Harness scorer + fixture lock
tracked
Official score
Upstream scorer or judge contract
explicit
Start
Send a live URL or codebase target.
This form creates a scoped run request you can email. The hosted site does not collect secrets, tokens, or repository credentials in the browser.