Repair broken browser tests without hiding real application bugs
Real results from python demo.py and the 50-scenario benchmark. Nothing on this page is simulated.
Correct repair rate
89.5%
34/38 refactors · Playwright alone 0/38
False-pass rate
0.0%
0/12 real bugs reported as passing
Fixed at $0.00
61
validated heals, offline, no API key
Ladder latency
921 ms
median per heal
One test, four states of the same checkout
Original
The checkout the test was written against
PASS
0 heals needed
Refactored UI
Every id, class and test id renamed; labels reworded
PASS (valid heal)
4 selector(s) healed, all assertions passed
Broken app
Refactor + a real bug: orders are silently not created
FAIL (heal rejected)
4 repair(s) applied but the assertions failed -> not counted
Broken: button removed
Refactor + the Place-order button was deleted
FAIL (refused to heal)
#place-order blocked: ambiguous
Inspect the repair evidence
Toggle between the original, the healed refactor and the blocked broken app: screenshots, markup diff, ladder trace, assertions.
Copied from `python demo.py` (artifacts/demo) and `python scripts/run_benchmark.py` (BENCHMARK_RESULTS.json). Not hand-written. Generated 10/16/2026, 12:31:55 PM.