An illustrative A/B · SIMULATED

Does the sandbox actually pay for itself?

A synthetic 60-item corpus (30 real, 30 slop/duplicate) whose outcomes are set in code: which reports reproduce, which read as convincing, what each costs. It illustrates the policy (pay only after a sandbox reproduction), not a measurement: the gap between the two arms follows from how the corpus is built. A measured comparison needs real reports from design partners.

Arm A · text-only triage

believes the prose

Would pay50 reports

False payouts20

Dollars exposed$5,000

Cost / 60 reports$0.60

Arm B · sandbox required

pays only on proof

Would pay26 reports

False payouts0

Dollars exposed$0

Cost / 60 reports$1.80

Honest misses (Arm B) — real bugs that didn't reproduce

These 4 are state-dependent flaws the sandbox couldn't reproduce from the recorded steps alone. They're shown, not hidden — the gap the roadmap closes with stateful replay.

C-real-1C-real-2C-real-3C-real-4

Verdict

Text-only triage would have paid $5,000 to 20 fake/duplicate reports. Requiring proof would have paid $0 to fakes and made 4 honest misses (shown). In this synthetic corpus, requiring proof removes every false payout; real reports from design partners will measure it.