Run the app-builder benchmark

The benchmark measures cost (exact wire token counts × pinned list prices), quality (60% fixed Playwright customer-journey suites + 25% security probes + 15% single LLM judge, per checkpoint, averaged over three milestones) and duration, for building three apps (Relay CRM, Deskhero, Portalis) with Dyad's local-agent mode. Everything is n=1 per cell, so a single bad line in one milestone moves an app column by 30 points; read every number with that in mind and say so when reporting.

Where things are

BENCH = the benchmarks/app-builder directory. It is not on every checkout: look in the current repo first, then the sibling checkout the user runs benchmarks from (as of 2026-09 that is /Volumes/essd/dyad-2, whose benchmarks/app-builder carries uncommitted hardening). Export BENCH for the bundled scripts if it is not under this repo. Key pieces: