# Reference-guided software engineering

This work introduces the ProgramDistill benchmark and a fully automated mine–craft–patch pipeline that factorizes 26 interactive web applications into 1,975 replay-verified behaviors and constructs 4,063 tasks, letting coding agents infer and restore missing functionality by interacting with a reference application whose source is hidden; across nine frontier agents, the best full-application reconstruction success is 49.2%, and partial-application reconstruction success falls from 100% to 64.0% and from 96% to 32% as restoration depth increases from 1 to 8.

This work introduces the ProgramDistill benchmark and a fully automated mine–craft–patch pipeline that factorizes 26 interactive web applications into 1,975 replay-verified behaviors and constructs 4,063 tasks, letting coding agents infer and restore missing functionality by interacting with a reference application whose source is hidden; across nine frontier agents, the best full-application reconstruction success is 49.2%, and partial-application reconstruction success falls from 100% to 64.0% and from 96% to 32% as restoration depth increases from 1 to 8.