Agent Andy

Agent Andy is an agent that checks a local rule before you act, and proves its answer. The problem: agents act first and explain later. For a question like "Can I detect at Myrtle Beach?", a confident wrong answer is not just unhelpful, it can lead to a fine. I run wheretodetect.org, a metal-detecting spot-rules site, so this is a real failure mode I care about. Andy answers only with a cited verdict, or it stages instead of answering. It checks the rule first, then shows its work: every action leaves a hash-chained, signed receipt, and verification names the exact receipt if one byte is changed. A staged answer is not a refusal, it is the honest state when the evidence is not there. Verified results, all built Oct 26 to Nov 4, 2026: 16/16 hostile inputs blocked or staged, 12/12 attacks caught, 8/8 legitimate requests allowed, and a 33-receipt recorded live chain that anyone can verify without installing anything at https://drvelvetfog.github.io/agent-andy/ The demo shows Myrtle Beach answered ALLOWED with a cited source, Tybee Island staged as UNVERIFIED when no primary rule was found, and a Witness packing check (book, cable, charger) that ships only if every item is seen in a photo. Plummy and aee-e2 inspired Andy and are not part of this submission. Andy is a solo entry: Tony Jagodka, Director, with Cosmo (Muse) as Build Manager and Claude (Claude Code) as Builder.