prototypelisted

# prototype: spike to learn, then throw it away Some uncertainty is cheaper to build through than to talk through. A prototype exists to answer ONE named question; the deliverable is the answer, never the code. ## Contract (agree before writing code) 1. **The question**: one sentence, falsifiable. "Can X stream 10k rows without buffering?" not "explore X". 2. **Timebox**: default 30 to 60 min of agent work. Hit the box with no answer → that IS a finding (question too big; split it or grill further). 3. **Kill criteria**: what result answers the question either way. 4. **Where it lives**: an authorised, manifest-owned scratch dir outside production source. Never import it from real code or touch external systems, credentials or user data without separate authority. ## Rules - Build only the checks needed to answer the question. Production quality gates do not apply to disposable scratch code; evidence, authority and containment do. - **Cheat aggressively**: hardcode data, stub auth, fake the network. Only the question's variable needs to be real. - **One question per spike.** A second question mid-spike → note it, finish, spike it separately if it still matters. - **Harvest, then retire.** Put the finding and evidence in the owning spec, decision record or scoping thread when artifact authority exists. Remove only paths proven run-owned and explicitly authorised; otherwise quarantine or hand them back with a named cleanup action. - **Prototype code