Large language models with web-search capability share a common failure mode: they stop at the snippet layer, fabricate missing specifics with plausible labels, and present secondhand summaries as primary evidence. This document — sammaₑsanā (Pali: "right searching") — specifies a runtime discipline that externalizes what RLHF training cannot internalize. It defines a four-tier evidence hierarchy, problem-type-keyed stopping conditions, a degradation grammar for unverified claims, a chunking protocol for long documents, explicit UNKNOWN rules, and a four-model division of labor for search tasks. The file is loaded into a model's persistent context (Project knowledge, system prompt, or custom instructions) and activated by a corresponding reference line in the model's persistent memory. Both are required; neither alone activates the discipline. Designed across four models in a single loop: GPT (skeleton, 19-section evidence hierarchy and degradation grammar), Gemini (red team: chunking protocol, scraping-failure flagging, per-turn self-declaration), Grok (reconnaissance design: UNKNOWN Rules 0–4, 6-angle 4-stage reconnaissance template), Claude (integration). Final judgment: the human author. Released under the MIT License. Fork, adapt, improve.
Akimitsu Takeuchi (Fri,) studied this question.