animaresearch/verdict-sweep

verdict-sweep is a small protocol that reshapes how AI research agents answer comparison questions. Instead of returning a long essay, it guides the agent to produce a decision matrix: a table where each row is a candidate and each column is a criterion the user cares about. The repo packages this as a "skill" file, called SKILL.md, that you copy into the skill directory of a compatible agent such as Claude or Codex. Once installed, you ask a question like "which local LLM runtime should I use?" and the agent follows the protocol to build a structured, evidence-backed table instead of a narrative summary. The core idea is to set the criteria before searching. The agent turns your stated concerns into columns, then searches for candidates and fills each cell with a value like yes, no, partial, or unclear. It checks repositories, activity levels, licenses, and implementation details when available. A running implementation outranks a paper-only claim. The protocol also separates the collection step from the final judgment step, so the agent that gathers a candidate does not also get to overstate its case. Blind spots, things the sweep did not cover, are written into the output as their own section. Installation is manual. You clone the repo and copy the skill folder into the appropriate directory for your agent. The README provides commands for Claude, Codex, and Windows PowerShell. After that, you interact with the agent in plain language and it follows the protocol automatically. The repo ships with two worked examples: one scoring 13 local LLM runtimes, and another surveying 22 systems related to expert offloading in mixture-of-experts models. Use cases include library selection, prior-art surveys, competitor mapping, technical due diligence, and vendor selection. The protocol is not limited to software topics, though the GitHub-based evidence columns matter less for non-software comparisons. The README is clear about what the tool does and does not do. It is not a replacement for deep research. It is a stricter output format for situations where you want a table you can argue with.