What problem does it solve? Reading a paper to make a decision (adopt a method, cite a result, revise a hypothesis) often produces a false sense of understanding from AI-generated summaries. This Skill forces retrieval-based reconstruction before any generated summary is shown, and mechanically checks the reconstructed argument for gaps and overclaims. ## Core Features & Use Cases - Decision-first framing: Fixes one concrete decision (adopt, cite, revise hypothesis, replicate, compare) and judges every later question against it. - Source freezing: Records title, authors, version, DOI/URL, and file SHA-256 in a source manifest so preprint and published versions are never blurred. - Closed-book reconstruction with checker feedback: The reader builds user-map and evidence-map JSON files from memory, then a deterministic checker (codes like S401, E303, X905) picks the next probing question. - Use Case: A researcher deciding whether to cite a result reconstructs the paper's claim-evidence-warrant chain, the checker flags an overclaim in the abstract (S401), and the citation decision is recorded with the evidence behind it. ## Quick Start Ask the agent to read a specific paper for a stated decision, for example: read paper.pdf to decide whether I can cite its main result, and walk me through reconstructing its argument.
