ai-fluency-discernmentlisted

# Discernment — evaluating output, process, and behavior Discernment is the flip side of Description: judging what came back, in three dimensions. It gets *harder* as agents get better, because polished output removes the obvious error signals — "it looks right" is never sufficient. ## The three dimensions - **Product** — accuracy (verifiable?), appropriateness (fits audience/purpose?), coherence (internally consistent?), relevance (addresses what was asked, or drifted?), completeness. - **Process** — did it follow the steps, or improvise? logical gaps? unauthorized assumptions? did it stay in scope? In an agentic run: what did it actually do, in what order? - **Performance** — did it flag uncertainty or present guesses as fact? ask good questions or barrel ahead? behave consistently with what was asked? **Coherence and accuracy are independent.** Something can be well-written, internally consistent, and still factually wrong. The most common failure is accepting a fluent output that doesn't reflect reality — only domain knowledge catches it, so the human is the essential check; the agent cannot fully self-evaluate. ## Output review (run before using a significant output) - Product: factual claims verifiable or flagged? matches the requested format/audience? no internal contradiction? answers the actual question? - Process: followed the steps in order? no silent scope expansion? assumptions stated, not hidden? - Performance: uncertainty surfaced? checked in at t