Audit Implementation

Verify a feature actually does what it claims, fits the stack idiomatically, and earns its complexity. Default is the simplest stack-native path; anything else must justify itself with a real, current reason.

Examples lean on a web stack (React + a data-fetching library + an ORM) because that is where these smells cluster, but the four axes apply to any stack — swap the tool names for whatever yours provides.

Glossary

• Label-truth — the visible action navigates or mutates exactly what the user reads. "Pending Orders" → list filtered to pending, not the unfiltered list.
• Justified query — a fetch whose shape, freshness, and key are explained by what the screen renders today; not "in case", not "for symmetry".
• Stack-native — uses what the chosen tool already provides (CSS variables, data-* variants, your data-fetching library's options, your ORM's relation includes, your API contract's input validation, component-library variants) instead of a hand-rolled equivalent.
• Misused tool — a tool used for the wrong job. Looks load-bearing, actually adds friction without paying back.
• Earns its keep — has a real current caller, a measured win, or a documented invariant. Hypotheticals do not count.

Axes

Each axis has a detail file with heuristics and do/don't examples. Read the relevant file before classifying a finding — the heuristics are the citation, not vibes.