severity

# Severity Assessment Rate how severe a finding or bug is **if it is real**, and emit one tier the caller acts on. > **Severity is not confidence.** `confidence` answers *is this real / correct*. > `severity` answers *how bad is it if it is*. The two are orthogonal — a caller > composes them (surface when severity clears the caller's bar for its tier AND > confidence clears that bar). Never fold them into one number. > **What is actually deterministic here.** One thing: the **path floor** in Step 2 > is an executable glob against the file path (a real `case` match), so it never > depends on model judgment. The **escalators** (data-loss / security / concurrency > shapes) are heuristic — an LLM reads the code shape. This is the honest split; > do not call the escalators deterministic. It mirrors `confidence`'s executable > rule checks only for the path floor. ## Contents - [Mode Detection](#mode-detection) - [Severity rubric](#severity-rubric) - [Step 1 — Base tier from axes](#step-1--base-tier-from-axes) - [Step 2 — Path floor and escalators](#step-2--path-floor-and-escalators) - [Step 3 — Combined tier](#step-3--combined-tier) - [Output Format](#output-format) - [Mapping to a reviewer's blocking flag](#mapping-to-a-reviewers-blocking-flag) - [How Callers Consume the Tier](#how-callers-consume-the-tier) - [Evals](#evals) --- ## Mode Detection Check the arguments: `$ARGUMENTS` | Argument | Default | Rates the severity of | Typical caller ...