← ClaudeAtlas

severitylisted

This skill should be used when rating findings in reviews, audits, and assessments. Preloaded on finding-producing agents.
lklimek/claudius · ★ 1 · AI & Automation · score 71
Install: claude install-skill lklimek/claudius
# Severity Classification Two independent axes per finding: - **severity** — how bad is the shipped defect? Derived from the `likelihood`/`impact` floats. - **merge_class** — does this stop THIS PR? Decided by the **blocker gates** below. Never encode one in the other. A LOW can block (it trips a gate); a CRITICAL can be `out_of_scope_follow_up` (pre-existing, untouched, no gate reachable through this PR). ## 1. Backstop zone (judge this first) Before rating anything, answer: **what stands between this defect and irreversible harm?** The answer sets the scope of G-INTENT and the ceiling on `impact`. | Zone | What backstops the code | Worst realistic outcome | `impact` ceiling | |---|---|---|---| | **Backstopped** | A server, consensus, or independent validator rejects bad output before it counts | Scary error, confusion, stuck or wedged client, wasted user time | ~0.7 | | **Sovereign** | Nothing. This code is the last line — key material, signing, entropy, local persistence, offline/standalone tools | Irreversible loss of funds, keys, or data | 1.0 | | **Boundary** | This code decides *what crosses into* sovereign territory — assembling the tx to be signed, choosing the destination, composing what the user is asked to approve | Same as Sovereign — the backstop validates the *signature*, not the *intent* | 1.0 | For **server-side code**, the code owns the data — read it as Sovereign over its own persistence, with the operator's monitoring as the only backstop against si