verify-badgelisted

## Measured evidence (already computed — do not re-derive) ```! python3 "${CLAUDE_SKILL_DIR}/scripts/measure.py" 2>&1 || echo "미측정 — 트랜스크립트 접근 불가" ``` # Verify Badge Attach **named accountability + verification history + a cold limits diagnosis** to a report. Skill instructions are English; **all user-facing output stays Korean** unless the user asks otherwise. ## Core principles — this is an audit, not a stamp 1. **Never invent verification.** If it wasn't done, write "미검증". If no errors were caught, write "발견 0건". 2. **The AI is the grader.** Do not inflate grades to please the user. Weak evidence gets a D. 3. **The limits section is mandatory.** Naming what can't be trusted is what makes the rest trustworthy. 4. **Real author name required.** Anonymous = the badge means nothing. --- ## Procedure ### Step 1 — Harvest evidence from the session (before asking anything) Pull **only what actually happened** from conversation history: | Item | Source | |---|---| | Search / document lookups | WebSearch·WebFetch call count | | Primary-source access | Did you open official docs, config.json, APIs, price pages directly? | | Cross-check rounds | Times the user said "다시 확인", "재검증", or you revisited values | | **Errors actually caught** | Value corrections in history (most important signal) | | Unresolved assumptions | Constants picked without grounding | | Prompt count · work period | Transcript — **count, don't estimate** (see below) | **If nothing was caught, say "발견 0건.