Scoring¶

VibeCode QA reports one number: a composite score from 0 to 100, mapped to a letter grade. It's designed to be stable across codebase sizes — no absolute-count cliffs — so a 200-file app and a 5-file library are graded on the same curve.

The formula¶

score = Σ(checkᵢ × weightᵢ) / Σ(weightᵢ)

Each check produces a 0–100 sub-score, multiplied by its weight; the weighted sum is divided by the total weight of the checks that ran. Checks that did not contribute are excluded from both sums, so they never penalize you: checks skipped as not applicable (e.g. the React check on a non-React project), Pro checks that are unavailable without a VCQA_PRO_KEY, synthetic checks derived from another check's data, and checks a project disabled in .vcqa.json.

Category weights¶

The seven AI Analysis checks carry weight 0 — they surface deeper findings (stale docs, contradictory patterns, fake tests) without moving the score. Several platform-specific checks (flutter, html-quality, container-health, cloudflare-workers, sqlite-d1) also carry weight 0 today: they report findings without moving the score.

Grades¶

The grade is presentation only. CI gating uses a score threshold, not a grade: --ci fails under 60 unless --fail-under N sets your own number.

Design principles¶

• Proportional, not absolute. Sub-scores scale to codebase size — duplication is a percentage, complexity is per-function, and so on. Adding code never tanks your score just for being bigger.
• Severity-weighted penalties. Where a check aggregates many findings (e.g. Best Practices), errors cost more than warnings, which cost more than infos — so missing a nice-to-have doesn't read like a critical bug.
• Trends over time. Every scan is saved to .vibe-check/history/ (last 30). The report shows your score delta, plus new and fixed issues versus the previous run.

See the 38 checks for each check's weight and what drives its sub-score.