21% of the code we scored in karpathy/llm.c — weighted by bytes, covering 100% of its code — sits in files no human has recently written, reviewed, or explained (comprehension score below 0.30 out of 1.00). Higher is worse. Non-code files — images, lockfiles, minified bundles, prose docs, generated files — are excluded throughout.
Every score decomposes — hover any cell.Methodology →github.com/karpathy/llm.c ↗
Byte-weighted average across 96 scored files (0 unscored). Below 0.30, the file has no human on its recent record.
96 scored · 0 unscored · 1,201,966 scored bytes · 0 unscored bytes
Comprehension score = 0.40 human review depth + 0.25 author recency + 0.25 bus factor + 0.10 question answerability. Bus factor caps at 3 distinct humans; recency decays over 180 days. A file counts as debt below 0.30.
Question answerability is 0 for every file Fathohm scores, here and in every installed repo. It was answered by the same people it graded, so it is collected and shown back to the team but never scored — the weight is held open for a check a teammate can verify. Every score therefore tops out at 0.90, not 1.00.
100% of scored code, by bytes, gets zero from this factor.
Each percentage is that directory's own debt — its indebted bytes over its scored bytes — so these do not add up to the repo number.
dev/cuda/33% debt(root)/5% debtllmc/9% debtdev/eval/100% debtdev/data/26% debtAuthorship is not the story. Across the gallery, the most indebted repo is almost entirely human-written and the healthiest is AI-era. Debt tracks how deeply humans review — not who, or what, wrote the code.
Share of scored bytes — the same denominator as the debt percent. Detection method is public; labels are correctable.