Grade a GitHub repo with the slop formulas, not a vibe score.

Method

Published formulas, stated approximations.

slopify scores a public snapshot. It does not ask a model how “sloppy” the code feels. The two structural metrics come from SlopCodeBench² and were restated by Earendil¹. The letter grade is slopify’s own scale on top of those two numbers.

1. Verbosity

   $$\mathrm{Verbosity}=\frac{|\text{flagged lines}\cup \text{clone lines}|}{\mathrm{LOC}}$$

   Share of source lines that are either a wasteful-code hit or part of a duplicated window. SlopCodeBench uses 137 ast-grep⁸ rules plus clone lines. We do not have that rule set. Flagged lines here are a smaller, listed set of patterns ($\mathtt{range}(\mathtt{len})$, comparisons to True/False, json.dumps/loads round-trips, swallowed except/catch, and similar). Clone lines are 6-line token-normalized windows⁶.

2. Mass and erosion

   $$\mathrm{mass}(f)=\mathrm{CC}(f)\sqrt{\mathrm{SLOC}(f)}$$

   $$\mathrm{Erosion}=\frac{{\sum }_{f:\mathrm{CC}(f)>10}\mathrm{mass}(f)}{{\sum }_f\mathrm{mass}(f)}$$

   How much of the codebase’s complexity mass sits in functions above McCabe’s limit of 10³⁴. SlopCodeBench chose that cutoff from Radon rank C⁵. Cyclomatic complexity on a connected control-flow graph is $v(G)=e-n+2p$; here $\mathrm{CC}$ is 1 plus decision keywords, boolean operators, and ternaries — a static count, not a graph walk.

3. ΔLOC

   $$\Delta \mathrm{LOC}=\text{additions}-\text{deletions}\text{(last 8 GitHub weeks)}$$

   Earendil’s simplest slop signal: change in line count¹. We read GitHub’s weekly code-frequency series⁷. Shown next to the grade, never mixed into it. Missing when GitHub returns 202 or an empty series.

4. Letter grade

   $$t(x)=\mathrm{clamp}\left(\frac{x-{\mu }_{\mathrm{h}}}{{\mu }_{\mathrm{a}}-{\mu }_{\mathrm{h}}},0,1.25\right)$$

   $$s=\frac{1}{2}t(\mathrm{Verbosity})+\frac{1}{2}t(\mathrm{Erosion})$$

   ${\mu }_{\mathrm{h}}$ and ${\mu }_{\mathrm{a}}$ are Earendil’s published human and agent means. Letters: $A$ if $s<0.2$, $B$ if $s<0.45$, $C$ if $s<0.75$, $D$ if $s<1$, else $F$. The letter is that average, so a quiet verbosity score can hide agent-range erosion (or the reverse). The report names which metric pulled the letter. This mapping is slopify’s, not a result from the paper.

Limits

• Public GitHub only. Default-branch zip, capped at 120 files and 22 MB.
• A profile skill scans the last four public, non-fork repos and templates a SKILL.md from recurring hits. It sees code output, not prompts or the agent harness.
• Comments and string literals are stripped before $\mathrm{SLOC}$, flags, clones, and $\mathrm{CC}$, so line numbers still match the file.
• Function extraction is regex-based. Nested bodies are punched out of the parent so mass is not double-counted. Unusual syntax is still missed.
• A low score is not a proof of taste. Earendil notes coupledness and churn are still unmeasured¹.