slop-eval-cli 0.1.0

Scores AI-generated UI for genericness ("slop") using an Anthropic LLM-judge rubric, with a --json mode for CI and agent quality gates.

slop-eval-cli (Python)

Score AI-generated UI for genericness with an Anthropic LLM judge, so a CI check catches the same "this looks like every other AI-built app" problem a human reviewer would flag on sight.

Why this exists

slop-eval scores a screenshot (or, as a fallback, a URL's raw HTML/text) against a versioned, public rubric using the Anthropic API, called with a forced tool call so the response comes back as structured JSON instead of a chat reply that has to be parsed apart. It complements deterministic AI-UI-tell catalogs (Impeccable's Slop, aislop) rather than replacing them: neither of those does holistic, judgment-based scoring ("does this layout feel novel," "does this component choice feel considered") the way an LLM judge can. Full background and the honest comparison table live in the project README.

This package is the Python distribution -- a genuine, independent port of the npm package's scoring logic, not a wrapper around the Node binary. It calls the Anthropic API directly via the official anthropic Python SDK.

Install

pip install slop-eval-cli

or with uv:

uv add slop-eval-cli

The complementary JS/TS distribution installs the same way on the npm side: npm install -g slop-eval-cli (or npx slop-eval-cli score ... to run it once without installing) -- see the project README for that package. Both are first-class, maintained together; neither is deprecated in favor of the other.

Quickstart

Requires Python 3.9+ and an ANTHROPIC_API_KEY (BYO key; get one at console.anthropic.com).

export ANTHROPIC_API_KEY="sk-ant-..."
slop-eval score --screenshot ./preview.png

For CI or agent consumption, add --json:

slop-eval score --screenshot ./preview.png --json

Illustrative shape of the output (not a captured live run -- no Anthropic API key was available in the environment this package was built in; the real score and evidence text come from whatever your screenshot and the live judge actually produce):

{
  "target": "./preview.png",
  "rubric": "v1",
  "compositeScore": 62,
  "findings": [
    {
      "ruleId": "llm-judge.layout-novelty",
      "category": "Layout novelty",
      "score": 4,
      "evidence": "Matches a common hero + 3-card grid + footer CTA pattern.",
      "status": "flag"
    }
  ],
  "summary": { "pass": 1, "flagged": 1, "notScored": 1 },
  "disclaimer": "This score is a heuristic quality signal from an LLM judge, not a certification..."
}

Without a key set, every run (human or --json) fails fast with exit code 2 and a clear "set ANTHROPIC_API_KEY" message:

slop-eval score --screenshot ./preview.png --json
# {"error": "ANTHROPIC_API_KEY environment variable is not set.\n..."}
# exit code 2

Or call the library directly (the agent-native path):

from slop_eval import score_composite, ScoreInput, LLMJudgeSource, ScreenshotDiffSource

sources = [LLMJudgeSource("v1"), ScreenshotDiffSource()]
result = score_composite(sources, ScoreInput(screenshot_path="./preview.png"))

print(f"Composite score: {round(result.composite_score)}/100")
for finding in result.findings:
    print(f"[{finding.status}] {finding.category}: {finding.score}/10 -- {finding.evidence}")

CLI reference

slop-eval score [options]

Options:
  --url <url>          URL to score (fetched as raw HTML/text -- v0.1 does not
                        render pages; see limitation below)
  --screenshot <path>  path to a screenshot image to score (preferred over --url)
  --rubric <name>       rubric version to use (default: "v1")
  --json                output structured JSON instead of a human-readable report
  --fail-below <n>      exit code 1 if the composite score is below this
                        threshold (0-100); no threshold by default

Exit codes: 0 success (no threshold, or score at/above --fail-below), 1 success but below threshold, 2 usage error or unrecoverable failure (missing API key, unreadable file, malformed rubric, mutually exclusive --url/--screenshot). Identical contract to the npm CLI.

--url and --screenshot are mutually exclusive; passing both or neither is a usage error (exit 2) in either output mode.

--url limitation (v0.1, by design): no bundled headless browser. --url fetches raw HTML/text via the Python standard library's urllib and hands it to the judge as a text fallback, reasoning over markup and copy rather than the rendered layout. --screenshot is the stronger signal; render the page yourself (Playwright, Puppeteer, or your CI's existing preview-screenshot step) and pass the image.

Real measured CLI overhead

Measured directly against the built slop-eval console script in this package (no scoring call -- --help and the no-API-key error path, which both exit before any network I/O):

Command	Real measured time
slop-eval score --help	0.67s
slop-eval score --url <x> (no API key, fails fast, exit 2)	0.71s
slop-eval score --screenshot <x> (no API key, fails fast, exit 2)	0.52s

These are consistently slower than the npm CLI's own documented numbers for the same three commands (0.49s, 0.26s, 0.19s) -- Python process startup plus importing the anthropic SDK's dependency tree costs more per invocation than Node's. Like the npm package's own README states for its numbers: the actual scored-run latency (a real LLM-judge call, fresh vs. cached) is a target, not a measured result in this environment, since it requires a live ANTHROPIC_API_KEY this environment does not have; the content-hash cache in src/slop_eval/cache.py guarantees a cache hit skips the API call entirely, but the number for a fresh call is not asserted here without having actually run one.

How this port works

--url / --screenshot
     |
     v
LLMJudgeSource            ScreenshotDiffSource (v0.1 stub,
(Anthropic API,             always "not_scored" --
content-hash cached)        no corpus yet)
     |                           |
     +-------------+-------------+
                    v
          composite scorer (average of
          non-"not_scored" 0-10 scores,
          scaled to 0-100)
                    v
          report writer (human text or --json)

• Same rubric, same prompt. src/slop_eval/rubric/v1.json is the same three-category rubric (layout novelty, visual-identity distinctiveness, component-pattern novelty) the npm package scores against, and the judge prompt/instructions text in src/slop_eval/sources/llm_judge.py is a direct port of the TypeScript original's wording.
• Anthropic API via the official SDK. LLMJudgeSource calls anthropic.Anthropic().messages.create(...) with tool_choice locked to a submit_slop_scores tool, the same forced-tool-call pattern the npm package uses, instead of reimplementing raw HTTP calls.
• Content-hash caching. src/slop_eval/cache.py hashes the input bytes (screenshot bytes, or URL + fetched text) and skips the API call on a repeat run against unchanged input -- a correctness guarantee, not just a cost saver, ported from the same reasoning in the TypeScript original's judge-cache.
• Composable RuleSource interface. src/slop_eval/sources/base.py defines the same plugin boundary (name + score()) the npm package uses, so a future rule catalog adapter or second LLM provider can slot in without touching the composite scorer.
• One documented difference: the TypeScript original runs its RuleSources concurrently via Promise.all (Node's I/O model is naturally async). This Python port runs them sequentially in list order. With exactly two v0.1 sources, the result (same findings, same order, same composite score) is identical either way -- sequential is simply the simpler implementation for that source count.

What a score means (and doesn't)

A slop-eval score is a heuristic quality signal from one LLM's read of your UI against a stated rubric. It is not a certification that something is or isn't AI-generated, and a clean score doesn't mean the UI is good by every measure, only that this rubric, at this version, didn't flag it. The same disclaimer is embedded in every report this CLI prints.

CI integration

- uses: actions/checkout@v4
- uses: actions/setup-python@v5
  with:
    python-version: '3.12'
- run: pip install slop-eval-cli
- run: slop-eval score --url "$PREVIEW_URL" --json --fail-below 50 > slop-eval-result.json
  env:
    ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}

Full walkthrough in docs/integrations/ci.md. The npm package additionally ships a ready-made composite GitHub Action (uses: RudrenduPaul/slop-eval/action@main) that wraps the Node CLI and posts a PR comment; the Python distribution has no equivalent bundled Action today (a plain CI step, as above, is the current path).

Security

slop-eval's whole job is calling an external LLM with user-supplied content (a screenshot or fetched page text) -- it never eval()s or exec()s anything read from a scan target, and the only network calls it makes are to the Anthropic API (--screenshot mode) or the one URL you pass it (--url mode). ANTHROPIC_API_KEY is read from the environment only, is never logged, and is never sent anywhere except the Anthropic API via the official anthropic SDK's own transport. See SECURITY.md for the full policy and the private disclosure process. Honest note: this project does not currently publish SLSA provenance, Sigstore signatures, or an SBOM, and has no OpenSSF Scorecard badge -- none of that infrastructure exists yet for either distribution, so it isn't claimed here.

Contributing

See CONTRIBUTING.md. There is no enforced minimum coverage threshold for the Python package today; the bar is that the full pytest suite (pytest from python/) passes and new behavior ships with tests, mocking the Anthropic API rather than calling it for real.

cd python
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
pytest

License

Apache 2.0, see LICENSE.