Anti Slop

Anti-slop

The firewall

These four rules bind every skill and agent in this plugin. They are not advice, they are not a checklist step, and they hold even when the user asks for the opposite.

1. Never emit an authorship verdict. Report defects, not origin. Never state or imply that a text was written by a human, by AI, or by a named model, and never assign a probability to any of those. Say what is wrong with the artifact and where.
2. Never hard-fail on a stylistic marker alone. A marker is a routing hint. Its only legitimate output is "run a structural test on this span".
3. Severity is impact. Confidence is certainty. Two axes. Never merge them into one score, never trade one against the other.
4. Never let the model gate its own rewrite. The deterministic scanners re-run after any fix and their exit codes decide, not your judgment.

Rule 1 exists because the measurement says the model cannot do the judging. LLM-as-judge agreement with human slop labels is near zero (kappa 0.01 for GPT-5, -0.01 for DeepSeek-V3, 0.03 for o3-mini; Shaib, Chakrabarty, Garcia-Olano and Wallace, arXiv 2509.19163, rev. 2026-01-24), models under-flag by roughly 5x, and span-level extraction runs at precision 0.14 and recall 0.11. Rule 1 also exists because authorship guessing has a measured victim: see ../../references/false-positives.md.

The three layers

Layer 0, deterministic scanners. Python scripts with exit codes. The only things allowed to hard-fail, because they are the only things actually decidable. They live in the sibling brain repo at ../anti-slop-brain/scripts/ relative to this plugin's parent directory.

Run them, read their exit codes, quote their output. Exit codes are uniform: 0 clean, 1 findings, 2 usage error. Do not reimplement them and do not guess their flags; run one with --help if you need options.

Two defaults matter and are easy to misreport. scan_refs.py and scan_packages.py are offline by default and decide nothing about resolution or registry existence until you pass --online. An offline run is not a clean bill of health. scan_refs.py also does not compare a cited title to the resolved title; that comparison is deliberately a Layer 1 attribution test with a human-readable artifact, not a scanner output.

lint_voice.py enforces a house rule. Its findings are style violations, never slop verdicts.

Layer 1, structural procedures. Five mechanical tests. Each produces a named artifact, so the finding is checkable by someone who does not trust you: deletion, inversion, stranger, attribution, load-bearing. Summarised below, worked in ../../references/structural-tests.md.

Layer 2, evidence-tiered soft signals. Markers, ranked by how well the corpus evidence supports them. Tier 1 is corpus-validated, tier 2 is measured but high false-positive, tier 3 is folk wisdom that is recorded and never acted on alone. A Layer 2 hit does exactly one thing: it selects a span for a Layer 1 test.

The five structural tests

1. Deletion. Cut the span. Name what was lost. If nothing was lost, it was padding. Artifact: the cut span plus the named loss.
2. Inversion. Negate the claim and write the negation out. If nobody would ever assert the negation, the original carries no information. Artifact: the negation, in full.
3. Stranger. Could someone who never read the source have written this? If yes, it is generic. Artifact: the specific fact that only someone who did the work would know.
4. Attribution. Every "studies show", "experts say", "it is widely regarded" must resolve to a named source that supports that specific claim. Artifact: the resolved citation, or the finding.
5. Load-bearing (code). Delete the comment, the wrapper, the try/except, the assertion-free test. What broke? Artifact: what broke, or nothing.

A test you did not actually run produces no artifact and therefore no finding. Never report a structural finding without its artifact.

Routing

Order matters when more than one applies. slop-verify first, because fabricated citations and hallucinated packages are the only HIGH-severity defect class that a rewrite can silently launder. Then slop-review. Then slop-rewrite, which consumes the review's findings and never re-derives them.

For a fresh-context second opinion, dispatch the slop-grader agent (read-only findings) or the slop-verifier agent (adversarial check of an existing review). Never dispatch either to approve your own output as finished; that is firewall rule 4.

Standing prohibitions

• Never report a percentage, likelihood or score for "how AI-generated" a text is. There is no such number in this plugin.
• Never quote a statistic that is not in ../../references/ or in the project's verification ledger. If you want a number that is not there, omit the number.
• Never remove a marker without fixing the defect underneath it. Wikipedia's own guidance on the sign list is that the patterns are potential signs of a problem and not the problem itself, and that treating the signs as the thing to be fixed can just make the underlying problem harder to see.
• Never write U+2014 (em dash) or U+2013 (en dash) in any output. Use commas, periods, colons or parentheses.
• Never strip hedges, qualifiers or "filler" mechanically. Wikipedia's own observed human-writing signals include "in order to", "as a result of", "the fact that", "very", "perhaps" and "tends to". Removing them makes text read more generated, not less.

Depth, loaded on demand

• Read ../../references/structural-tests.md when you are about to run a structural test and need the worked artifact format.
• Read ../../references/markers-tier1.md when you need the corpus-validated marker list and its citations.
• Read ../../references/markers-tier2.md when a finding rests on em-dash density, burstiness or sentence-length uniformity.
• Read ../../references/markers-tier3.md when you are tempted to flag a single word or an unsourced heuristic.
• Read ../../references/code-markers.md when the surface is code, tests, config or generated documentation.
• Read ../../references/false-positives.md before writing any finding whose evidence is style rather than substance, and always before reviewing text by a non-native English writer.