sloprefine 1.6.1__tar.gz

sloprefine-1.6.1/.github/workflows/ci.yml

@@ -0,0 +1,35 @@
+name: ci
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+
+jobs:
+  test:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python-version: ["3.11", "3.12", "3.13"]
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python-version }}
+      # mcp too, or the nine MCP server tests skip and CI reports green
+      # without ever having exercised the agent-facing surface.
+      - run: pip install -e ".[dev,mcp]"
+      - run: pytest -q
+      - run: ruff check .
+      - name: dogfood (the docs must pass their own linter)
+        run: sloprefine README.md CONTRIBUTING.md --no-color
+      - name: agent surfaces smoke test
+        run: |
+          sloprefine rules
+          sloprefine prompt > /dev/null
+          sloprefine prompt --profile talk > /dev/null
+          sloprefine check corpus/slop_control.txt --format agent || true
+          sloprefine drift corpus/slop_control.txt corpus/clean_control.txt || true
+          sloprefine prompt --audience expert > /dev/null
+          sloprefine check corpus/choppy_control.txt --max-choppiness 0.17 && exit 1 || true
+          sloprefine audit --ai corpus/audit-demo/machine --human corpus/audit-demo/human

sloprefine-1.6.1/.github/workflows/publish.yml

@@ -0,0 +1,47 @@
+name: publish
+
+# Trusted publishing: PyPI verifies this workflow's OIDC identity, so there is
+# no API token stored in the repository and none to rotate or leak.
+# Configure the publisher once at
+# https://pypi.org/manage/account/publishing/ before the first release.
+
+on:
+  release:
+    types: [published]
+
+permissions:
+  contents: read
+
+jobs:
+  publish:
+    runs-on: ubuntu-latest
+    permissions:
+      # Job-level permissions REPLACE the workflow-level block rather than
+      # adding to it, so contents:read has to be restated here or checkout
+      # runs with contents:none and cannot fetch the repository.
+      contents: read
+      id-token: write          # required for trusted publishing
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+
+      # Publish nothing that does not pass its own suite and its own linter.
+      - run: pip install -e ".[dev,mcp]"
+      - run: pytest -q
+      - run: ruff check .
+      - run: sloprefine README.md CONTRIBUTING.md --no-color
+
+      # The tag must match the version the package actually declares, or the
+      # release and the artifact disagree for the rest of time.
+      - name: tag matches the declared version
+        run: |
+          tag="${GITHUB_REF_NAME#v}"
+          declared=$(python -c "import sloprefine; print(sloprefine.__version__)")
+          test "$tag" = "$declared" || {
+            echo "tag $tag != declared version $declared"; exit 1; }
+
+      - run: pip install build
+      - run: python -m build
+      - uses: pypa/gh-action-pypi-publish@release/v1

sloprefine-1.6.1/.gitignore

@@ -0,0 +1,8 @@
+__pycache__/
+*.py[cod]
+.pytest_cache/
+.ruff_cache/
+dist/
+build/
+*.egg-info/
+.venv/

sloprefine-1.6.1/.sloprefine.toml

@@ -0,0 +1,20 @@
+# sloprefine's own configuration, which doubles as a worked example.
+# "landscape" and "robust" are on the excess-vocabulary list and are also
+# ordinary words in this repository's subject matter.
+[sloprefine]
+allow = ["landscape", "robust"]
+
+# The sentence floor. Off by default in code because it is the one rule with
+# no citation behind it and it refuses a device human writers legitimately
+# use. Enabled here as this repository's own house policy.
+#   runt_mode = "verbless" refuses "One plot." but keeps "It worked."
+#   runt_mode = "all"      refuses everything under the floor
+# Parallelism runs allowed per 1000 words. Documentation enumerates things
+# for a living, so this repo's own prose runs higher than a talk would.
+# A conference talk should sit at 1.0 or below.
+parallel_budget_per_1k = 3.0
+
+min_sentence_words = 5
+runt_mode = "verbless"
+allow_runts = ["Thank you."]
+max_hits = 0

sloprefine-1.6.1/CITATION.cff

@@ -0,0 +1,66 @@
+cff-version: 1.2.0
+title: "sloprefine: a slop refinery for prose"
+message: "If you use this software, please cite it as below."
+type: software
+authors:
+  - family-names: Rumenovski
+    given-names: Filip
+    affiliation: "Wayne State University"
+    email: "frumenovski@gmail.com"
+repository-code: "https://github.com/filiprumenovski/sloprefine"
+license: MIT
+version: 1.6.1
+date-released: 2026-09-22
+abstract: >-
+  A dependency-free prose linter that flags the documented surface markers of
+  machine-generated text, measures cadence and syntactic template reuse, and
+  calibrates its own rule weights against a labelled corpus. Rules carry
+  citations to the studies that motivate them; rules that fail to discriminate
+  on a supplied corpus are reported rather than silently applied.
+keywords:
+  - writing
+  - linter
+  - large language models
+  - stylometry
+  - text diversity
+references:
+  - type: article
+    title: "Delving into LLM-assisted writing in biomedical publications through excess vocabulary"
+    authors:
+      - family-names: Kobak
+        given-names: Dmitry
+    year: 2025
+    journal: "Science Advances"
+  - type: conference-paper
+    title: "Detection and Measurement of Syntactic Templates in Generated Text"
+    authors:
+      - family-names: Shaib
+        given-names: Chantal
+      - family-names: Elazar
+        given-names: Yanai
+      - family-names: Li
+        given-names: "Junyi Jessy"
+      - family-names: Wallace
+        given-names: "Byron C."
+    year: 2024
+    collection-title: "EMNLP"
+  - type: article
+    title: "AI Writers Have a Consistent Stylometric Footprint, but AI Editors Do Not"
+    authors:
+      - family-names: Shan
+        given-names: Zhengyang
+      - family-names: Lee
+        given-names: Yukyung
+      - family-names: Hao
+        given-names: Sophie
+    year: 2026
+    journal: "arXiv:2608.27855"
+  - type: article
+    title: "GPT detectors are biased against non-native English writers"
+    authors:
+      - family-names: Liang
+        given-names: Weixin
+      - family-names: Zou
+        given-names: James
+    year: 2023
+    journal: "Patterns"

sloprefine-1.6.1/CONTRIBUTING.md

@@ -0,0 +1,72 @@
+# Contributing
+
+## Adding a rule
+
+Every rule needs a `fix`: one imperative sentence a model can act on without
+reading the rationale. `why` explains the rule to a person; `fix` is what gets
+shipped into the revision loop and into `sloprefine prompt`. A test asserts
+every rule has one.
+
+Every rule also needs a source, meaning not a vibe but something someone put
+their name on: a study, published corpus statistics, or an argument someone
+wrote down and signed. Add the short key and full
+reference to the docstring at the top of `src/sloprefine/rules.py`, then declare
+the rule with `_rule(...)` and give it a `why` that a writer can act on.
+
+Two more things are required.
+
+1. A positive test in `tests/test_sloprefine.py::test_rule_fires`.
+2. No new hits on `corpus/clean_control.txt`. That file is a regression fence.
+   `test_clean_control_is_silent` asserts zero. If your rule fires there,
+   either the rule is too broad or the fence was wrong; argue for one.
+
+## False positives are the whole game
+
+A prose linter that cries wolf gets uninstalled in a week, and then it protects
+nobody. Prefer a rule that misses half the instances over one that fires on
+ordinary writing. If a word is load-bearing in some field, add it to
+`COMMONLY_LEGITIMATE` rather than dropping it from the lexicon.
+
+## Adding a reader-preference signal
+
+The highest bar in the repository. A positive signal needs a direction
+measured after 2022 by comparing human and machine text against real reader
+judgments. "Good writers do this" is not admissible, and neither is a
+pre-2022 prescription, because prescription is what put the pattern in the
+training data in the first place. That is how the rule of three became a
+detection marker.
+
+Cite the corpus, say who the readers were, and give the direction of the
+effect. Say which
+audience it applies to; `reader.py` takes one because the evidence says the
+two reader clusters want different things. If you approximate a feature the
+source measured with a model, name the proxy in the code and in the output.
+
+Then run `sloprefine audit` against a real corpus pair and report the
+enrichment. A signal that does not separate machine text from human text on
+current models is not a signal yet.
+
+## Adding a stylometric feature
+
+The bar here is different. Features go in `stylometry.py`. They must be
+computable offline, with no language model and nothing that needs
+corpus-level statistics, so that a draft can be checked on a laptop and
+the number still means the same thing tomorrow.
+
+Length-correct anything derived from token counts. Type-token ratio and
+Shannon entropy both fall with document length for arithmetic reasons, so an
+uncorrected feature measures how long the document is and reports it as style.
+
+Do not add a threshold. `stylometry.py` deliberately ships without population cutoffs. The reason
+is in `voice.py`, where a cutoff transferred between corpora turns out to be
+the documented failure mode of deployed detectors, landing hardest on
+non-native speakers. If a feature is only meaningful against a threshold, it belongs in
+the voiceprint layer as a deviation, not in the report as a judgement.
+
+## Running everything
+
+```bash
+pip install -e ".[dev]"
+pytest -q
+ruff check .
+```

sloprefine-1.6.1/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Filip Rumenovski
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.