stopslop

Measure a paper's prose against how a venue actually writes — before and after LLMs existed.

Two baselines, always both: fossil (published up to 2022, so it cannot have been machine-assisted) and poisoned (the last full year plus this one, which cannot be assumed clean). The fossil window is the target; the poisoned window is context, never an excuse — a field that moved because it is being written with the same tools is not a neutral yardstick.

It measures and locates. It never edits your paper: a tool that strips tells on the author's behalf produces text that passes a detector and still reads as machine-written, so check reports where the prose sits and locate gives the character offsets, and the writing stays yours.

Install

The wheel carries the baselines for all seven venues — 105 venue-years of measurements — so there is nothing to crawl and nothing to configure.

pip install stopslop              # pypdf, BSD-3-Clause, no copyleft
pip install "stopslop[pymupdf]"   # adds PyMuPDF (AGPL-3.0, opt-in)

Then, from any directory, in any project:

stopslop check paper.pdf --venue tmlr --out report.md

No system tools required. Everything else is stdlib re and statistics.

Extractors

The extractor is part of the measurement. Across 24 papers the four supported extractors agree to within ~1% on prose metrics — first person 18.1–18.3, -ly adverbs 13.2–13.4, passives 10.3–10.5 — and disagree sharply on layout-sensitive ones: numerals runs 43.4 under pymupdf against 61.4 under pypdf. So measurements are stored per extractor and never mixed, and stopslop check refuses rather than compare across them.

Baselines for all of them ship in the base wheel, so an extra only decides what you can measure with, never what you can read. Pick one with --extractor.

pip install stopslop carries no copyleft: PyMuPDF is AGPL and is opt-in, so taking on that licence is always your deliberate choice.

Use

stopslop check paper.pdf --venue tmlr --out reports/paper.md   # score one paper against TMLR fossils
stopslop check paper.pdf --venue all --out reports/paper.md    # score one paper against combined fossils

--venue all is the default and the wider comparison: 1,033 papers in the fossil window against 20 to 226 for a single venue. More papers, and a broader distribution, so fewer metrics stand out. Naming your venue is the sharper comparison and the one to use when you know where the paper is going.

Recrawl yourself

The shipped corpus is read-only: it is replaced by the next pip install --upgrade, and a crawl landing among it would be indistinguishable from it afterwards. So a corpus of your own is something you start on purpose:

stopslop init                                    # creates ./.stopslop/
stopslop crawl tmlr 2022 2025 2026 --limit 20    # fetch and measure
stopslop stats tmlr 2022                         # inspect one crawled year
stopslop where                                   # which corpus am I using?

.stopslop/pdf/<venue>/<year>/<paper-id>.pdf              the papers
.stopslop/stats/<extractor>/<venue>/<year>/stats.json    their measurements

Two subtrees, split by lifecycle: the PDFs are large, re-fetchable and disposable — rm -rf .stopslop/pdf frees the disk without touching a number — while the measurements are small and are what a report is built from. PDFs sit above the extractor split because one copy of a paper feeds all four.

.stopslop/ is found from the working directory or any parent, the way git finds .git. That is the only switch: no environment variables, no flag, no directory adopted because of its name. Everything that changes a number is written down where a reader can see it.

A corpus of your own replaces the shipped one; the two are never mixed. Two crawls made with different venues, limits and dates are two different samples, and averaging them would put a number in your report that nobody could audit. So you choose the venues, the years and the depth — and until you have crawled a window, check says so instead of comparing against something thin. Delete .stopslop/ to go back to the shipped corpus.

crawl is the only command that touches the network, and it needs no key and no account. Everything else is offline.

Where the baselines come from

stopslop where prints which mode is in force and every resolved path, and every report names its corpus in its header — so a number can always be traced to the measurements it came from.

Scope

Seven venues today — TMLR, NeurIPS, ICML, ICLR, ACL, EMNLP and JMLR — each measured per year, with the fossil window reaching back to 2010 where the papers exist. A venue is one entry in a registry, and stopslop crawl builds any of them from scratch, so a corpus of your own is a command rather than a fork.

Every number is auditable. Each report prints the baseline it used, the extractor, the date, and the n behind each window, and the same measurements sit in plain JSON inside the package. locate gives the character offsets behind a rate, so a claim about your prose can be checked against the sentence that produced it.

It compares prose to a corpus. It does not classify authorship. A metric outside the fossil range is a question about a sentence, not a verdict about a writer — academic English is written well in many registers, and this tool measures distance from one venue's historical distribution, nothing more.

Analytics

The commands above score your paper. These read the corpus itself — how a metric moved between the fossil and poisoned windows, per venue or across them — which is where the baselines the checker uses come from.

stopslop trend tmlr --out reports/tmlr_trend.md                # one venue's drift
stopslop crosstrend --out reports/total_trend_points.md        # venues kept apart
stopslop pooltrend  --out reports/total_trend.md               # venues pooled

crosstrend keeps the venues apart, so agreement between them is visible — a metric moving the same way in six of seven venues is a different kind of evidence from one moving a lot in one. pooltrend merges them into a single pair of windows.

Each of the three takes --plot, which renders the figure beside the table and needs matplotlib. It is deliberately neither a dependency nor an extra: the measurement path has no use for it, so install it yourself if you want figures.

MIT licensed.