humanizer-ru

Verifiable hygiene for chat insertions into Russian text

40 regex markers of chat-interface insertion artifacts, of which 38 have a complete record of evidence. Zero class A false positives across 12,314 non-carrier texts, class B — 8, i.e. 0.00065 (Wilson 95% CI 0.0003 to 0.0013; measured 04.09.2026 against a frozen preregistration). Every figure with its date and reproduction command — in the “Project numbers” section.

Insertion-artifact cleaning and fact reconciliation are also available for English text: add --language en to humanizer-clean, humanizer-polish, humanizer-facts or humanizer-report (or --language auto). This does not enable the Russian stylistic heuristics and does not deliver authorship verdicts; the profile preserves code, URLs, Markdown and verifiable facts. The Russian profile remains the default for compatibility.

Who this is for

• Editors and instructors: check text before publishing: humanizer-markers --scan файл.md.
• Developers and CI: a gate for chat-interface insertions: action and contract.
• AI-assistant users: the same check inside an agent environment: MCP in a single config or demo.

Try it in 30 seconds

• Browser demo: nothing to install, text does not leave the browser.
• Report an issue or share your experience: issue in the repository; user text is not transmitted automatically by either the demo or the feedback collector.
• Check a specific insertion:

The shortest path “found → removed → verified”:

pip install humanizer-ru
humanizer-markers --scan input.txt       # найти следы (rc=1 = находка)
humanizer-clean --in-place input.txt     # снять поддержанные артефакты
humanizer-markers --scan input.txt       # убедиться, что остатка нет (rc=0)

For English input add --language en; for mixed input — --language auto. Cleaning does not rewrite style or meaning: verify the result manually and use humanizer-facts diff до.txt после.txt --no-additions after editorial changes.

python -c "open('primer.txt','w',encoding='utf-8').write('Согласно отчёту :contentReference[oaicite:3]{index=3}, рост заявок.\n')"
humanizer-markers --scan primer.txt; echo "rc=$?"
  primer.txt:1 [contentReference] Согласно отчёту :contentReference[oaicite:3]{index=3}, рост заявок.
  Найдено маркеров: 1.
  rc=1

rc=1 means “markers found” — this is the expected result of a check against a sample with an insertion trace, not an error; rc=0 — no traces, rc=2 — input cannot be read (with --json the error envelope is printed to stdout).

Full scenario: find, clean safely, verify, report

1. Find an insertion artifact. humanizer-markers --scan файл.md prints findings with coordinates and class: A — hard copy-paste artifacts from chat interfaces, B — contextual indicators such as invisible characters and hidden layout; rc=1 means findings. The demo page does the same in the browser without installation and highlights the original ranges. humanizer-scan counts soft signs of machine writing: they calibrate the scope of the edit and do not deliver a verdict.

2. Clean safely. humanizer-clean --in-place файл.md runs, in a single command, the before check, removal of supported artifacts, the after check and fact reconciliation; the original is kept in a .bak copy. It removes invisible layer-A markers and visible class-A artifacts outside protected regions (code, frontmatter, URLs, HTML attributes, ZWJ emoji clusters). Style and meaning are not rewritten; unsupported findings such as class B, wiki markup and placeholders remain as explicit residue with exit code 1, and full cleanliness is not claimed. If the invariants of protected regions are violated, the result is not written. Invisible characters are also removed pointwise, by the risk classes from the invisible_classes field of markers.v1.json: humanizer-markers --remove файл.md removes safe automatically, ambiguous — only with the explicit flag --include-ambiguous, dangerous shows and does not remove. humanizer-polish normalizes typography without changing meaning; on markup use the modes --preserve-markup and --typographic.

3. Check the diff. humanizer-clean --diff файл.md prints a unified diff before and after without writing; the --json envelope carries the fact reconciliation and a list of residue. After manual editing, verify facts separately: humanizer-facts diff до.txt после.txt --no-additions compares numbers, dates, URLs, names, quotes, negations and modalities; the lost and changed fields must be empty, otherwise restore the facts to the text.

4. Report the result. humanizer-report до.txt после.txt prepares a report on the edit with fact reconciliation. To pass a finding to a colleague use humanizer-markers --scan --json файл.md: forward the file, line, marker, class and fragment fields, not the whole document. A class-A finding establishes the fact of insertion, not the author: mark the source that the artifact pointed to as “needs verification”. There are no authorship verdicts, neither in the tools nor in the skill — that is the Principal Rule of SKILL.md.

The scenario in a single block:

humanizer-markers --scan вставка.md              # 1: найти; rc=1 = находки есть
humanizer-clean --diff вставка.md                # 2: показать, что будет снято
humanizer-clean --in-place вставка.md            # 2: очистить; оригинал в .bak
humanizer-facts diff вставка.md.bak вставка.md --no-additions  # 3: сверить факты
humanizer-report вставка.md.bak вставка.md       # 4: отчёт о правке

Details of commands and modes — in docs/USAGE.md.

MCP in a single config

{
  "mcpServers": {
    "humanizer-ru": {
      "command": "uvx",
      "args": ["--from", "humanizer-ru==3.36.4", "humanizer-mcp"]
    }
  }
}

The uvx form installs the pinned PyPI release and starts the stdio server. With a local install, pip install humanizer-ru and running humanizer-mcp are equivalent.

Matrix of verified capabilities and boundaries

No leadership claims: as of the recording date there is no comparable external research in the niche (see LEADERBOARD.md). Rows — what is actually verified by gates and tests of the cycle; boundaries — what the surface does not do.

Surface	Known artifact check	Safe cleaning	Fact reconciliation	Machine envelope	Boundary
CLI (humanizer-markers, -polish, -facts, -report)	yes, with coordinates and A/B classes	strip / --preserve-markup / --typographic modes with preservation invariants	humanizer-facts (fact categories)	--json, exit codes per contract	does not check semantics; no authorship verdicts
MCP (humanizer-mcp, tool set from contract.v1.json)	same commands via stdio	same modes via humanizer_polish	humanizer_facts	JSON-RPC envelopes, isError per contract	text does not leave the process
Pages demo	yes, highlights original ranges in the browser	yes, preview + explicit Apply + Undo for supported artifacts	no	copying the report from a single result	offline in the browser, no installation
GitHub Action	insertion gate + auto-fix of the text path (class A)	action_fix outside fenced/code regions	no	gate rc	fix does not touch protected regions
Text skill (SKILL.md)	agent procedures per references	stylistic edits only on explicit request	no	no (skill prose)	no guarantees of naturalness or meaning preservation

What this does NOT do

• Rewritten text: the theoretical detection ceiling under paraphrase [bib:sadasivan2023]; paraphrase zeroes out detectors [bib:dipper2023].

• Natively smooth machine text without artifacts: the documented boundary is the same [bib:sadasivan2023]; population-level detection is possible only on large samples [bib:chakraborty2023], and per-document verdicts are not claimed.

• Short text: fewer signals than words, watermarks and statistics require length [bib:anthropic2026wm], [bib:synthid2024].

• Watermarks without a key: a distortion-free watermark is invisible to a third-party observer by construction [bib:kuditipudi2023]; cryptographic indistinguishability without the key [bib:cgz2023]; the SynthID-Text detector requires a developer key [bib:synthid2024]; Anthropic confirms: without the key the watermark cannot be verified, the detector API is in closed preview [bib:anthropic2026wm].

• [bib:…] keys are listed in research/BIBLIOGRAPHY.md.

• Do not run polish on Markdown or markup: it strips ##, **, «ёлочки» quotes, em dashes; for markup use the --preserve-markup mode.

Why you can trust it

• Methodology and benchmark: figures with confidence intervals.
• Public benchmark: table with CIs, reproduction commands and a “where we are worse” column.
• Threat model and detector boundaries.
• Parity gate for Python and JS rules.
• Self-audit: figures, statuses and errata.
• Status of the last successful run and deploy: status.json on Pages (generated by a deploy artifact from the exact SHA; updated only on green runs).

Installing the skill in browser clients

• The demo works without installation: https://vladimir-human.github.io/humanizer-ru/ — text does not leave the browser.
• Claude.ai and Claude Code: add the skill from the dsh/skills/humanizer-ru directory following the installation instructions in docs/USAGE.md.
• Agent clients with agentskills.io support (opencode, DeepSeek Harness): unpack the text bundle from the release archive.
• Browser extension rejected: a new surface outline (permissions, store review) does not pay for itself; the idea backlog — research/BACKLOG.md.

Directories: Glama MCP · skills.sh.

Same-named projects

On GitHub there are skills with the same name but different content. Snapshot 2026-09-11 (verification: gh repo view <владелец>/humanizer-ru --json stargazerCount):

• ilyautov/humanizer-ru — 333 stars: positions as “removes neural-network signs,” no public registry of numbers; invited to a joint public benchmark (issue 220).
• smixs/humanizer-ru — 154 stars: a deterministic linter; the only same-named project included in LEADERBOARD.md as a candidate (paired run 2026-09-03).
• This project — verifiable chat-interface insertion hygiene: every figure from deterministic snapshots and the fact registry, boundaries — in THREAT-MODEL, false positives — in the benchmark.

Came by the name — choose by verification method, not by stars.

Project numbers

• 58 machine-writing patterns and 40 regex markers (classes A and B).
• Evidence records: 38 out of 40 markers (registry research/fixtures/marker-sources.json).
• Gates: 156 gates in full check_all (145 in --quick); fixtures in tests/fixtures/, documentation is checked by check_docs.py, persona is described in PERSONA.md.

Why it is named this way: the name was inherited from the first function, which stripped the copy-paste layer after a chatbot and returned text to its human look. The second function of the product: diagnostics, highlighting machine traces and explaining the reason for each flag, with no authorship verdicts. Both functions work offline; text does not leave your machine.

Class breakdown of FP, exploratory, outside preregistration F16: class A: 0 cases on 12314 non-carrier texts; class B: 8 cases on 12314, i.e. 0.00065, Wilson 95% CI from 0.0003 to 0.0013; control set of 40 texts: 0 flags; heavy domain S4 legal and official, n=381, volume deficit recorded in the preregistration: 18 cases on 381, i.e. 0.0472, Wilson 95% CI from 0.0301 to 0.0734; denominators: 12354 full F16 corpus, 12314 validation stratum.

More details

• What to feed it and how to rewrite
• Manual installation and usage
• Architecture and patterns
• Security and version differences
• Sources

Regex markers: classes A and B

Class A — hard copy-paste artifacts: service links and citation markers from chat interfaces. Class B — contextual indicators: invisible characters, hidden layout, placeholder fields; a single B match is not enough. Marker class — copypaste_artifacts; marker retirement is possible only on failure within its own class; statuses and dates — in markers.v1.json.

Changelog

Changelog — in CHANGELOG.md and on GitHub Releases.

License

MIT

Project status

Dogfooding — the project checks its own texts with its own rules: the style-marker threshold in delivery files is verified by the gate scripts/check_own_style.py (the current maximum is printed on its run).