Skip to content
CAI
Software that uses CAICheck a score

guillaumemeyer/watermarks-remover

62.8

Adequate · 16 November 2026

24.2k

lines of production code

Python

primary language

5

measurements over time

CAI band scale
CAI trend line
CAI lens gauges

What this system is

Features

Add Cursor text hygiene skill

A new 'Clean user-facing text' skill has been added to the Cursor integration. This skill instructs the AI to apply text hygiene to natural-language content (such as articles, documentation, and emails) before finalizing it, while strictly preserving facts, formatting, code, and URLs. It also includes guidelines on authorized processing and handling statistical watermarks.

integrations · high confidence

Add black-box watermark-stealing module and prompt corpus downloader

Introduces a new \stealer\ package that implements a black-box attack against SynthID-class text watermarks. The module provides a CLI (\steal.py\) to query watermarked models, collect replies, and derive a reusable scorer \s\*\ that estimates which tokens the watermark boosts. It includes a prompt corpus downloader (\download\_prompts.py\) that fetches data from Hugging Face with checkpointing and retry logic, and a scorer (\scorer.py\) that can be used to demote green tokens during scrubbing or sign-flipped for spoofing.

stealer · high confidence

Distributed service with hardened Docker images

The service is now distributed as a standalone HTTP application with a set of hardened Docker images. The main image exposes an HTTP API on port 8765 and includes system tools like ffmpeg, Ghostscript, and exiftool. Additional optional images are provided for specific watermark removal and detection capabilities: CtrlRegen (pixel watermark removal), MarkDiffusion (image watermark harness), MarkLLM (text watermark detection and rewriting), and a SynthID scorer. All images are hardened by pinning base image digests, using unprivileged users, and restricting file permissions.

service · high confidence

Introduce directory and website audit scripts with SARIF export

Added \audit\_dir.py\ and \audit\_website.py\ to recursively scan local directories and remote sitemaps for AI-provenance marks (C2PA, AI metadata, stylometry). The directory audit supports concurrent scanning and outputs human-readable, JSON, or OASIS SARIF 2.1.0 reports. The website audit downloads URLs from sitemaps and applies the same detection logic, with safety checks against SSRF and entity expansion. Both scripts rely on \audit\_lib.py\ for normalized file inspection and aggregation.

service/scripts · high confidence

New SynthID-text and Polish benchmark corpora with evaluation scripts

Added a reproducible benchmark suite for evaluating SynthID-text watermark removal, including a large English corpus (30 documents) and a new Polish corpus (50 documents) designed to test rewrite robustness in inflectional languages. The release includes shell scripts for running full and smoke tests, along with documentation explaining how to use the seed documents and interpret results.

benchmarks · high confidence

New stylometry and Unicode cleaning scripts for text sanitization

Added two new Python scripts to the \skills/clean-user-facing-text\ module: \score\_stylometry.py\ and \text\_unicode.py\. The stylometry script provides a zero-LLM statistical detector that scores text for AI-generated patterns (such as specific phrase usage, burstiness, and lexical diversity) to flag suspicious content. The Unicode script implements 'Layer A' hardening by stripping invisible control characters, normalizing space homoglyphs, and handling reserved ignorable code points to prevent steganography or broken pastes.

python · high confidence

New text inspection and cleaning scripts for invisible Unicode and stylometry

Added new CLI scripts (\clean\_text.py\, \inspect\_text.py\) and shared utilities (\common.py\) to the \skills/clean-user-facing-text\ area. These tools allow users to inspect text files for invisible Unicode characters, space homoglyphs, and stylometric AI markers, and to clean such content by normalizing Unicode (NFKC), mapping confusable characters, and stripping exotic spaces. The scripts include safeguards against processing binary files, enforce input size limits to prevent resource exhaustion, and support in-place editing with backups. An audit mode is available to generate structured JSON reports of detected issues without modifying the source text.

skills/clean-user-facing-text/scripts · high confidence

Repository restructured as a distributable Claude Code plugin with HTTP service and Docker support

The project has been reorganized to ship as a Claude Code plugin and marketplace offering, introducing a dedicated HTTP service (wr-core) that the agent skills drive over HTTP, allowing the agent host to remain dependency-free. A new \install\_skill.py\ script and Makefile targets support installing skills into Claude Code, Cowork, and Cursor, while a \compose.yaml\ file and Dockerfiles enable local full-stack bring-up with optional harness and heavy profiles for watermark detection and rewriting backends. The repository also adds a \PostToolUse\ hook for deterministic file cleaning, pre-commit hooks for staged-file checking and auto-cleaning, and a \.coderabbit.yaml\ configuration to automate PR reviews.

(repo-wide) · high confidence

Removals

Removal of the 'remove-claude-marks' skill and its cleaning scripts

The 'remove-claude-marks' skill has been removed from the system. This deletion eliminates the capability to strip invisible Unicode characters, statistical text watermarks, and C2PA/AI provenance metadata from text and image files. Consequently, users can no longer use the associated Python scripts (such as \clean\_text.py\, \clean\_image.py\, and \inspect\_text.py\) to perform Layer A (codepoint stripping) or Layer B (statistical rewrite) cleaning operations on their content.

skills/remove-claude-marks · high confidence

Behavioural changes

Cross-platform support for PostToolUse file-provenance hook

The system now reliably executes the PostToolUse hook (which reports or strips AI/C2PA provenance marks after the agent writes files) on Windows, macOS, and Linux. A new Node.js launcher script (\hooks/run\_hook.js\) replaces platform-specific assumptions by probing for available Python interpreters (\py\, \python\, \python3\) and verifying they meet the minimum Python 3.10 requirement before invoking the underlying \hook\_written\_file.py\ script, ensuring consistent behavior regardless of the host OS.

hooks · high confidence

Test coverage

Expanded test coverage for AI watermark detection and cleaning

Added a comprehensive test suite for the watermarks-remover tool, including shared pytest fixtures and sample data for HTML, Markdown, SVG, and stylometry inputs. New tests validate AI generator product-name detection in PNG metadata (e.g., ChatGPT, Midjourney), confidence classification for audit findings, SARIF 2.1.0 export formatting, and multi-worker concurrency in directory audits. The suite also covers AV/ISOBMFF metadata inspection and stripping (MP4, MOV, WAV, MP3, FLAC, AVIF, HEIC), backup preservation logic, binary-input guards for text-only tools, and robust handling of c2patool probe failures.

tests · high confidence

Dependencies

Introduce isolated, pinned dependency files for optional benchmarks and backends

The project now uses separate, pinned requirement files (requirements-dev.txt, requirements-ctrlregen.txt, requirements-markdiffusion.txt, requirements-markllm.txt, requirements-semantic.txt, and requirements-synthid-scorer.txt) to isolate dependencies for optional components such as CtrlRegen, MarkDiffusion, MarkLLM, semantic divergence metrics, and the SynthID scorer. This ensures that optional features do not drift or conflict with the core service, allows targeted security auditing (e.g., via pip-audit) for known vulnerabilities in pinned research backends, and clarifies which packages are required for specific benchmarks or optional pipelines.

(dependencies) · high confidence

Written by cca.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

How this codebase got here

Score

  • CAI 64.1 → 62.8 (-1.3)
  • Rubric changed (rubric-2026.11.2 → rubric-2026.11.4) — scores are not directly comparable.

Lenses

  • Code Health 75.2 → 66.5 (-8.7)
  • Architecture 100.0 → 100.0 (0.0)
  • Maturity 61.6 → 61.5 (-0.1)
  • Readiness 56.0 → 55.9 (-0.1)
  • Security 81.3 → 81.3 (0.0)

Resolved (12)

  • Duplicated block (10 lines × 2) (service/scripts/text_watermark.py)
  • Duplicated block (10 lines × 2) (service/scripts/text_watermark.py)
  • Duplicated block (11 lines × 2) (service/scripts/text_watermark.py)
  • Duplicated block (18 lines × 2) (service/scripts/clean_ctrlregen.py)
  • Duplicated block (18 lines × 2) (service/scripts/clean_image.py)
  • Duplicated block (21–22 lines × 2) (service/scripts/bench_synthid_text.py)
  • Duplicated block (5–8 lines × 2) (service/scripts/bench_synthid_text.py)
  • Duplicated block (6–9 lines × 2) (service/scripts/detect_text_watermark.py)
  • Duplicated block (7 lines × 2) (service/scripts/bench_synthid_text.py)
  • High CVE: [GHSA redacted] (service/scripts/requirements-markllm.txt)
  • Hotspot: service/scripts/clean_ctrlregen.py (service/scripts/clean_ctrlregen.py)
  • Off-boarding risk: anonymized user #1

New (23)

  • Edited copies of one member (3 members, 13+ corresponding lines) (service/scripts/server.py)
  • Edited copy of a member (11 corresponding lines) (service/scripts/server.py)
  • Edited copy of a member (12 corresponding lines) (service/scripts/container_meta.py)
  • Edited copy of a member (14 corresponding lines) (service/scripts/detect_text_watermark.py)
  • Edited copy of a member (18 corresponding lines) (service/scripts/bench_synthid_text.py)
  • Edited copy of a member (19 corresponding lines) (service/scripts/container_meta.py)
  • Edited copy of a member (21 corresponding lines) (service/scripts/synthid_score_server.py)
  • Edited copy of a member (28 corresponding lines) (service/scripts/format_dispatch.py)
  • Edited copy of a member (31 corresponding lines) (service/scripts/clean_audio.py)
  • Edited copy of a member (4 corresponding lines) (service/scripts/text_unicode.py)
  • Edited copy of a member (61 corresponding lines) (service/scripts/image_meta.py)
  • High CVE: [GHSA redacted] (service/scripts/requirements-markllm.txt)
  • Low coverage: install_skill.py (install_skill.py)
  • Low coverage: service/scripts/clean_audio.py (service/scripts/clean_audio.py)
  • Low coverage: service/scripts/clean_file.py (service/scripts/clean_file.py)
  • Low coverage: service/scripts/clean_video.py (service/scripts/clean_video.py)
  • Low coverage: service/scripts/detect_text_watermark.py (service/scripts/detect_text_watermark.py)
  • Low coverage: service/scripts/hook_written_file.py (service/scripts/hook_written_file.py)
  • Low coverage: service/scripts/markdiffusion_harness.py (service/scripts/markdiffusion_harness.py)
  • Low coverage: stealer/init.py (stealer/init.py)
  • …and 3 more

Architecture

  • Unchanged — 0 containers · 1 contexts · 0 edges

Written by cca.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

Survey your own repository

guillaumemeyer/watermarks-remover was measured the same way every project in this corpus was: the same rubric, at a pinned commit, with the result published in full. Point a surveyor at a repository you know and see whether you agree with it.

About this page

  • The score is its most recent published measurement, taken on 16 November 2026 at a pinned commit. It is not a live figure and does not change until the project is measured again.
  • Measured at commit 5a62be6338d2812372b1da440e55c82922f97d53 — the exact code this score is about.
  • Scored under rubric-2026.11.4 — the same rubric and the same method as every other entry in this index.
  • Measured by Canine Code Analyzer (cca.canine.dev) using cca measure-2026.11.83+preprod-a489942cf94b.