v2.0.0.a 20260820 Release

[unixwzrd]

20260818_00 - v2.0.0

Ghostmark Edition: Evidence-Based Local Audit and Safe Cleanup

• Release identity: v2.0.0 is the Ghostmark Edition—named for finding hidden carriers and scheme-specific statistical signals while keeping provenance, probabilistic evidence, typography, and formatting in their proper categories.

• Breaking support baseline: Python 3.10+ is now required; the former NLTK extras, semantic/AI-likeness scoring, ai_score, entropy, repetition, burstiness, stop-word, and type-token-ratio reports are removed.

• Shared findings model: human, JSON, and CSV reports now derive from the 2.0 findings schema with provenance, unicode_security, known_watermark, authorship_signal, typography, and formatting categories, locations, confidence, removability, and planned action.

• Deterministic audit: --metrics now reports reproducible document facts and anomaly inventory only. --dry-run previews the complete cleanup pipeline without writing, and --diff presents a human-readable unified diff.

• Policy-ready reporting: category-aware threshold selection supports CI gates without treating informational typography or Markdown wrapping as a failure.

• Provenance safety: recognized C2PA text carriers are reported independently of AI authorship and remain protected during generic cleanup. --strip-provenance removes only complete recognized local carriers; the tool never dereferences external manifests.

• Markdown safety: opt-in --unwrap-markdown joins soft breaks within paragraphs and list/block-quote containers while preserving separate elements, hard breaks, tables, code, front matter, HTML, and reference definitions. A valid C2PA carrier blocks reformatting unless explicitly stripped.

• Source safety: --source classifies comments, strings, identifiers, and syntax; it only cleans supported comment payloads and checks parsing before and after transformation.

• Offline watermark profiles: optional explicit local profiles report profile-specific detector outcomes for supported KGW/SynthID configurations and deterministic fixtures. No generic detector, proprietary-watermark removal claim, vendor API, or runtime artifact download is provided.

• Calibrated authorship signals: optional local causal-model profiles report paragraph-level likelihood, perplexity, selected-token rank, and top-10 rate separately from watermark evidence. Probabilities require explicit held-out calibration, and scores never trigger cleanup.

• Documentation and research: added a vendor evidence matrix, feasibility ledger, and local research harness guidance that distinguish publicly reproducible schemes from production deployment claims.

• Statistical-text research refresh: documented KGW green-list logit bias, SynthID Text tournament sampling, the downstream trajectory effect of changing an early sampled token, and the separate DetectGPT, Fast-DetectGPT, and Binoculars authorship-detection family. Updated Anthropic's status to reflect its documented model-level Claude text marking while retaining not_publicly_detectable for local use until detector mechanics or compatible tooling are published.

• Release hardening: in-place writes now use a synced, same-directory atomic replacement, retain file permissions while updating modification time, and never overwrite an existing preserved backup. Structured C2PA parsing follows the fixed 2.4 delimiters and URI/data-URI forms, supports single-line comments and front matter, and retains malformed lookalikes. Source mode routes every provenance removal through comment classification so strings and identifiers remain untouched. CI validates the installed console script directly, treats the coverage run as authoritative, builds without a hard-coded wheel filename, and ShellChecks every maintained shell script.

---

This discussion was created from the release v2.0.0.a 20260820 Release.