Module: De-AI Editing

Trigger: deai, humanize, reduce AI traces, natural writing, tone cleanup

Purpose: Detect likely AI-writing traces in visible prose while preserving LaTeX structure and technical claims.

Commands

uv run python -B scripts/deai_check.py main.tex --section introduction
uv run python -B scripts/deai_check.py main.tex --analyze
uv run python -B scripts/deai_batch.py main.tex --all-sections

Raw Script Output

• deai_check.py emits section-level analysis, trace scores, and optional fix suggestions.
• deai_batch.py supports broader batch inspection across sections.
• The tense category ([Script] LOW) flags present-tense reporting verbs in Methods / Experiments / Results, gated to those sections; see tense-guide.md.
• The overclaim category ([Script] LOW) flags unambiguous causal / firstness / universality phrasing; see over-claim-guard.md.

Skill-Layer Response

• Treat the script output as analysis, not as permission to rewrite the paper by default.
• Return % DE-AI ... style findings or a short risk summary unless the user explicitly asks for source edits.
• Preserve \cite{}, \ref{}, \label{}, custom macros, and math environments.
• Never invent new claims, metrics, baselines, or references while smoothing the prose.

Claim-Evidence-First Humanization

Before reducing AI tone, preserve the academic payload:

• Facts/evidence: numbers, datasets, experiments, figures, tables, citations, equations, and metrics.
• Claims/stance: the paper's real contribution, uncertainty, design choice, and limitation.
• Logic: paragraph role, section role, and claim-evidence map.
• Boundaries: assumptions, scope, missing evidence, and unsupported claims.

Only then remove rhetorical scaffolds such as not merely A, but B, essentially, the key is, The conclusion is:, or vague this/things/factors. Keep a contrast when it names a real baseline, criterion, and evidence; otherwise state the claim directly. The module should not promise lower detector scores or replace venue AI-use disclosure.

Disclosure obligation (read before de-AI editing)

This module improves readability; it does not remove a disclosure obligation. If an LLM had a non-trivial role in producing the paper, the target venue may require you to disclose it (in a dedicated section, a checklist, the acknowledgements, or the cover letter). See ai-disclosure.md for the per-venue policy matrix. Do not treat "reducing AI traces" as a substitute for required disclosure.

Reference: guide.md

Graded mode (--tier) and D1-D5 dimensions

--tier {light|medium|heavy} is opt-in. Without it, the default output is exactly as before. When present, it:

• scales thresholds — light flags fewer items (looser caps), heavy flags more (stricter caps); medium keeps the current thresholds;
• enables the D1 sentence-length check — flags sections whose sentence-length coefficient of variation is suspiciously low (machine-even cadence);
• labels every finding with its AIGC dimension D1-D5 and attaches a one-line teaching note (why detectors flag the pattern).

uv run python -B scripts/deai_check.py main.tex --analyze --tier heavy

The five dimensions are readability-oriented, not tuned to evade any specific detector: D1 sentence-length variety, D2 paragraph structure, D3 information density, D4 connector frequency, D5 term-context matching. Thresholds (including sentence_length.cv_threshold) remain overridable via references/deai/tone-thresholds.yaml.