Social Science Paper De-AI and Academic Quality Calibration

Problem

Social-science manuscripts often trigger AIGC detectors or reviewer comments saying the writing “sounds machine-generated”: repetitive high-level filler, overly regular sentence rhythm, weak argumentation, and vague contribution claims. Letting a model rewrite the whole paper is risky because it can alter numbers, citations, interviewee quotes, or the author’s substantive argument. This skill targets md, txt, and docx social-science drafts, aiming to reduce AI-like traces to submittable academic quality while preserving facts and authorial stance.

How It Works

The workflow follows diagnose–align–edit. First, scan.py creates sentence indexes, an AI-trace score, layer-level work orders, and flagged.md; the model reads only flagged sentences instead of the whole text. It then calibrates terminology, frameworks, and common traps by discipline. Mechanical issues such as filler words and rhythm are suggested for quick acceptance, while structural issues—argumentation, stance, title, contribution—are presented as options and edited only after author confirmation. Changes are written as edits.json patches by sentence ID, then validated by verify.py against numbers, percentages, citations, direct quotes, and word-count tolerance; any violation returns FAIL.

Boundary

It is suited for social-science papers and research proposals in sociology, management, education, law, economics, and history. It is not intended for code, pure engineering experiments, poetry, or fiction. It does not guarantee passing every detector and does not rewrite the author’s conclusion; if the source file changes, rescan is required.