academic-humanizer

An academic variant of blader/humanizer (MIT): the same AI-tell detection, plus a register guard so scholarly prose is not damaged.

Setup: run bash scripts/setup.sh once to fetch the pinned upstream base into vendor/ (gitignored). This skill composes vendor/SKILL.md (the 33 patterns) with rules/academic-overrides.md (the academic layer).

When to use: when an academic English draft (paper section, abstract, rebuttal, cover letter) carries AI idioms (em-dashes, significance inflation, promotional language, chatbot residue) and you want to clean the style before submission.

Output: <input>.humanized.md (the final rewrite plus a <!-- HUMANIZE-SUMMARY --> meta comment). It is a suggestion, not a final decision — review before adopting.

Hard rules

1. vendor/ is unmodified. All academic adaptation lives in rules/academic-overrides.md. Never edit the vendored files.
2. Meaning is invariant. Statistics (F, t, p, β, η², d, r, χ², α, N, M, SD, CI), numbers, dates, in-text citations, references, direct quotations, range en-dashes (pp. 12–18, 2019–2021), and field abbreviations (LLM, API, fMRI, ANOVA, IRB) are preserved 100% (see the Do-NOT list in academic-overrides.md).
3. Register is preserved. Formal academic stays formal academic. PERSONALITY AND SOUL is not applied — never inject opinion, first person, colloquialism, "have opinions", or "let some mess in" (upstream §63 self-limit: for reference/technical text, neutral and plain is the correct human voice).
4. Academic protection (EXCLUDE / DEMOTE). Do not strip a single epistemic hedge (may / appears / suggests / is likely), method passive (were analyzed / recruited), or logical connective (therefore / thus / consequently / moreover). Clean up only over-nested hedges and "let's dive in"-style openers.
5. Verify after. For text containing statistics or citations, re-verify the numbers and citations after humanizing (an AI rewrite can silently drift a value or reference). The donot_violations = 0 gate must hold.
6. Change rate above 25% = warning, above 50% = stop and roll back.

Non-goal (ethics guard)

This is a quality tool, not an authorship-disguise tool. Do not use it to hide AI assistance where a venue requires disclosure. academic-humanizer cleans AI idioms from a draft (quality improvement); evading detection is not its purpose. Upstream's README self-limits the same way ("quality improvement, not detection evasion"). Follow your venue's AI-assistance disclosure policy.

Workflow (single inline invocation — based on the vendor/SKILL.md monolith)

Given a <path.md>, perform the following in-session (no separate agent calls, no global registration):

1. Load input

• Read <path.md> → keep the original; count words, sentences, paragraphs.
• Read vendor/SKILL.md → internalize the 33 patterns and the Process loop.
• Read rules/academic-overrides.md → internalize EXCLUDE/DEMOTE/KEEP, the Do-NOT list, the §14 en-dash exception, and the audit guard.

2. Detect patterns (in memory)

• Scan for the 33 vendor patterns. Apply overrides to each match: EXCLUDE (§59 PERSONALITY, §38 voice, §24 over-hedging) → not detected; DEMOTE (§13 passive, §28 signposting, §7 vocabulary, §9 negative parallelism) → only on clusters/overuse; KEEP → base as-is.
• Record each match (§ID, span, fix). Strictly exclude Do-NOT spans (statistics, citations, references, range en-dashes, abbreviations).

3. Rewrite (in memory — vendor Process: draft → audit → final)

• Draft: replace KEEP/DEMOTE matches with natural academic alternatives. Preserve meaning and register; keep the word and paragraph count (5 source paragraphs → 5 result paragraphs).
• Audit: ask "What makes this AI generated?" — but apply the academic-overrides.md Process guard: do not name epistemic hedge, method passive, logical connectives, or legitimate academic vocabulary as an AI tell.
• Final: clean up audit residue and scan for zero em-dashes (§14). Range en-dashes are not hits. Monitor the change rate (25% warning, 50% stop).

4. Self-check (in memory, 6 academic points)

• Run the 6-point self-check in academic-overrides.md (meaning invariant, em-dash zero, register, hedge preserved, passive preserved, no arbitrary additions).
• On any violation, roll back that edit → re-run step 3 for that span (at most once).

5. Output

• Write <path>.humanized.md — the final rewrite plus a <!-- HUMANIZE-SUMMARY v1.0 ... --> block at the end (source/result word counts, change rate, §pattern detections before→after, self-check N/6, grade A–D, 3–5 highlights, remaining findings, Do-NOT-violations = 0 confirmation).
• Return four lines: status (Done. change rate X% / grade Y / self-check N/6), 4–6 §pattern detections, 1 highlight, and (if statistics/citations present) "recommend re-verifying numbers and citations".

Partial re-run

"just this pattern" / "just this paragraph" / "a second pass" → re-process that scope only, through the same single-invocation flow. The humanizer is a pure monolith with no strict pipeline, so only the fast inline mode exists.

References

• Academic override: rules/academic-overrides.md / base monolith: vendor/SKILL.md
• Pattern source: vendor/README.md (33-pattern summary table), Wikipedia "Signs of AI writing"
• Upstream pin and re-fetch: UPSTREAM.md