Sensitive Expression Guard

Purpose

In DSR papers — above all self-referential case studies analyzing the author's own practice — symptom-adjacent vocabulary (diagnosis, spectrum, hyperfocus, etc.) invites readers to misidentify the author's medical status, and rhetorical absolutes can exceed the paper's declared modality ceiling. Some readers skip concessive clauses and short-circuit on keywords alone. This skill addresses the risk across three layers:

1. Mechanical layer (regression prevention): scripts/check_sensitive_expression.py enforces registered rules. Only reappearance of banned_terms (author-finalized prohibitions) is FAIL; everything else is WARN
2. LLM layer (semantic discovery): the machine only sees registered terms and sentence-level co-occurrence. The agent simulates a careless adversarial reader, checking metaphor-level hyperbole, concessive-clause skipping, and out-of-context quotation resilience
3. Accumulation layer (learning): author decisions (fix / allowlist / ban) are recorded with rationale in design/sot/sensitive-expressions.yml and fed back into the mechanical layer

Trigger timings

• When a chapter draft or revision is completed (mechanical layer is mandatory)
• When assemble_manuscript.py aborts a build on FAIL
• When the pre-submission gate returns WARN/FAIL for sensitive-expression
• When the author flags an expression risk (misidentification, overstatement, deficit-model vocabulary) — LLM review plus accumulation

Principles

• FAIL (banned reappearance) always blocks: a regression of an expression the author already calibrated. Never triaged as "no action needed"
• WARN means confirm intent: negations, mentions (inside 「」), and quoted text are already exempted by the script. Remaining WARNs are triaged by the author
• Reading the co-occurrence rule: misidentify fires when a medicalizing term and an author self-reference share a paragraph. A disclaimer pattern in the same paragraph exempts it. Note that a pair-definition living in another chapter does NOT exempt — careless readers may read a chapter in isolation
• The abstract is a separate register: terms in abstract_flagged may be allowed in the body with pair definitions yet still WARN in the abstract (the ## 1. section of paper-outline.md). This institutionalizes the policy of narrowing medically connoted vocabulary in summaries
• Presence check: any paper using medicalizing vocabulary at all must contain a non-diagnostic disclaimer sentence ("regardless of diagnostic status", etc.). Absence of bad words is not sufficient

Procedure

[Step 1: Mechanical scan] ➔ [Step 2: Pattern-match triage] ➔ [Step 3: LLM adversarial-reader review] ➔ [Step 4: Author decision ➔ accumulate in yml]

Step 1: Mechanical scan

# All chapters (abstract section is auto-extracted from paper-outline.md and scanned under the abstract register)
python3 scripts/check_sensitive_expression.py docs/<paper-id>/chapters
# Single chapter
python3 scripts/check_sensitive_expression.py docs/<paper-id>/chapters/<file>.md

Step 2: Pattern-match triage of findings

Output pattern	Classification	Judgment criterion
FAIL/banned (banned reappearance)	Must fix	Regression of an author-finalized calibration. No exceptions
WARN/absolute (absolute/unbounded quantifier)	Confirm	Calibrate to a verifiable scope. Negations/mentions already exempt
WARN/quantifier (vague population quantifier, no citation)	Confirm	Attach a citation or weaken to an existence claim ("exist" not "not a few")
WARN/modality (over-ceiling claim verb)	Confirm	Align with the paper's declared modality ceiling (explore / suggest / existence proof)
WARN/deficit (deficit-model vocabulary)	Confirm	Check consistency with the neurodiversity stance; use established vocabulary such as "barrier"
WARN/misidentify (misidentification coupling)	Confirm	Add a disclaimer in the paragraph or separate the vocabularies; "no action needed" is acceptable if defined elsewhere (record rationale)
WARN/abstract (summary register violation)	Fix-leaning	Replace per the narrowed-abstract policy (body may keep the term)
WARN/disclaimer (missing non-diagnostic disclaimer)	Must fix	The disclaimer sentence is mandatory whenever medicalizing vocabulary is used

Step 3: LLM adversarial-reader review (short-circuits the machine cannot see)

1. Concessive-clause skipping: check how each clause reads in isolation when "not ... but ..." constructions are read partially
2. Metaphor hyperbole: semantically discover rhetorical exaggeration not covered by registered terms
3. Quotation resilience: check how a sentence reads when quoted alone, out of context (e.g., on social media)
4. Author-attribute inference resistance: check whether any path remains by which readers could infer "the author is diagnosed" from the self-referential case description

Step 4: Author decision and accumulation in sensitive-expressions.yml

1. Record the author's decision in docs/<paper-id>/design/sot/sensitive-expressions.yml with a rationale comment

banned_terms:
  - 無尽蔵 # 2026-08-18: calibrated — unbounded absolute, unverifiable
allowlist:
  - 障害|二項対立 # meta-discussion criticizing the deficit model (mention, not use)

2. After fixing the prose, re-run the script and confirm FAIL=0 and that remaining WARNs are justified
3. Append a decision summary to docs/<paper-id>/design/logs/friction-log.md (date, finding, classification, rationale, action)

Related

• Detection script: scripts/check_sensitive_expression.py
• Enforced invocation: assemble_manuscript.py (build aborts on FAIL; bypass with --force), pre-submission gate check_pre_submission.py check #7
• Terminology drift: terminology-consistency skill (coordinates with the §2.5 modality ceiling)
• General WARN/FAIL triage: pre-submission-triage skill

Outputs

• Sensitive-expression triage summary (presented in chat)
• docs/<paper-id>/design/sot/sensitive-expressions.yml (accumulated rules and decisions — source of truth)
• docs/<paper-id>/design/logs/friction-log.md (audit trail of decisions)