feat(academic): publisher disclosure, artefacts, and bibliography refuse

SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: authentext
-description: "Remove signs of AI-generated writing from text. Use when editing or reviewing text to make it sound more natural and human-written. Based on Wikipedia's \"Signs of AI writing\" guide. Detects and fixes inflated symbolism, promotional language, superficial -ing analyses, vague attributions, em dash overuse, rule of three, AI vocabulary, negative parallelisms, reasoning failures, and LLM artifacts. Includes severity classification, technical literal preservation, and density-aware detection guidance."
+description: "Repair editorial defects in prose with verified diffs. Remove recurring AI-writing patterns while preserving meaning, voice, citations, and technical literals. Do not optimise against detector scores or claim undetectability. Based on Wikipedia's \"Signs of AI writing\" guide. Detects inflated symbolism, promotional language, superficial -ing analyses, vague attributions, em dash overuse, rule of three, AI vocabulary, negative parallelisms, reasoning failures, and LLM artifacts."
 license: MIT
 metadata:
   version: "3.2.0"

conductor/tracks.md

@@ -19,11 +19,11 @@ This file tracks all major tracks for the project. Each track has its own detail
   Phase 3 PR D is [PR #287](https://github.com/edithatogo/authentext/pull/287).
   _Link: [tracks/editorial-safety-invariants_20260811/index.md](./tracks/editorial-safety-invariants_20260811/index.md)_
 
-- [ ] **domain-and-citation_20260811** (P0) - Clinical, legal, and creative
-      references; academic disclosure upgrade; sourceright/citeweft handoff.
-      [#277](https://github.com/edithatogo/authentext/issues/277). Depends on
-      the safety track.
-      _Link: [tracks/domain-and-citation_20260811/index.md](./tracks/domain-and-citation_20260811/index.md)_
+- [~] **domain-and-citation_20260811** (P0) - Clinical, legal, and creative
+  references; academic disclosure upgrade; sourceright/citeweft handoff.
+  [#277](https://github.com/edithatogo/authentext/issues/277). Depends on
+  the safety track.
+  _Link: [tracks/domain-and-citation_20260811/index.md](./tracks/domain-and-citation_20260811/index.md)_
 
 - [~] **pattern-registry-contracts_20260811** (P1) - Machine-readable
       pattern registry and a stricter Agent Skills contract, extending

conductor/tracks/domain-and-citation_20260811/index.md

@@ -3,7 +3,7 @@
 - **Track ID:** `domain-and-citation_20260811`
 - **Type:** Feature
 - **Priority:** P0
-- **Status:** New
+- **Status:** In progress
 - **Depends on:** `editorial-safety-invariants_20260811`
 - **Current phase:** Phase 1 — Positioning and academic upgrade
 - **GitHub track:** [#277](https://github.com/edithatogo/authentext/issues/277)

conductor/tracks/domain-and-citation_20260811/metadata.json

@@ -4,9 +4,9 @@
   "description": "Add clinical, legal, and creative references, upgrade academic disclosure and artefact rules, route profile families to real files, and hand citation work to sourceright/citeweft.",
   "type": "feature",
   "priority": "P0",
-  "status": "new",
+  "status": "in_progress",
   "created_at": "2026-08-11",
-  "updated_at": "2026-08-11",
+  "updated_at": "2026-08-12",
   "current_phase": "Phase 1: Positioning and academic upgrade",
   "parallel_safe": false,
   "depends_on": ["editorial-safety-invariants_20260811"],

conductor/tracks/domain-and-citation_20260811/plan.md

@@ -2,13 +2,14 @@
 
 ## Phase 1: Positioning and academic upgrade
 
-- [ ] Task: Write failing tests for detector-evasion refusal wording and for
+- [x] Task: Write failing tests for detector-evasion refusal wording and for
       academic disclosure / interface-artefact / tortured-phrase / bibliography
       refuse rules.
-- [ ] Task: Update `conductor/product.md` and the compiled skill description
+- [x] Task: Update `conductor/product.md` and the compiled skill description
       so Authentext is editorial defect repair with verified diffs, not a
-      humanizer that chases detector scores.
-- [ ] Task: Upgrade `src/modules/SKILL_ACADEMIC.md` with publisher-branching
+      humanizer that chases detector scores. `product.md` already stated the
+      refusal; the compiled skill description now matches.
+- [x] Task: Upgrade `src/modules/SKILL_ACADEMIC.md` with publisher-branching
       disclosure, interface-artefact hard errors, tortured-phrase lexicon,
       and out-of-scope reference lists.
 - [ ] Task: Phase Verification & Checkpoint (Refer to `workflow.md`).

references/academic.md

@@ -3,6 +3,9 @@
 ## Navigation
 
 - [Description](#description)
+- [PUBLISHER DISCLOSURE](#publisher-disclosure)
+- [INTERFACE ARTEFACTS](#interface-artefacts)
+- [TORTURED PHRASES](#tortured-phrases)
 - [ACADEMIC VOICE](#academic-voice)
 - [ACADEMIC PATTERNS](#academic-patterns)
 - [CITATION AND REFERENCING](#citation-and-referencing)
@@ -11,7 +14,9 @@
 
 ## Description
 
-This module applies to academic writing: research papers, essays, dissertations, grant proposals, and formal research prose. It maintains scholarly rigor while removing AI voice patterns.
+This module applies to academic writing: research papers, essays, dissertations, grant proposals, and formal research prose. It repairs editorial defects with verified diffs. It does not optimise against detector scores and does not claim to make text undetectable.
+
+Prefer subtractive and reordering edits over generative ones. Do not invent a citation, statistic, finding, or source.
 
 **When to Apply:**
 
@@ -24,17 +29,68 @@ This module applies to academic writing: research papers, essays, dissertations,
 
 **When NOT to Apply:**
 
-- Creative writing
+- Creative writing (load the creative reference when that module is present)
 - Technical documentation
 - Business communications
 
 ---
 
-## ACADEMIC VOICE
+## PUBLISHER DISCLOSURE
+
+Policies differ by venue. Check the named journal or publisher before treating an Authentext pass as undeclared copy-editing. This table was checked on 2026-08-12. Research remains permissioned; do not refresh the table from the web unless the user has granted research permission.
+
+| Venue           | Disclose                                                                                          | Do not treat as automatic exemption                                                                                                   | Where authors usually declare                                                                                                           |
+| --------------- | ------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------- |
+| Elsevier        | AI use that makes substantive changes to sentence structure or organization of a part of the text | Basic grammar, spelling, and punctuation checks                                                                                       | Separate declaration before the references                                                                                              |
+| Springer Nature | Generative editorial work and autonomous content creation                                         | AI-assisted copy editing of human-generated text for readability, style, grammar, spelling, punctuation, tone, wording, or formatting | Methods, Acknowledgements, Introduction, or Preface when declaration is required                                                        |
+| Wiley           | AI used to substantially edit, develop, or translate any part of a manuscript                     | AI tools used solely for spelling, grammar, and general editing                                                                       | Acknowledgements for drafting or editing; Methods for study design, analysis, or figures; also at submission                            |
+| ICMJE           | Any AI-assisted technology used to produce the submitted work                                     | None stated as a copy-editing exemption                                                                                               | Cover letter and the submitted work: writing assistance in Acknowledgements; data collection, analysis, or figure generation in Methods |
+
+**Elsevier warning:** If an edit rewrites sentence structure or reorganizes a passage, report that the change crosses Elsevier's disclosure line even when Springer Nature would treat the same pass as undeclared copy-editing. Do not silently assume the venue is Springer Nature.
+
+Authentext does not file the declaration, invent the tool-use statement, or certify compliance.
+
+---
+
+## INTERFACE ARTEFACTS
+
+Treat leftover chatbot interface language as a **hard error**. Remove it. Do not soften it into scholarly hedging.
+
+Hard-error strings and close relatives:
+
+- `regenerate response`
+- `as an AI language model`
+- `as of my last knowledge update`
+- `I hope this helps`
+- `let me know if you'd like me to expand`
+
+These are Critical. A manuscript that still contains them is not ready for review or rewrite of surrounding prose until they are gone.
+
+---
 
-**Scholarly precision matters.** Academic writing has specific conventions: hedging where appropriate, acknowledging limitations, citing sources properly. The goal is to remove AI patterns while preserving legitimate academic style.
+## TORTURED PHRASES
 
-**Rule:** Keep legitimate academic hedging ("may suggest", "appears to indicate"). Remove AI filler ("it is worth noting that", "it is important to emphasize").
+Tortured phrases are synonym-substituted technical terms that no specialist would write. They are an integrity signal, not a style preference.
+
+**Document threshold:** report the document when **5 or more** distinct tortured phrases appear, matching the Problematic Paper Screener listing rule. Flag each instance even below the threshold. Do not "fix" a phrase by guessing a prettier synonym. If the intended term is already named in the source, restore that term. If it is not, report the span and leave it.
+
+Seed lexicon (documented examples; not exhaustive):
+
+| Tortured form                  | Ordinary term                |
+| ------------------------------ | ---------------------------- |
+| vegetative electron microscop* | scanning electron microscopy |
+| bosom peril                    | breast cancer                |
+| kidney disappointment          | kidney failure               |
+| fake neural organizations      | artificial neural networks   |
+| lactose bigotry                | lactose intolerance          |
+
+`vegetative electron microscop*` matches `vegetative electron microscopy` and `vegetative electron microscope`.
+
+---
+
+## ACADEMIC VOICE
+
+**Scholarly precision matters.** Keep legitimate academic hedging ("may suggest", "appears to indicate"). Remove AI filler ("it is worth noting that", "it is important to emphasize").
 
 ---
 
@@ -53,13 +109,15 @@ This module applies to academic writing: research papers, essays, dissertations,
 - "Experts agree"
 - "It has been demonstrated"
 
+**Action:** Do not invent a paper, year, or DOI to replace the vague attribution. Keep the claim only if the source already supports it. Otherwise cut the claim or report that the citation is missing.
+
 **Before:**
 
 > Studies have shown that climate change significantly impacts biodiversity. Research indicates that immediate action is necessary.
 
 **After:**
 
-> Smith et al. (2023) found that climate change reduced local biodiversity by 40% over two decades. Immediate conservation measures are recommended (Jones, 2024).
+> The manuscript claims climate change affects biodiversity and that action is necessary, but it names no study. Do not invent Smith et al. (2023) or any other source. Ask the author for the citation or cut the claim.
 
 ---
 
@@ -75,7 +133,9 @@ This module applies to academic writing: research papers, essays, dissertations,
 
 **After:**
 
-> Prior work established the foundation for this study (Smith, 2022; Jones, 2023). However, these studies were limited to laboratory conditions. Our field study addresses this limitation.
+> Prior work is cited in the manuscript's existing reference markers. Those studies were limited to laboratory conditions. The present field study addresses that limitation.
+
+Do not invent "Smith, 2022" or "Jones, 2023" to fill a template.
 
 ---
 
@@ -109,6 +169,8 @@ This module applies to academic writing: research papers, essays, dissertations,
 
 > This study demonstrates X under conditions Y. Future work should test whether X holds in real-world settings. The methodology may apply to similar problems in Z domain.
 
+Replace X, Y, and Z only with claims already in the source.
+
 ---
 
 ### Pattern A5: Promotional Abstract Language
@@ -129,7 +191,7 @@ This module applies to academic writing: research papers, essays, dissertations,
 
 **After:**
 
-> We present a method achieving 95% accuracy on dataset X, improving on prior work by 12%.
+> We present the method and the accuracy figures already reported in the manuscript. Do not invent a 95% accuracy claim or a 12% improvement.
 
 ---
 
@@ -147,6 +209,8 @@ This module applies to academic writing: research papers, essays, dissertations,
 
 > We analyzed the data using ANOVA.
 
+Keep "ANOVA" only if the source already names it.
+
 ---
 
 ### Pattern A7: Artificial Signposting
@@ -183,27 +247,31 @@ This module applies to academic writing: research papers, essays, dissertations,
 
 **After:**
 
-> 73 of 100 participants (73%) showed improvement (p < 0.01).
+> Report the count, denominator, and test statistic already in the source. Do not invent "73 of 100" or a p-value.
 
 ---
 
 ## CITATION AND REFERENCING
 
+Reference lists are **out of scope**. Do not generate, complete, reformat, or "fix" a bibliography. Do not invent missing references, DOIs, PMIDs, page ranges, or author lists. Leave the reference list untouched and report the gap.
+
+In-text citation tokens that already exist in the source are protected spans. You may flag a vague attribution (Pattern A1) or a padded span (Pattern A10). You may not rewrite the reference list to match.
+
 ### Pattern A9: Fake or Inaccurate Citations
 
 **Problem:** AI generates plausible-looking but fake or inaccurate citations.
 
 **Severity:** Critical
 
-**Action:** Verify every citation against real databases (Google Scholar, DOI, PubMed).
+**Action:** Do not invent a replacement citation. Do not complete or reformat the reference list. Report the suspect token and stop.
 
 **Before:**
 
 > (Smith et al., 2023) found significant effects.
 
 **After:**
 
-> [Verify: Does Smith et al. 2023 actually exist? Check DOI.]
+> Keep `(Smith et al., 2023)` as written. Report it as unverified. Do not add a DOI, expand the author list, or rewrite the bibliography entry.
 
 ---
 
@@ -219,26 +287,30 @@ This module applies to academic writing: research papers, essays, dissertations,
 
 **After:**
 
-> Global average temperature has increased 1.1°C since 1880 (NASA, 2023).
+> Keep the claim the manuscript already makes. Drop the padded `[1-15]` span only when those markers are not tied to specific sentences in the source. Do not replace them with a new invented citation.
 
 ---
 
 ## SEVERITY CLASSIFICATION
 
 ### Critical (must fix)
 
+- Interface artefacts (`regenerate response`, `as an AI language model`, `as of my last knowledge update`, close relatives)
 - Pattern A9: Fake or inaccurate citations
+- Tortured-phrase documents at the Problematic Paper Screener threshold of 5 or more
 
 ### High (strong AI signals)
 
 - Pattern A1: Vague literature citations
+- Elsevier-crossing sentence-structure or organization edits on an Elsevier manuscript (disclosure warning)
 
 ### Medium (moderate AI signals)
 
 - Pattern A2: Formulaic literature review sections
 - Pattern A4: Generic conclusions
 - Pattern A5: Promotional abstract language
 - Pattern A8: Vague quantitative claims
+- Individual tortured phrases below the document threshold
 
 ### Low (weak AI signals)
 
@@ -253,23 +325,25 @@ This module applies to academic writing: research papers, essays, dissertations,
 
 ### Do
 
-- Cite specific sources with verifiable references
-- Use appropriate hedging for claims
-- Report exact statistics and p-values
-- Acknowledge limitations clearly
-- Use field-standard terminology
-- Follow journal/conference style guides
+- Prefer cuts and reordering over new sentences
+- Keep existing citations, qualifiers, numbers, and limitations
+- Flag vague attributions without inventing a source
+- Warn when an edit would require Elsevier disclosure
+- Report tortured phrases; restore a term only when the source already uses it
+- Follow the named journal's disclosure location when the author asks for a declaration draft they will own
 
 ### Don't
 
-- Use vague citations ("studies have shown")
+- Optimise wording against a detector score
+- Generate, complete, or reformat a reference list
+- Invent Smith et al., a DOI, a p-value, or a sample size
+- Leave interface artefacts in submitted prose
+- Treat Springer Nature copy-editing exemption as universal
+- Use vague citations ("studies have shown") as a rewrite target that you then "solve" with a fake paper
 - Add promotional language ("groundbreaking", "novel")
-- Over-hedge beyond legitimate academic caution
-- Pad citations unnecessarily
-- Use marketing language in abstracts
 
 ---
 
 _Module Version: 3.2.0_
-_Last Updated: 2026-03-03_
+_Last Updated: 2026-08-12_
 _Applies to: Research papers, essays, dissertations, grant proposals, literature reviews_

scripts/compile-skill.js

@@ -45,7 +45,7 @@ const REFERENCE_FILES = {
   foio: 'foio-editorial.md',
 };
 
-const STANDARD_DESCRIPTION = `Remove signs of AI-generated writing from text. Use when editing or reviewing text to make it sound more natural and human-written. Based on Wikipedia's "Signs of AI writing" guide. Detects and fixes inflated symbolism, promotional language, superficial -ing analyses, vague attributions, em dash overuse, rule of three, AI vocabulary, negative parallelisms, reasoning failures, and LLM artifacts. Includes severity classification, technical literal preservation, and density-aware detection guidance.`;
+const STANDARD_DESCRIPTION = `Repair editorial defects in prose with verified diffs. Remove recurring AI-writing patterns while preserving meaning, voice, citations, and technical literals. Do not optimise against detector scores or claim undetectability. Based on Wikipedia's "Signs of AI writing" guide. Detects inflated symbolism, promotional language, superficial -ing analyses, vague attributions, em dash overuse, rule of three, AI vocabulary, negative parallelisms, reasoning failures, and LLM artifacts.`;
 
 /**