Academic Research Skills for Claude Code

简体中文版 | 繁體中文版 | 日本語版 | 한국어

A comprehensive suite of Claude Code skills for academic research, covering the full pipeline from research to publication.

Install in 30 seconds (Claude Code CLI / VS Code / JetBrains, v3.7.0+):

/plugin marketplace add Imbad0202/academic-research-skills
/plugin install academic-research-skills

Then try /ars-plan to walk through your paper structure via Socratic dialogue, or jump to Quick install for prerequisites and the traditional symlink flow.

AI is your copilot, not the pilot. This tool won't write your paper for you. It handles the grunt work — hunting down references, formatting citations, verifying data, checking logical consistency — so you can focus on the parts that actually require your brain: defining the question, choosing the method, interpreting what the data means, and writing the sentence after "I argue that."

Unlike a humanizer, this tool doesn't help you hide the fact that you used AI. It helps you write better. Style Calibration learns your voice from past work. Writing Quality Check catches the patterns that make prose feel machine-generated. The goal is quality, not cheating.

Why human-in-the-loop, not full automation?

Lu et al. (2026, Nature 651:914-919) built The AI Scientist — the first fully autonomous AI research system to publish a paper through blind peer review at a top-tier ML venue (ICLR 2025 workshop, score 6.33/10 vs workshop average 4.87). Their Limitations section enumerates the failure modes that any fully-autonomous AI research pipeline inherits: implementation bugs, hallucinated results, shortcut reliance, bug-as-insight reframing, methodology fabrication, frame-lock, citation hallucinations.

ARS is built on the premise that a human researcher augmented by AI avoids these failure modes better than either alone. Stage 2.5 and Stage 4.5 integrity gates run a 7-mode blocking checklist (see academic-pipeline/references/ai_research_failure_modes.md); the reviewer offers an opt-in calibration mode that measures its own FNR/FPR against a user-supplied gold set.

Zhao et al. (2026-05) audited 111M references across 2.5M papers on arXiv, bioRxiv, SSRN, and PMC. Their conservative estimate is 146,932 hallucinated citations for 2025 alone, with an observed mid-2024 inflection; for the bioRxiv-to-PMC pairing they report 85.3% preprint-to-published persistence. The paper describes "real citations deployed to support claims the cited references do not actually make" as an open challenge. ARS v3.7.1 added trust-chain frontmatter for source provenance; v3.7.3 added locator infrastructure (three-layer citation anchors) for future claim-level audits and surfaces advisory risk signals at cite time (ARS labels the claim-faithfulness gap internally as "L3"; this is ARS terminology, not the paper's). v3.7.x is motivated by Zhao et al.'s corpus-scale findings; corpus-scale evaluation of ARS itself remains future work.

v3.8 closes the second half of the L3 gap. v3.7.3 made every citation carry a locator anchor; v3.8 adds an opt-in audit pass (ARS_CLAIM_AUDIT=1) that fetches the cited source against each anchor and judges whether the claim is actually supported. Five new HIGH-WARN classes (claim-not-supported, negative-constraint-violation, fabricated-reference, anchorless, constraint-violation-uncited) gate-refuse output through the formatter terminal hard gate. Calibration is shipped as a 20-tuple gold set with FNR<0.15 + FPR<0.10 acceptance thresholds; ramp-on plan is deferred to post-calibration evidence per v3.8 spec §5.

Ren et al. (2026, Self-Improvements in Modern Agentic Systems: A Survey) supplies a third, survey-level anchor. Its scientific-discovery synthesis (§7.4) concludes that discovery agents cannot easily verify novelty, correctness, or reproducibility on their own and may exploit weak proxies instead, must manage evidence across heterogeneous tools and literature, and raise governance issues — "scientific writing can also amplify misinformation when the evidence is weak." Its generation-loop chapters (§5.1–§5.2) list human auditing and retained human anchors among the practical safeguards for self-generated evaluation loops, and its historical chapter (§2.2) records the oldest form of the same lesson: the practical success of Lenat's EURISKO depended heavily on the user serving as the external evaluation signal, pruning unproductive heuristic drift — a limitation the survey notes persists in modern agentic systems. ARS cites the survey as design rationale for its human-in-the-loop stance, not as empirical proof that human-in-the-loop pipelines outperform autonomous ones; the survey's actionable deltas for ARS are tracked in #539–#541 and #547–#550.

v3.3 was inspired by PaperOrchestra (Song, Song, Pfister & Yoon, 2026, Google): Semantic Scholar API verification, anti-leakage protocol, VLM figure verification, and score trajectory tracking.

---

Architecture & pipeline

👉 docs/ARCHITECTURE.md — the full pipeline view: flow diagram, stage-by-stage matrix, data-access flow, skill dependency graph, quality gates, and mode list.

The architecture doc supersedes the sprawling pipeline description that used to live here. Everything about what runs in which stage now lives in one place.

Quick install

Prerequisites

• Claude Code (latest; plugin packaging requires recent versions)
• ANTHROPIC_API_KEY exported, or set on first claude run
• Optional: Pandoc for DOCX, tectonic + Source Han Serif TC for APA 7.0 PDF (Markdown output works without either)
• Optional (real Python): The core skills (research / write / review) need no Python — they are prompt-driven. A real Python interpreter is needed only for: the PreToolUse write-scope guard (optional subagent hardening — if no real Python is found it cleanly no-ops and the guard is simply inactive; core skills are unaffected), plus a few opt-in features that shell out to Python (revision-patch mode, the submission-package verifier, and the /ars-cache-invalidate / /ars-mark-read / /ars-unmark-read commands). On Windows, note that python3 is often a non-functional Microsoft Store placeholder rather than real Python; install Python from python.org (or via winget) so the launcher can find a real interpreter. The guard launcher is a POSIX shell script and hooks.json invokes it through bash, so on Windows it needs Git Bash (bundled with Git for Windows). With Git Bash present, a missing real Python degrades cleanly (the guard no-ops, silently). Without Git Bash, Claude Code falls back to PowerShell, which cannot run the .sh launcher at all: the guard is inactive and the PreToolUse hook will log an error per call rather than no-op quietly (accepted degradation — the guard is optional and never blocks your writes, but the hook noise is the trade-off until Git Bash is installed).

Plugin install (v3.7.0+, recommended):

/plugin marketplace add Imbad0202/academic-research-skills
/plugin install academic-research-skills

Verify it works: run /ars-plan and describe a paper you're working on — ARS will start a Socratic dialogue to map out chapter structure. For a single-shot test instead, try /ars-lit-review "your topic".

👉 docs/SETUP.md — full guide: install Claude Code, set up API keys, optional Pandoc/tectonic for DOCX/PDF, cross-model verification (ARS_CROSS_MODEL), and six installation methods (Plugin, project skills, global skills, claude.ai Project, repo-cloned, Claude Science import).

Using Claude Science? The four skills import directly: Skills → Import from GitHub, paste https://github.com/Imbad0202/academic-research-skills, Preview, then Import 4 skills (requires v3.14.0+ of this repo — the importer reads the explicit skill paths in the marketplace manifest). Imports are point-in-time snapshots: re-import after ARS updates. Imported skills carry the ARS methodology (research / writing / review protocols); Claude Code-specific machinery — slash commands, hooks, subagent orchestration — does not transfer. See docs/SETUP.md Method 5 for details.

Using Pi? Install the in-tree, community-maintained wrapper with pi install git:github.com/Imbad0202/academic-research-skills. It keeps the original ARS content authoritative and documents Pi-specific orchestration and hook limitations. See pi/README.md.

Using Codex CLI? Install the sibling distribution instead: Imbad0202/academic-research-skills-codex — same workflow content, Codex-native packaging as a single $academic-research-suite skill with ars-* aliases.

Third-party platforms and integrations that wrap or host ARS are listed in THIRD_PARTY.md — community-submitted and not reviewed or endorsed by the maintainer.

Performance & cost

👉 docs/PERFORMANCE.md — per-mode token budgets, full-pipeline estimate (~$4–6 for a 15k-word paper), and recommended Claude Code settings (Auto mode; Agent Team optional).

Guides & articles

• Academic Writing Shouldn't Be a Solo Act — full pipeline walkthrough (English)
• 學術寫作不該是一個人的事：一套開源 AI 協作工具如何改變研究者的工作流 — 完整使用指南（繁體中文）

---

Features at a glance

• Deep Research — 13-agent research team with Socratic guided mode, PRISMA systematic review, intent detection, dialogue health monitoring, optional cross-model DA, Semantic Scholar API verification.
• Academic Paper — 12-agent paper writing with Style Calibration, Writing Quality Check, LaTeX hardening, visualization, revision coaching, citation conversion, anti-leakage protocol, and VLM figure verification.
• Academic Paper Reviewer — 7-agent multi-perspective peer review with 0–100 quality rubrics (Journal-Fit Reviewer + 3 dynamic reviewers + Devil's Advocate), concession threshold protocol, attack intensity preservation, optional cross-model DA critique / calibration, R&R traceability matrix, read-only constraint.
• Academic Pipeline — 10-stage pipeline orchestrator with adaptive checkpoints, claim verification, Material Passport, optional repro_lock, optional cross-model integrity verification, mid-conversation reinforcement, and score trajectory tracking.
• Data Access Level Metadata (v3.3.2+) — every skill declares data_access_level (raw / redacted / verified_only); enforced by scripts/check_data_access_level.py. Pattern adapted from Anthropic's automated-w2s-researcher (2026). See shared/ground_truth_isolation_pattern.md.
• Task Type Annotation (v3.3.2+) — every skill declares task_type (open-ended or outcome-gradable). All current ARS skills are open-ended.
• Benchmark Report Schema (v3.3.5+) — JSON Schema + lint for honest benchmark comparisons. See shared/benchmark_report_pattern.md.
• Artifact Reproducibility Lockfile (v3.3.5+) — optional repro_lock sub-block on Material Passport. Configuration documentation, not replay guarantee — LLM outputs are not byte-reproducible. See shared/artifact_reproducibility_pattern.md.
• Model Tiering (#517, v3.16+) — optional ARS_MODEL_TIERING switch with two directions: economy (execution-type agents dispatch one tier below the session model, floor Opus-class) and quality-boost (judgment-type agents at integrity gates and final review step up to the frontier tier). Default unset = byte-equivalent to pre-#517 behavior. See shared/model_tiering.md.
• Canonical Cross-Model Handoff Envelope (#527, v3.17+) — the owner→dispatcher→owner blind-checkpoint transport path (#523) now has a machine-stable [CROSS-MODEL-HANDOFF v1] envelope with a normative Python grammar (scripts/cross_model_handoff.py) instead of prose-only enforcement, pinning agreement/divergence/malformed-result routing across all three checkpoint owners. See shared/cross_model_verification.md §"Cross-model handoff envelope".
• Experiment Provenance Intake (#260) — optional experiment_provenance[] on the Material Passport records experiments the scholar ran externally (ARS never runs experiments), and manuscript claims join to them via claim_intent_manifest.planned_experiment_ids[]. The integrity gate (Stage 2.5/4.5) audits each experiment-backed claim against declared provenance — ALIGNED / OVERSTATED / NOT_SUPPORTED_BY_PROVENANCE / PROVENANCE_INSUFFICIENT — without judging whether the experiment itself was correct. A fail-closed experiment_intake_declaration makes "did you run experiments?" an explicit Stage 1 decision (even literature-only runs declare no_experiments_declared). See shared/handoff_schemas.md §"Experiment Provenance Intake (#260)".

Integrity and verification boundary: ARS checks the manuscript and the reported process—including citation existence, claim–source alignment, reported methodology, declared experiment–result alignment, figure/table fidelity, and reporting/process/package conformance. Some checks are sampled or LLM-mediated. ARS does not establish that procedures were actually performed, raw data are authentic, or results reproduce; a consistently reported fabrication can pass these checks. See POSITIONING.md § Integrity checks and the empirical-work boundary.

---

Showcase: real pipeline output

See the complete artifacts from a real 10-stage pipeline run — peer review reports, integrity verification reports, and the final paper:

Browse all pipeline artifacts →

Artifact	Description
Final Paper (EN)	APA 7.0 formatted, LaTeX-compiled
Final Paper (ZH)	Chinese version, APA 7.0
Integrity Report — Pre-Review	Stage 2.5: caught 15 fabricated refs + 3 statistical errors
Integrity Report — Final	Stage 4.5: zero regressions confirmed
Peer Review Round 1	Journal-Fit Reviewer + 3 Reviewers + Devil's Advocate
Re-Review	Verification after revisions
Peer Review Round 2	Follow-up review
Response to Reviewers	Point-by-point author response
Post-Publication Audit Report	Independent full-reference audit: found 21/68 issues missed by 3 rounds of integrity checks

---

Companion: Experiment Agent

If your research involves running experiments (code or human studies) before writing, the Experiment Agent skill fills the gap between ARS Stage 1 (RESEARCH) and Stage 2 (WRITE).

ARS Stage 1 RESEARCH  →  RQ Brief + Methodology Blueprint
        ↓
  experiment-agent     →  run/manage experiments → validate results
        ↓
ARS Stage 2 WRITE     →  write paper with verified experiment results

What it does: executes code experiments (Python, R, etc.) with real-time monitoring, manages human study protocols with IRB ethics checklist, interprets statistics with 11-type fallacy detection, and verifies reproducibility.

How to use together: pause the ARS pipeline after Stage 1, run experiments in a separate experiment-agent session, then bring the results (with Material Passport) back to ARS Stage 2. ARS requires zero modification. See the experiment-agent README for setup instructions.

Stage 1 intake declaration (#260): at Stage 1, ARS detects whether the run will carry experiment-backed claims and sets a fail-closed experiment_intake_declaration on the Material Passport. If you ran experiments externally, the scholar enters one experiment_provenance[] entry per experiment (experiment_id, nested repro_lock, planned_vs_executed[], negative_results[], known_limitations[]) and the declaration is set to experiments_declared; if not, it is set to no_experiments_declared. The declaration is required on every post-#260 passport — a run that touches no experiments still declares no_experiments_declared, so the integrity gate can never be silently bypassed by a forgotten provenance block. The experiment_ids are frozen at this intake point; the writers later reference them via planned_experiment_ids[].

Teaching-side companion: Teaching Skills applies the ARS architecture (skill ensembles, shared contracts, staged gates, a Course Passport) to the teaching side of academic life — course design → lessons → assessment → delivery → reflection; its sotl mode hands classroom-inquiry projects off to ARS deep-research / academic-paper for the publication phase.

---

Usage

Quick Start

# Start a full research pipeline
You: "I want to write a research paper on AI's impact on higher education QA"

# Start with Socratic guidance
You: "Guide my research on AI in educational evaluation"

# Write a paper with guided planning
You: "Guide me through writing a paper on demographic decline"

# Review an existing paper
You: "Review this paper" (then provide the paper)

# Check pipeline status
You: "status"

Individual Skills

Deep Research (8 modes)

"Research the impact of AI on higher education"       → full mode
"Give me a quick brief on X"                          → quick mode
"Do a systematic review on X with PRISMA"             → systematic-review mode
"Guide my research on X"                              → socratic mode (guided)
"Fact-check these claims"                             → fact-check mode
"Do a literature review on X"                         → lit-review mode
"Compare these papers in WHY/HOW/WHAT format"         → three-way-scan mode
"Review this paper's research quality"                → review mode

Academic Paper (11 modes)

"Write a paper on X"                                  → full mode
"Guide me through writing a paper"                    → plan mode (guided)
"Build a paper outline"                               → outline-only mode
"I have a draft, here are reviewer comments"          → revision mode
"Parse these reviewer comments into a roadmap"        → revision-coach mode
"Write an abstract for this paper"                    → abstract-only mode
"Turn this into a literature review paper"            → lit-review mode
"Convert to LaTeX" / "Convert citations to IEEE"      → format-convert mode
"Check citations"                                     → citation-check mode
"Generate an AI disclosure statement for NeurIPS"     → disclosure mode
"Audit my rebuttal draft against the reviews"         → rebuttal-audit mode

Academic Paper Reviewer (6 modes)

"Review this paper"                                   → full mode (Journal-Fit Reviewer + R1/R2/R3 + Devil's Advocate)
"Quick assessment of this paper"                      → quick mode
"Guide me to improve this paper"                      → guided mode
"Check the methodology"                               → methodology-focus mode
"Verify the revisions"                                → re-review mode
"Calibrate this reviewer against my gold set"         → calibration mode

Academic Pipeline (Orchestrator)

"I want to write a complete research paper"           → full pipeline from Stage 1
"I already have a paper, review it"                   → mid-entry at Stage 2.5 (integrity first)
"I received reviewer comments"                        → mid-entry at Stage 4

Pipeline ends with Stage 6: Process Summary — auto-generates a paper creation process record with 6-dimension Collaboration Quality Evaluation (1–100 scoring).

Supported Languages

• Traditional Chinese (繁體中文) — default when user writes in Chinese
• English — default when user writes in English
• Bilingual abstracts (Chinese + English) for academic papers

Using a different language? Socratic mode (deep-research) and Plan mode (academic-paper) use intent-based activation — they detect the meaning of your request, not specific keywords. This means they work in any language without modification.

However, the general Trigger Keywords section (which determines whether the skill is activated at all) still lists English and Traditional Chinese keywords. If you find the skill isn't activating reliably in your language, you can add your language's keywords to the ### Trigger Keywords section in each SKILL.md file to improve matching confidence.

Supported Citation Formats

• APA 7.0 (default, including Chinese citation rules)
• Chicago (Notes & Author-Date)
• MLA
• IEEE
• Vancouver

Supported Paper Structures

• IMRaD (empirical research)
• Thematic Literature Review
• Theoretical Analysis
• Case Study
• Policy Brief
• Conference Paper

---

Skill Details

Per-agent responsibilities and per-stage artifacts now live in docs/ARCHITECTURE.md. Version numbers are anchored here so release metadata stays in one place.

Deep Research (v2.11.0)

13-agent research team. Modes: full, quick, review, lit-review, three-way-scan, fact-check, socratic, systematic-review. Full agent roster and artifacts: see ARCHITECTURE.md §3.

Academic Paper (v3.2.0)

12-agent paper writing pipeline. Modes: full, plan, outline-only, revision, revision-coach, abstract-only, lit-review, format-convert, citation-check, disclosure, rebuttal-audit. Output: MD + DOCX (via Pandoc when available) + LaTeX (APA 7.0 apa7 class / IEEE / Chicago) → PDF via tectonic. Full agent roster and per-phase responsibilities: see ARCHITECTURE.md §3.

Academic Paper Reviewer (v1.10.0)

7-agent multi-perspective review with 0-100 quality rubrics. Modes: full, re-review, quick, methodology-focus, guided, calibration. Decision mapping: ≥80 Accept, 65-79 Minor Revision, 50-64 Major Revision, <50 Reject. First-round review panel vs. contract-governed re-review dispatch boundary: see ARCHITECTURE.md §3 Stage 3 / Stage 3'.

Academic Pipeline (v3.19.0)

10-stage orchestrator with integrity verification, two-stage review, Socratic coaching, and collaboration evaluation. Pipeline guarantees: every stage requires user confirmation checkpoint; integrity verification (Stage 2.5 + 4.5) cannot be skipped; R&R Traceability Matrix (Schema 11) independently verifies author revision claims. v3.4 added the Compliance Agent (PRISMA-trAIce + RAISE) at Stage 2.5 / 4.5. v3.5 adds the Collaboration Depth Observer (collaboration_depth_agent, advisory only — never blocks) at every FULL/SLIM checkpoint and at pipeline completion. MANDATORY integrity gates (2.5 / 4.5) explicitly skip the observer so compliance checks are not diluted. Based on Wang & Zhang (2026), IJETHE 23:11. Stage-by-stage matrix with agents, artifacts, and gates: see ARCHITECTURE.md §3.

---

v3.0 Optimizations: What We Discovered About AI's Structural Limits

What happened

While using ARS to write a reflection article about AI in higher education, I ran into three structural problems that no amount of prompt engineering could fix:

1. Frame-lock: I asked the AI to run a devil's advocate debate against its own thesis. It did — four rounds, each more refined than the last. But every round stayed inside the frame I'd set. The DA attacked arguments, never premises. It never asked "are we even discussing the right question?" This is the same pattern that caused the 31% citation error rate in v2.7's stress test: the verifying AI and the generating AI share the same cognitive frame.

2. Sycophancy under pushback: Every time I challenged the DA's attacks, it conceded too quickly. It retracted findings faster than it launched them. The model's training rewards conversational harmony — so "the user pushed back" was treated as evidence that the attack was wrong, when often it just meant the user was persistent.

3. Intent misdetection: The Socratic Mentor kept trying to converge and produce deliverables ("Want me to write this up?") when I was still exploring. It couldn't distinguish "the user wants a deep philosophical discussion" from "the user wants an RQ brief." Both look like engagement, but they need opposite AI behaviors.

What we changed (v3.0)

Devil's Advocate — Concession Threshold Protocol (deep-research + academic-paper-reviewer)

• DA must now score every rebuttal on a 1-5 scale before responding
• Concession only allowed at score ≥4 (rebuttal directly addresses core attack with evidence)
• Score ≤3: hold position and restate the original attack
• Anti-sycophancy rules: no consecutive concessions, concession rate tracking, frame-lock detection after each checkpoint

Socratic Mentor — Intent Detection Layer (deep-research)

• Classifies user intent as exploratory vs. goal-oriented at dialogue start and every 3 turns
• Exploratory mode: disables auto-convergence, raises max rounds to 60, prohibits "want me to summarize?" prompts
• Goal-oriented mode: standard convergence behavior
• Anti-premature-closure rules: in exploratory mode, the user decides when to stop

Socratic Mentor — Dialogue Health Indicator (deep-research)

• Silent self-assessment every 5 turns on three dimensions: persistent agreement, conflict avoidance, premature convergence
• Auto-injects challenging questions when agreement pattern detected
• Invisible to user (to prevent gaming), but log available for post-session review

Why this matters

These optimizations don't solve AI's structural limits — they make the limits visible and manageable. The DA will still eventually concede if pushed hard enough. The Socratic Mentor will still have some convergence bias. But now there are explicit checkpoints that slow down the sycophancy, force the DA to justify concessions, and prevent the Mentor from wrapping up before the user is ready.

The deeper lesson: AI literacy isn't about learning to use AI as a tool, following ethics rules, or fearing AI risks. It's about engaging AI deeply enough to discover its structural limits yourself — and your own thinking limits in the process.

---

License

This work is licensed under CC-BY-NC 4.0.

You are free to:

• Share — copy and redistribute the material
• Adapt — remix, transform, and build upon the material

Under the following terms:

• Attribution — You must give appropriate credit
• NonCommercial — You may not use the material for commercial purposes

Attribution format:

Based on Academic Research Skills by Cheng-I Wu
https://github.com/Imbad0202/academic-research-skills

---

Contributors

Cheng-I Wu (吳政宜) — Author and maintainer

aspi6246 — Contributor. The v3.1 optimization was inspired by patterns from Claude-Code-Skills-for-Academics: read-only constraint pattern, anti-pattern codification as first-class design, cognitive framework approach (teaching "how to think" not just procedures), and lean skill size philosophy.

mchesbro1 — Contributor. Originally proposed and drafted the IS Basket of 8 journals for academic-paper-reviewer/references/top_journals_by_field.md (Issue #5).

cloudenochcsis — Contributor. Extended the IS section from the Basket of 8 to the full Senior Scholars' Basket of 11 — adding Decision Support Systems, Information & Management, and Information and Organization (Issue #7, PR #8). Sourced from the AIS Senior Scholars' List of Premier Journals.

eltociear (Ikko Eltociear Ashimine) — Contributor. Translated the Japanese README (README.ja-JP.md) (PR #161).

xpfo-go (xpfo) — Contributor. Translated the Simplified Chinese README (README.zh-CN.md) (PR #181).

devCharlotte — Contributor. Translated the Korean README (README.ko-KR.md) (PR #469).

Yaobin29 — Contributor. Proposed reviewer-response tooling in PR #433; the deep-research three-way-scan mode and the academic-paper rebuttal-audit mode (rescued from the PR's audit concept) were integrated from that contribution in v3.12.1.

ktao732084-arch — Contributor. Expanded the academic-paper disclosure system with nine medical-publishing policy targets, target-specific required-fact intake, and fail-closed standalone rendering (Issue #596, PR #599); expanded the EQUATOR clinical-reporting reference with condensed CARE, STARD and TRIPOD+AI guidance plus a fail-closed study-design routing sequence (Issue #594, PR #601); and designed and contributed the standalone Chinese-literature resolver, API protocol, and synthetic transport-fixture suite (Issue #595, PR #600).

---

Changelog

v3.19.0 (2026-07-22) — Revision-round claim-drift guards, PDF read-integrity preflight, read-scope attestation

Added: three advisory-or-opt-in integrity layers plus a launcher fix. Revision-round claim-drift guards (#569/#570): a claim-strength ladder ("no silent move along is associated with < predicts < causes without an authorizing roadmap item") wired into revision drafting and a new advisory Phase E6, plus a deterministic numeric/citation token-conservation checker — together they stand watch over the epistemic and token halves of the #390 honest-claim residual (a touched block's interior had no fidelity check; whether the guards reduce the measured drift awaits the #652 re-measurement). Baseline measured on the current frontier model first (evals/heldout/revision_claim_drift/), mechanism shape credited to Yila-AI/sci-ssci-skills. PDF read-integrity preflight (#512): a three-signal page-count cross-check so a truncated / mispaginated PDF read cannot mint an apparently-valid page anchor. read_scope attestation (#513): an optional honest-coverage declaration on the human-read ledger (full_text / sections / abstract_only / toc_only) that makes the finalizer's citation promotion read-scope-aware. Launcher watchdog fix (#545): removes a pipe-stall that blocked every healthy PreToolUse write-scope-guard call for the full wall-clock bound. Suite → v3.19.0; the three underlying skill versions are unchanged.

v3.18.0 (2026-07-18) — Self-improvement survey integration: advisory quality layers, risk-stratified claim gate, cross-model reviewer & judge tracks

Added: eight quality mechanisms motivated by Ren et al. (2026, arXiv:2607.13104, Self-Improvements in Modern Agentic Systems: A Survey): per-sub-question scope bindings + a Phase E scope-conformance advisory (#547) and search-bounded novelty claims + an E5 novelty classification (#548) — both advisory-only, displayed per-row at the MANDATORY integrity checkpoints; risk-stratified Stage 2.5 claim verification (100% of HIGH-IMPACT claims + a random sentinel, extending the #518 reference tiers to claim level, #549); cache-through wired into the citation-verification gate with an age-based staleness advisory + opt-in live re-validation (#541, closing the v3.11 Delta-2 forward-decl); a consent-gated cross-model reviewer track — one seat of the fixed five-seat panel on the second model family (#540) — and re-review judge independence with a transparent Judge Record (#539); a metamorphic routing/gate robustness eval seed set (#550), which also shipped the reviewer skill's missing zh-TW trigger aliases; and the survey itself as a third human-in-the-loop literature anchor (#542). Independently of the survey track, plugin installs also gain the #544 SessionStart update-available reminder (/plugin update announce when behind; ARS_UPDATE_CHECK=0 kill switch). academic-pipeline tracks the suite at v3.18.0; the other three skill versions are unchanged.

v3.17.0 (2026-07-16) — Pipeline boundary semantics, canonical cross-model handoff envelope, executable panel checker

Fixed: two under-specified pipeline boundaries closed (#528) — Stage 5's "before finalization: always MANDATORY" now names exactly one checkpoint (the entry gate between Stage 4.5 PASS and Stage 5 dispatch), and Stage 6 gains a defined terminal-acknowledgement vocabulary (finish/end/done/confirm) plus an explicit decline path; all five pipeline surfaces now carry whole-file sha256 content locks (#529) so any further prompt-surface drift fails CI until the hash is updated in the same commit. Blind-checkpoint transport moved to the dispatching layer (#523) — the Bucket A checkpoint owners were being told to execute cross-model transport themselves, which is unexecutable under the runtime Bash deny; the dispatching layer now owns the transport call. Added: a canonical [CROSS-MODEL-HANDOFF v1] envelope + normative Python grammar (#527) replaces prose-only enforcement of the owner→dispatcher→owner transport path, pinning agreement/divergence/malformed-result routing across all three checkpoint owners. A defrift lock for the #514 tools allowlist (#524, 74 mutation tests) closes the drift path where a symmetric edit to an agent + its mirror could silently re-add Bash. An executable sprint-contract panel checker (#510) recomputes both v3.6.2 decision layers from the primary artifacts and catches a transcription error in the majority-vote formula. A machine-readable degradation registry (#511 Part A) indexes every graceful-degradation mechanism in the suite, plus a hermetic transport-fixture integration test for the citation-verification gate (#511 Part B) exercising all four resolver clients end-to-end against checked-in synthetic API bodies. academic-pipeline tracks the suite at v3.17.0; the other three skill versions are unchanged.

v3.16.0 (2026-07-12) — Model tiering, cross-model gate hardening, WP advisory sharpening

Added: opt-in model tiering (#517) — a new ARS_MODEL_TIERING switch with two directions (economy dispatches the 13 execution-type agents one tier below the session model, floor Opus-class; quality-boost steps the judgment-type agents at the integrity gates and final-review surfaces up to the frontier tier); default unset stays byte-equivalent, with the frozen 39-agent classification pinned by a new manifest + lint. Cross-model gate hardening (#518) — risk-stratified verification sampling (HIGH-IMPACT references verified 100% at both gates), blind disagreement checkpoints at the two irreversible decisions (design freeze + final editorial decision), an id-status allowlist for verifier model ids, and a promotion bakeoff protocol; the once-planned generic 6th reviewer is retired, not deferred. GPT-5.6 Sol listed as a provisional cross-model verifier with explicit reasoning-effort control (#515). Korean trigger keywords + routing boundary fixtures by devCharlotte (#452/#509). A CARS introduction-rhetoric + title-crafting reference for the paper writer (#500). Changed: the WP research-question advisory generalizes beyond its 20-shell table via the noun-swap test (#501) and a sharpened exemption clause that catches decorated title-form shells (#505) — held-out miss rate 0.34–0.38 → 0.094 with false-fire 0/16 preserved; reviewer calibration protocol now documents the LLM-as-judge leniency direction (FARS anchor, #484); OpenAlex API-key auth + budget-aware 429 handling + arXiv ToU-aligned backoff (#495/#496). Docs: THIRD_PARTY.md community directory (#497/#498). academic-pipeline tracks the suite at v3.16.0; the other three skill versions are unchanged.

v3.15.0 (2026-07-04) — Release-gate hardening, prompt-debt retirement round 2, defrift locks

A release-discipline-and-hygiene release; no skill-behavior changes. Added: three CI gates — the CHANGELOG-covers-merges pre-tag gate (#483), version-consistency invariants 9-11 plus a tag-time re-run gate (#487), and a command-invariants gate pinning the SessionStart announce list to the actual 16-command inventory (#486) — plus two defrift locks: the Phase Boundary enforcement sentence is pinned verbatim across all 23 Bucket A agent blocks, and the SETUP cross-model examples are pinned to each other and to the canonical model tables (#491 → #492). Changed: prompt-debt retirement round 2 deep-scans the 17 agents the first pass deferred (#489 → #490): two live self-contradictions fixed in both socratic_mentor agents (stale 15-round quit rules vs the documented typical 20-30-round run), the repo-wide stale enforcement-status sentence corrected at 29 surfaces, few-shot and duplicated-process scaffolds trimmed across 7 agents — verified by a 4-batch parallel audit + independent codex cross-model challenge; audit report under audits/. Fixed: DOI badge served from shields.io (#482). academic-pipeline tracks the suite at v3.15.0; the other three skill versions are unchanged.

v3.14.0 (2026-07-02) — Claude Science importability, eval-comment rendering, prompt-debt retirement

A portability-and-polish release; no skill-behavior changes. Added: Claude Science importability — the marketplace manifest declares explicit skill paths, so GitHub-API importers that cannot traverse the symlinked skills/ directory (Claude Science "Import from GitHub", Windows checkouts) now find all four skills; verified end-to-end on Claude Science, with an import guide in README + SETUP (#480). Eval-harness PR comments render as a one-line verdict + per-task table with the raw JSON folded into <details>, replacing the raw report dump — display layer only, gate logic byte-identical (#479). Changed: expired writing-harness scaffolds retired from four writer-surface agents after the 2026-07 harness-retirement audit (#476/#477 → #478, net −111 prompt lines); a remind-don't-block Platform Port Reminder surfaces the platform-ports policy when a PR adds a new top-level directory (#473). Docs: native-reviewed Korean README by devCharlotte (#469/#471); GitHub Copilot repository instructions (#465); auto permission mode recommended over Skip Permissions (#464). The accumulated [Unreleased] backlog (16 entries whose code shipped before the v3.13.0 tag — diff/patch revision mode #390, submission-package verifier #394, eval gold sets #215/#216, and more) is rolled into the versioned record; see CHANGELOG.md. academic-pipeline tracks the suite at v3.14.0; the other three skill versions are unchanged.

v3.13.0 (2026-06-18) — Hook portability, provider-agnostic verification, guard correctness

A minor release hardening the install/runtime surface and extending cross-model reach. Fixes: the write-scope guard no longer false-denies a user's own CLAUDE.md under the git-clone + symlink install layout (#459, closing the residual half of #448/#449 — CLAUDE.md is documentation, not a load-bearing enforcement file, so it leaves the infra-protected list while every load-bearing file stays protected); Windows Python hook portability + graceful no-Python degradation via a cross-platform hooks/run_guard.sh launcher that rejects the 0-byte Microsoft Store python3 stub and never spams the hook log (#454); draft_writer dual-phase static union documented + POSIX-safe Windows path matching (#451). Added: provider-agnostic cross-model verification accepting OpenAI-compatible endpoints (MiMo, DeepSeek, self-hosted) alongside grounded first-party OpenAI, which is never silently downgraded (#455); an opt-in Socratic adjacent-framing probe (STORM-borrowed perspective expansion, ARS_SOCRATIC_ADJACENT_PROBE=1, default OFF, prose-layer only — deep-research 2.10.0 → 2.11.0) (#461). academic-pipeline tracks the suite at v3.13.0; academic-paper and academic-paper-reviewer are unchanged. See CHANGELOG.md for the per-issue detail.

v3.12.1 (2026-06-15) — Reviewer-response triage modes (PR #433 integration)

A patch release folding the genuinely-novel parts of an external contribution into existing skills as modes, per ARS's mode-based architecture. New modes: deep-research three-way-scan — a lightweight WHY/HOW/WHAT paper-comparison triage between quick and lit-review, with per-paper shortlists + a cross-paper synthesis (deep-research 2.9.4 → 2.10.0); academic-paper rebuttal-audit — standalone advisory QA of an author's existing rebuttal/response draft against the reviewer comments (per-comment coverage table + gap list + tone/evidence/misread risk flags), which generates nothing and explicitly suppresses Schema 11 / Material Passport writes / ready_to_submit when run standalone (enforced by a check_rebuttal_audit_guard() lint with mutation coverage); plus a revision-coach scope extension to pushback/disagreement posture and non-journal scopes, and /ars-3w + /ars-rebuttal-audit slash commands. Routed by input shape: reviewer comments AND a draft → rebuttal-audit; comments only → revision-coach. Integrated from @Yaobin29's PR #433. Suite mode count 25 → 27 (still 4 skills). See CHANGELOG.md for the per-issue detail.

v3.12.0 (2026-06-08) — Kong auto-research feature track: experiment provenance, figure fidelity, cross-paper contradiction, partial-evidence decomposition

A minor release shipping the Kong et al. (2026, arXiv:2605.18661) auto-research feature track plus the partial-evidence-trap decomposition work, each reviewed and merged independently. New features: Experiment Provenance Intake + claim→experiment alignment — a schema-first evidence-ledger layer for experiment-backed claims, intake-and-alignment only (the scholar runs experiments externally; ARS never executes them) (#260); a Figure/Table Fidelity Gate that checks whether a caption's interpretation follows from the data and whether the manuscript cites the artifact for a claim it supports (#261); a structured Cross-Paper Contradiction inventory making assessed paper-pairs enumerable for scholar confirmation (#262); and sub-claim decomposition before judgment in both the citation judge (#213) and the editorial synthesizer (#214), closing the §F.3.2 partial-evidence trap on both layers. Guidance + interpretive layer: concise-output + pressure-stable boundary reinforcement across the report-producing reviewers (#274); a same-family / rubric-aware calibration epistemic note (#273); the retrieved-content instruction/data boundary stated as a standing principle (#367). Negative scope: the Kong META (#255) closed with a "Rejected mechanisms" section in POSITIONING.md enumerating the five autonomous mechanisms ARS does not do, plus two Tier D design-lesson docs. Release-discipline lint: version-consistency invariants 5–7 (#357) and ARCHITECTURE component-version policing (#345). Plus correctness fixes across the cross-model grounding guards (#346 / #349 / #351), the citation-gate cache key and rationale bounding (#359 / #360 / #361), the eval gold set (#250), and ACL/EMNLP disclosure regrounding (#242). The new schemas, manifest field, and all invariants are additive and backward-compatible. academic-pipeline tracks the suite at v3.12.0; the other three skill versions are unchanged. See CHANGELOG.md for the per-issue detail.

v3.11.1 (2026-06-06) — Post-ship correctness, hardening & provenance rollup

A patch release rolling up the post-ship fixes surfaced after v3.11.0, each reviewed and merged independently: a cross-model consent-gate extension to the integrity-verification + collaboration-depth paths (#322), a per-entry OpenAlex + Crossref backfill parallelization (#138), and seven correctness/hardening fixes across the citation-existence gate, the v3.10 policy layer, the eval harness, the domain evidence profiles, and the #310 security-boundary edge cases (#323 / #327 / #328 / #329 / #331 / #332 / #333) — including two P1 fixes (#327 domain-profile activation on the no-handoff path, #328 the eval-harness per-class threshold gate). No new features and no breaking schema changes. See CHANGELOG.md for the per-issue detail.

v3.11.0 (2026-06-04) — Deterministic citation verification gate (#182)

Adds a deterministic citation-existence verification gate that runs independently of LLM peer review. Every cited reference is cross-checked against up to four bibliographic indexes — Semantic Scholar + OpenAlex + Crossref + the new arXiv resolver (scripts/arxiv_client.py, no API key needed) — and a per-citation lookup_verified status ({true, false, unresolvable}) is written to a unified summary, so a fabricated citation with a provably-bogus DOI/arXiv ID is caught by lookup rather than by hoping a reviewer agent notices. The gate inherits the v3.10 terminal_policies opt-in model: detection always runs, but a lookup_verified == false row is terminal only when a user opts into terminal_policies.citation_existence == strict — default behavior is advisory and /ars-mark-read-acknowledgeable. false is narrowed to ID-keyed unmatched (an exact DOI/arXiv lookup that provably fails), so legitimately-unindexed humanities / non-English / regional citations stay unresolvable and never block (a documented precision-over-recall tradeoff). Ships a persistent SQLite verification cache (~/.cache/ars/verification.db, 90-day TTL) with an /ars-cache-invalidate command, a standalone verification_gate API + verify_passport.py CLI, and a four-index extension (k=0..4) of the v3.9.0 contamination triangulation matrix (all advisory). academic-pipeline tracks the suite at v3.11.0; the other three skill versions are unchanged. Spec: docs/design/2026-05-21-v3.10-182-promote-citation-gate-spec.md (§0 amendment + C-V6).

v3.10.0 (2026-06-01) — Triangulation policy layer, Kong survey adoptions, eval harness, scoped-write guard

Minor release bundling: the opt-in contamination-triangulation terminal policy layer (#127 — default citation behavior byte-equivalent to v3.9.0); Kong et al. 2026 survey adoptions — the Rebuttal Commitment Ledger (#256/#266/#268/#269) and discipline-relative domain evidence profiles (#259); the v3.10 measurement infrastructure — a generalized eval gold set + ranking-lift CI gate (#184); the scoped-write guard MVP (#134) — a deterministic PreToolUse hook that fences the 23 single-phase agents to their own phase directory and denies them Bash (they use the Grep/Glob and structured editing tools instead); the /ars-mark-read plugin commands (#190) plus a broken-on-arrival fix (#195); a Simplified-Chinese README (#185); and CI hardening (#156/#155). academic-paper → v3.2.0 and academic-paper-reviewer → v1.10.0 for the Commitment-Ledger and domain-profile additions; academic-pipeline tracks the suite at v3.10.0. Default skill behavior is unchanged unless a strict policy mode is opted into; the one default-on change is the #134 guard, which constrains the fenced subagents, not user-facing outputs.

v3.9.4.2 (2026-05-19) — post-ship hotfix for PR #149 CI discipline gates (codex post-ship)

Codex post-ship review of PR #149 (7 CI discipline gates) surfaced 4 P2 findings; v3.9.4.2 hardens 3 of 4. F1: harness-retirement-monthly.yml adds GH_REPO so scheduled runs have repo context for gh issue create. F2: release-cooldown.yml filters PREV_TAG lookup to v* tags so non-release tags cannot bypass cooldown. F3: release-cooldown.yml also reads annotated tag subject + accepts hot-fix spelling (v3.9.2 was previously a false-negative hotfix). PR #157 follow-up: [skip-cooldown] override now read from both commit message AND annotated tag message (self-bootstrapping fix — this tag's cooldown bypass demonstrates F2+F3 work end-to-end). F4 (test-count-monotonic harden) reverted because it surfaced pre-existing scripts/ package issue, tracked as #154 (since fixed by PR #158) + re-attempt #155. Closes #152. Follow-ups: #155, #156.

v3.9.4.1 (2026-05-19) — post-ship hotfix for v3.9.4 temporal verification (#135 codex post-ship)

Codex post-ship review of v3.9.4 caught 4 real bugs that per-task subagent reviewers missed. Hotfix patches all 4: (1) audit() now wires citation_provenance through to P2 and P4 — when a ref slug has confidence: low or conflict, the verifier emits TEMPORAL-METADATA-MISSING instead of using timeline dates as ground truth (spec §3.4 first-party safety check was broken). (2) _date_to_interval parses all schema-valid date shapes including YYYY-MM (Crossref month precision) and YYYY-MM-DD..YYYY-MM-DD (interval); v3.9.4 silently ValueError'd on these and skipped the check. (3) P4 now binds direct date captures when ref markers are absent — sentences like "The 2026 policy enabled the 2020 rollout" actually trigger now. (4) citation_provenance.schema.json confidence:high allOf now requires presence (then.required) in addition to non-null, closing the absent-property bypass. 1561 passed (+12 new tests vs v3.9.4 baseline, 0 regression). ARCHITECTURE.md aligned to current state (was stale at v3.8.0).

v3.9.4 (2026-05-18) — #135 temporal verification layer (advisory)

Deterministic advisory verifier at the Phase 4 → 5 boundary covering 5 temporal failure modes (P1 retrospective arithmetic, P2 anachronistic citation, P3 comparator unmaterialized, P4 causal inversion, P5 deictic present). New Phase 2 sibling timeline_extraction_agent owns phase2_investigation/timeline.yaml + phase2_investigation/citation_provenance.yaml. Verifier script scripts/temporal_integrity_audit.py runs 5 passes deterministically. M3 Temporal Integrity Iron Rule added to report_compiler_agent + draft_writer_agent. M6-minimal: Crossref issued + pdftotext cover first-party verification. M7-minimal: date provenance + comparator materialization. M5-stub: user-declared version_family_id only. Zero modification to literature_corpus_entry, claim_audit_result, claim_intent_manifest. bibliography_agent unmodified (F2 invariant). 3 new sidecar schemas. Coverage estimate: 55-70% baseline / 65-75% with M7 minimal. 1549 passed (+44 new, 0 regression).

v3.9.3 (2026-05-18) — #128 housekeeping (shared client utilities + dedup resolvers)

Pure refactor + one latent-bug fix from the v3.9.0 /simplify review backlog. Extracts scripts/_text_similarity.py (3-way client dedup: normalize / similarity / threshold / retry constants) + scripts/_passport_yaml.py (2-way migration tool dedup: ruamel.yaml round-trip config) + private _resolve_by_doi_then_title helper (2-way resolver body dedup, §3.4 / §3.5 API surface preserved). Standardizes throttle measurement on time.monotonic across OpenAlex + Crossref (was time.time, NTP-unsafe), aligning with Semantic Scholar. Dual-path import infrastructure on all 5 module-level cross-imports (sibling-first, namespace-package fallback) preserves class identity for SemanticScholarUnavailable and bonus-fixes 2 latent-broken import scripts.X paths. 1505 passed (+23 new, 0 regression). #128 §4 (parallelize OA + CR per-entry) carried to #138.

v3.9.2 (2026-05-18) — #133 phase boundary hot-fix

#133 closure (hot-fix layer). Long-term architectural fix tracked as v3.10 active conductor in #134. Adds: routing clarification gate in CLAUDE.md (cross-phase materials → clarify with a-d options, not silent dispatch), 22 single-phase agents get prompt hard fence (## Phase Boundary (v3.9.2)), 16 multi-phase / phase-orthogonal / cross-phase-meta agents intentionally NOT fenced (honest framing — prose placebo creates false-enforcement illusion), advisory verifier scripts/check_pipeline_integrity.py detects #133 pattern post-hoc. Behavioral smoke tests with cross-model spot-check (100% Opus 4.7, ≥75% Sonnet + GPT-5.5).

v3.9.1 (2026-05-18) — #129 + #130 client hardening

v3.9.0 hot-fix. Wraps OpenAlex / Crossref response-read failures as *Unavailable (#129); guards check_claim_audit_consistency against non-string manifest_id (#130). No spec change.

v3.9.0 (2026-05-17) — #102 cross-index triangulation measurement

#102 closure. v3.7.3 shipped single-index (Semantic Scholar) contamination detection; v3.9.0 extends to three-index triangulation (S2 + OpenAlex + Crossref) as advisory evidence only. Two new optional booleans (openalex_unmatched, crossref_unmatched) on contamination_signals; manual-entry not-rule extended symmetrically. Finalizer adds a 4-tier advisory matrix (k=0/1/2/3 over present *_unmatched fields) with v3.7.3 legacy CONTAMINATED-UNMATCHED preserved for the k=1/k_max=1 S2-only case. Formatter pass-through allowlist extends 3 → 9 suffixes; refusal rules 1-10 unchanged per R-L3-2-E. The policy layer (strict modes, hard-block tier, venue_type / triangulation_policy) is deferred to v3.10 per spec §2.3. k=3 marker is CONTAMINATED-TRIANGULATION-UNMATCHED (describes observable, not inferred cause). 3 new firm rules: R-L3-2-C (k computed over present fields), R-L3-2-D (no API-inferred classification), R-L3-2-E (refusal list unchanged; pass-through allowlist extends).

Migration: v3.7.3 corpora — run python scripts/migrate_literature_corpus_to_v3_9_0.py PATH to backfill the two new fields. Pre-v3.7.3 corpora — run migrate_literature_corpus_to_v3_7_3.py FIRST, then v3.9.0 migration (daisy-chained per spec §3.7; the v3.9.0 tool only acts on entries that already carry contamination_signals.semantic_scholar_unmatched).

v3.8.2 (2026-05-17) — #118 uncited audit_tool_failure surface

#118 closure. The ARS_CLAIM_AUDIT=1 uncited constraint-judging path used to silently substitute {"judgment": "NOT_VIOLATED"} on JudgeInvocationError, suppressing HIGH-WARN constraint checks on transient judge outage. v3.8.2 routes those failures through a dedicated uncited_audit_failures[] aggregate at MED-WARN advisory tier, mirroring the cited path INV-14 row but using a dedicated schema because claim_audit_result.ref_slug is required and the uncited path has no ref to bind. The four option-1..4 trade-offs from the #118 issue body landed on option 2 (new aggregate) — option 4 (re-raise and abort) was rejected for the audit-coverage hit on flaky judge endpoints.

• New uncited_audit_failure.schema.json aggregate (spec §3.6). One entry per uncited sentence × manifest pair where the constraint judge raised JudgeInvocationError. Same fault-class enum as cited-path INV-14 (judge_timeout / judge_api_error / judge_parse_error / cache_corruption / retrieval_api_error / retrieval_timeout / retrieval_network_error). rule_version: D4-c-v1-uaf-v1.
• UAF-INV-1..UAF-INV-6 lint (spec §6 rule 4d). finding_id uniqueness, scoped_manifest_id cross-array integrity, (M, C) pair integrity when manifest_claim_id non-null, per-(sentence, manifest) dedup, rationale fault_class prefix, cross-aggregate exclusivity vs constraint_violations[].
• Finalizer §5 MED-WARN advisory row: annotation [CLAIM-AUDIT-TOOL-FAILURE-UNCITED — <fault-class>], gate passes (retry-next-pass remediation). Formatter REFUSE list unchanged — UAF is advisory.
• Pipeline integration (scripts/claim_audit_pipeline.py): swallow site at line 1211-1224 removed; JudgeInvocationError now emits a UAF row + continues to the next (sentence, manifest) pair. No fake NOT_VIOLATED reaches constraint_violations[].
• Tests: 18 new (15 schema/lint TSUAFUncitedAuditFailureInvariants + 3 pipeline integration TP23UncitedJudgeOutageEmitsUAF). Baseline 694 → 712 tests, 0 regression.
• Agent doc (academic-pipeline/agents/claim_ref_alignment_audit_agent.md): Output emission table grows seventh row; Error handling table grows from 3 surfaces to 4 surfaces with the uncited-path UAF row.

v3.8.0 (2026-05-16) — L3 Claim-Faithfulness Locator + Audit (paired milestone)

v3.7.3 + v3.8 close the L3 (claim-faithfulness) gap end-to-end. v3.7.3 ships the locator infrastructure — every citation carries a three-layer anchor so future audits can fetch the cited passage. v3.8 ships the audit pass that consumes those anchors, judges whether the cited source supports the claim, and gate-refuses HIGH-WARN violations at the formatter terminal hard gate. The release also bundles 5 audit-trail-shipped feature PRs accumulated since v3.7.0 (#104 / #105 / #108 / #111 / #115).

• #103 — claim_ref_alignment_audit_agent (v3.8 PR #121). Opt-in (ARS_CLAIM_AUDIT=1, default OFF) Stage 4→5 audit agent. Judges every sampled citation against retrieved excerpt; emits claim_audit_results[] + claim_intent_manifests[] + claim_drifts[] + uncited_assertions[] + constraint_violations[] aggregates. 8-row finalizer matrix routes HIGH-WARN classes (CLAIM-NOT-SUPPORTED / NEGATIVE-CONSTRAINT-VIOLATION / FABRICATED-REFERENCE / ANCHORLESS / CONSTRAINT-VIOLATION-UNCITED) through the formatter REFUSE rules 6-10. Calibration runner ships with 20-tuple gold set (T-C1 FNR<0.15 + FPR<0.10, T-C2 per-class, T-C3 shape integrity). 8 rounds of dual-track review (R1 codex + Gemini-3.1-pro-preview, R2-R8 codex-only after Gemini quota exhausted); trajectory R1 4P1+2P2 → R8 0P1+4P2 ship gate.
• v3.7.3 — Three-Layer Citation Emission + contamination signals (PR #98). synthesis_agent / draft_writer_agent / report_compiler_agent gain ## Three-Layer Citation Emission (v3.7.3) H2. Every <!--ref:slug--> carries <!--anchor:<kind>:<value>--> with <kind> ∈ {quote, page, section, paragraph, none} (quote anchors capped at 25 words, URL-encoded). pipeline_orchestrator_agent finalizer becomes 5-cell with precedence-zero NO-LOCATOR check. formatter_agent adds explicit hard-gate refusal for [UNVERIFIED CITATION — NO QUOTE OR PAGE LOCATOR]. literature_corpus_entry.schema.json adds optional contamination_signals: { preprint_post_llm_inflection, semantic_scholar_unmatched } object. bibliography_agent computes both signals at ingest. 11-round review trajectory (Codex×10 + Gemini cross-model×1) closed 22 findings. Spec: docs/design/2026-05-12-ars-v3.7.3-claim-faithfulness-and-contaminated-source-spec.md. External motivation: Zhao et al. arXiv:2605.07723 (2026-05).
• #108 — AI disclosure policy-anchor renderer (audit-trail-shipped 2026-05-14). Adds PRISMA-trAIce / ICMJE / Nature / IEEE policy-anchor disclosure paths alongside the existing venue-track renderer.
• #111 — slr_lineage emission on systematic-review → academic-paper handoff (2026-05-15). Schema 9 optional boolean slr_lineage field; producer pipeline_orchestrator_agent writes at every handoff transition; consumer disclosure mode dispatches --policy-anchor=prisma-trAIce per the §4.3 G2 invariant track gate.
• #104 — README motivation: Zhao et al. corpus-scale evidence anchor (2026-05-15). README + README.zh-TW.md motivation section frames the v3.7.x line against Zhao et al.'s 146,932 hallucinated-citation finding.
• #105 — v3.7.3 contamination_signals backfill migration tool (2026-05-15). scripts/migrate_literature_corpus_to_v3_7_3.py retro-computes both contamination signals across pre-v3.7.3 passports.
• #115 — Semantic Scholar client maturity (2026-05-15). scripts/semantic_scholar_client.py adds 1-req/s throttle (drops to 0.1s when S2_API_KEY detected), outage latch on URLError, and reset_outage_latch() for long-running cross-passport batches.

v3.7.0 (2026-05-05) — Claude Code Plugin Packaging

Plugin packaging upgrade: ARS now installs in one line on Claude Code CLI / VS Code / JetBrains via /plugin marketplace add Imbad0202/academic-research-skills + /plugin install academic-research-skills. The traditional git clone + symlink to ~/.claude/skills/ flow continues to work — both tracks are first-class.

• Plugin manifest + marketplace metadata (Phase 1, PR #68). .claude-plugin/plugin.json declares the suite (4 skills auto-discovered from skills/ directory via relative symlinks). .claude-plugin/marketplace.json registers the plugin so a single GitHub-hosted endpoint serves both the marketplace listing and the plugin source. README + README.zh-TW.md + docs/SETUP.md carry dual-track install instructions.
• 10 slash commands at commands/ars-*.md (Phase 2.1, PR #69) mapping MODE_REGISTRY.md entries to /ars-<mode> triggers. Model routing is pinned in each command's frontmatter — opus for full and revision-coach (architectural / review-interpretation depth), sonnet for the other 8. No Haiku per project policy.
• 3 plugin-shipped agents at agents/*_agent.md (Phase 2.1, PR #69) as relative symlinks to the v3.6.7-hardened downstream agents in deep-research/agents/: synthesis_agent, research_architect_agent, report_compiler_agent. Underscore filenames preserved to keep scripts/check_v3_6_7_pattern_protection.py hard-pinned paths and INV-3 manifest-confined Clause 1 invariant intact. Symlinks (not copies) preserve a single source of truth and prevent the Pattern C3 attack surface that v3.6.7 §6 inversion sweep + INV-1/2/3 lint closes. (Materialized to real byte-identical copies in #413 — relative symlinks break Windows checkouts without core.symlinks and zip-download installs; the single-source guarantee moved to the scripts/check_agents_mirror_sync.py byte-equality CI lint.)
• model: inherit added to those three source agent frontmatters. Inherit chosen over pinning sonnet so an opus session running ARS full pipeline keeps opus agents (instead of being capped). The user's ~/.claude/hooks/warn-agent-no-model.sh PreToolUse hook gates Haiku at the dispatching boundary, so inherit resolves through an already-Haiku-free model.
• SessionStart announce hook at hooks/hooks.json + scripts/announce-ars-loaded.sh (Phase 2.2, PR #70). When the plugin loads, the hook injects an additionalContext listing the 10 slash commands, the 3 plugin agents, and a token-budget pointer into the LLM's first turn. startup and clear source values get the full announce; resume and compact get a one-line ack to avoid burning context. Bash 3.2 compatible — runs on macOS stock /bin/bash with no brew install bash requirement.
• Phase 2.2 scope reduction: a SubagentStop → run_codex_audit.sh codex audit hook was scoped out for v3.7.0 due to a contract gap (the SubagentStop payload carries no stage/deliverable info, so the wrapper would have to half-infer required arguments) and an invoker-class boundary (run_codex_audit.sh lines 4–7 forbid same-session in-LLM invocation; PostToolUse fires inside the producing session). Real audit-hook integration deferred to a future release when ARS gains a stage/deliverable propagation contract. See docs/design/2026-04-30-ars-v3.7.0-plugin-packaging-roadmap.md Update note 2026-05-05 (Phase 2.2 scope reduction).
• docs/PERFORMANCE.md + .zh-TW.md gain a "v3.7.0 Plugin agents and model routing" subsection explaining the inherit semantics and current 3-agent scope boundary.
• Codex review chain across the three PRs: 8 inline iterative rounds + 3 fresh PR-level rounds, all converging to 0 P0/P1/P2 findings before merge. The Phase 2.2 fresh PR review caught one P2 (unquoted ${CLAUDE_PLUGIN_ROOT} breaking install paths with spaces) that the inline rounds missed — confirms the value of separating implementation review (inline) from contract review (fresh).
• What did NOT change: the four skill directories, all 25 modes, agent prompts, schema files, and lint contracts. Plugin packaging only adds new top-level surface (commands/, agents/, hooks/, .claude-plugin/, skills/ symlink dir, three plugin-agent model: inherit frontmatter additions). Existing 4.3k clone-install users see no breaking change.

v3.6.8 (2026-05-03) — Generator-Evaluator Contract Gate (v3.6.6 spec ship)

Naming note: this release ships the v3.6.6 generator-evaluator contract spec and implementation. The v3.6.6 work landed after v3.6.7 due to project sequencing; the design doc retains the v3.6.6 internal naming for the contract gate version, while the suite release is tagged v3.6.8 to keep the CHANGELOG monotonic.

• Schema 13.1 (shared/sprint_contract.schema.json) extends Schema 13 with two new mode enum values (writer_full + evaluator_full), two new optional top-level fields (pre_commitment_artifacts writer-only, disagreement_handling evaluator-only), and 12 allOf branches enforcing reviewer- / writer- / evaluator-conditional gates. Existing reviewer contracts validate byte-equivalent under Schema 13.1 (§3.6 zero-touch promise).
• Two new shipped contract templates under shared/contracts/writer/full.json (D1–D7, F1/F4/F2/F3/F0) and shared/contracts/evaluator/full.json (D1–D5, F1/F2/F3/F6/F4/F5/F0). Promoted from design-time artefacts on the spec branch to live shipped status atomically with the Schema 13.1 upgrade.
• Two-phase orchestration inside academic-paper full: Phase 4 splits into Phase 4a (writer paper-blind pre-commitment) + Phase 4b (writer paper-visible drafting + self-scoring); Phase 6 splits into Phase 6a (evaluator paper-blind pre-commitment) + Phase 6b (evaluator paper-visible scoring + decision). Phase-numbered <phase4a_output> / <phase6a_output> data delimiters mirror the v3.6.2 reviewer pattern. Lint count summary: writer 3+4 / evaluator 5+5 / reviewer 5+6 (reviewer remains zero-touch).
• academic-paper SKILL + agent files gain a verbatim ## v3.6.6 Generator-Evaluator Contract Protocol block (101 lines in SKILL.md plus 47 lines in draft_writer_agent.md + 57 lines in peer_reviewer_agent.md). SKILL.md also adds a new ## Known limitations section carrying graceful-degradation + cross-session resume forward notes for v3.6.7+.
• Validator extensions: scripts/check_sprint_contract.py SC-* mode-gating audit (SC-5 + SC-11 reviewer-only; SC-9 extended across all three mode families). 17 new tests bring the validator unit-test count from 54 to 71 (positive + 5 schema-branch negative + 2 §3.6 reviewer regression + 6 mode-gating tests).
• Manifest CI lint: scripts/check_v3_6_6_ab_manifest.py enforces §6.2 manifest schema + §6.5 git-tracked invariants on tests/fixtures/v3.6.6-ab/manifest.yaml. .github/workflows/spec-consistency.yml extends the sprint contract validation loop to iterate writer + evaluator template directories alongside the existing reviewer loop, plus runs the new manifest CI lint.
• A/B evidence fixture stub at tests/fixtures/v3.6.6-ab/ (30 files): manifest + README + 6 paper-A inputs/baseline + 1 paper-C inputs/baseline + Stage 3 reviewer excerpt + 6 codex-judge baseline placeholders. Real fixture data populates in follow-up commits before the implementation work fully completes.

v3.6.7 (2026-04-30) — Downstream-Agent Pattern Protection (Step 1+2)

• Three downstream agents hardened against 13 of 17 documented hallucination/drift patterns: synthesis_agent (A1–A5 narrative-side), the survey-designer mode of research_architect_agent (B1–B5 instrument-side), and the abstract-only mode of report_compiler_agent (C1–C3 publication-side). Each agent prompt now carries a PATTERN PROTECTION (v3.6.7) block.
• Four reference files in shared/references/: irb_terminology_glossary.md, psychometric_terminology_glossary.md, protected_hedging_phrases.md, word_count_conventions.md. The reference files carry operational contracts that the agent prompts cite by path.
• Cross-model audit prompt template at shared/templates/codex_audit_multifile_template.md with seven audit dimensions and a mandatory three-part Section 4(f) check for report_compiler_agent bundles. Failure of any sub-check is a P1 finding.
• Static lint + 29-test mutation suite: scripts/check_v3_6_7_pattern_protection.py enforces protection-clause presence and obligation-phrase shape; scripts/test_check_v3_6_7_pattern_protection.py preserves codex review evidence so future checker regressions surface in CI. Both are wired into .github/workflows/spec-consistency.yml.
• Codex review history: seven rounds of gpt-5.5 + xhigh cross-model review reached SHIP-OK with zero P1+P2 findings. Step 6 (orchestrator runtime hooks) and Step 8 (synthetic eval case) ship in a follow-up PR.

v3.6.5 (2026-04-27) — Material Passport literature_corpus[] Consumer Integration

• Two Phase 1 literature consumers wired: deep-research/agents/bibliography_agent.md and academic-paper/agents/literature_strategist_agent.md. Both follow the same five-step corpus-first, search-fills-gap flow when the passport carries a non-empty literature_corpus[] and the same four Iron Rules (Same criteria / No silent skip / No corpus mutation / Graceful fallback on parse failure).
• PRE-SCREENED reproducibility block in Search Strategy reports: enumerates included / excluded / skipped corpus entries, with F3 zero-hit note and F4a–F4f provenance reporting that compose around partial declaration of obtained_via / obtained_at. final_included = pre_screened_included[] ∪ external_included[] stays neutral — no provenance tags on bibliography entries or literature matrix rows.
• Consumer protocol reference at academic-pipeline/references/literature_corpus_consumers.md with the canonical PRE-SCREENED template, BAD/GOOD examples, four Iron Rules, and per-consumer reading instructions.
• CI lint scripts/check_corpus_consumer_protocol.py enforcing nine protocol invariants with manifest-driven consumer list (scripts/corpus_consumer_manifest.json).
• Schema 9 caveat retired: shared/handoff_schemas.md retired the v3.6.4 "Consumer-side integration deferred to v3.6.5+" caveat; replaced with backpointer to the consumer protocol.
• Presence-based, no schema change, no new env flag. Parse failures fall back to external-DB-only flow with a [CORPUS PARSE FAILURE] surface. citation_compliance_agent corpus integration deferred (target version TBD post-v3.8).
• No breaking changes. Existing user adapters work without modification.

v3.6.4 (2026-04-25) — Material Passport literature_corpus[] Input Port

• literature_corpus[] field added to Schema 9 as an optional input port for user-owned literature. Each entry conforms to shared/contracts/passport/literature_corpus_entry.schema.json (CSL-JSON authors, year, title, source_pointer + private optional abstract / user_notes).
• Language-neutral adapter contract at academic-pipeline/references/adapters/overview.md: any program (any language) reading a user corpus source can produce conformant passport.yaml + rejection_log.yaml. Fail-soft entry-level errors, fail-loud adapter-level errors, deterministic ordering.
• Three reference Python adapters under scripts/adapters/: folder_scan.py (filesystem of PDFs), zotero.py (Better BibTeX JSON export), obsidian.py (vault frontmatter). Starting points only; users are expected to write their own adapters for non-reference sources.
• Rejection log contract at shared/contracts/passport/rejection_log.schema.json with closed enum of categorical reason values; always emitted (empty when no rejections).
• CI gates: scripts/check_literature_corpus_schema.py validates schemas + adapter examples; scripts/sync_adapter_docs.py --check prevents schema→docs drift; new pytest.yml workflow runs scripts/adapters/tests/ on path-filtered triggers.
• Input-port-only at v3.6.4: v3.6.4 shipped the schema and adapter contract without consumer integration. bibliography_agent and literature_strategist_agent were wired in v3.6.5.
• No breaking changes.

v3.6.3 (2026-04-23) — Opt-in Passport Reset Boundary

• Opt-in passport reset boundary (ARS_PASSPORT_RESET=1). Promotes every FULL checkpoint to a context-reset boundary. New resume_from_passport=<hash> mode lets users resume in a fresh Claude Code session from the Material Passport ledger alone. systematic-review mode with the flag ON makes reset mandatory at every FULL checkpoint; other modes treat reset as the flag-gated default. Flag OFF preserves pre-v3.6.3 behavior byte-for-byte.
• Schema 9 gains an append-only reset_boundary[] ledger with two entry kinds (kind: boundary + kind: resume). Hash uses JSON Canonical Form + SHA-256 with canonical placeholder for self-reference safety. Optional pending_decision handles MANDATORY branch choices.
• New scripts/check_passport_reset_contract.py CI lint: every mention of the flag must co-locate a pointer to the authoritative protocol doc.
• Protocol doc: academic-pipeline/references/passport_as_reset_boundary.md.
• docs/PERFORMANCE.md updated with long-running-session guidance.
• No breaking changes. Flag default is OFF.

v3.6.2 (2026-04-23) — Reviewer Sprint Contract Hard Gate

v3.6.2 introduces Schema 13 sprint contracts and a hard-gate orchestration that forces reviewers to pre-commit their scoring plan before reading the paper. Reviewer-only first test case; writer/evaluator deferred to v3.6.4. See CHANGELOG.

• Schema 13 sprint contract with panel_size, acceptance_dimensions, failure_conditions (with severity precedence + panel-relative cross_reviewer_quantifier), measurement_procedure, optional override_ladder, bounded agent_amendments. Validator: scripts/check_sprint_contract.py.
• Two-call hard gate. Reviewers run paper-content-blind Phase 1 + paper-visible Phase 2; Phase 1 output is wrapped in <phase1_output>...</phase1_output> data delimiter to narrow the self-injection surface.
• Synthesizer three-step mechanical protocol. Build cross-reviewer matrix → evaluate each failure_condition with panel-relative quantifier + recognised expression vocabulary → resolve precedence by severity. Forbidden-ops list explicit in editorial_synthesizer_agent.
• Two reviewer templates ship (shared/contracts/reviewer/full.json panel 5; shared/contracts/reviewer/methodology_focus.json panel 2). reviewer_re_review, reviewer_calibration, reviewer_guided are reserved in the schema enum but ship without contract templates in v3.6.2; they retain pre-v3.6.2 behaviour. reviewer_quick is excluded from the enum entirely.
• academic-paper-reviewer SKILL version: 1.8.1 → 1.9.0. academic-pipeline SKILL version: 3.5.1 → 3.6.2 (suite-version invariant). Suite version bumped to 3.6.2.
• See spec docs/design/2026-04-23-ars-v3.6.2-sprint-contract-design.md and protocol academic-paper-reviewer/references/sprint_contract_protocol.md.

v3.5.1 (2026-04-22) — Opt-in Socratic Reading-Check Probe

v3.5.1 adds an opt-in honesty probe to the Socratic Mentor (ARS_SOCRATIC_READING_PROBE=1). Default off. See CHANGELOG.

• Opt-in reading-check probe: when ARS_SOCRATIC_READING_PROBE=1 is set, the Socratic Mentor fires a one-time honesty probe during goal-oriented sessions where the user has cited a specific paper. Decline is logged without penalty. Outcome flows into the Research Plan Summary and Stage 6 AI Self-Reflection Report. No new agent, no schema change.
• deep-research SKILL version: 2.9.0 → 2.9.1. academic-pipeline SKILL version: 3.5.0 → 3.5.1. Suite version bumped to 3.5.1.

v3.5.0 (2026-04-21) — Collaboration Depth Observer

• New agent: collaboration_depth_agent in academic-pipeline (Agent Team grows from 3 to 4). Invoked at every FULL/SLIM checkpoint and at pipeline completion; scores user-AI collaboration against a 4-dimension rubric. Advisory only — never blocks progression. MANDATORY checkpoints (Stages 2.5 / 4.5 integrity gates) do NOT invoke the observer.
• New rubric: shared/collaboration_depth_rubric.md v1.0. Dimensions: Delegation Intensity, Cognitive Vigilance, Cognitive Reallocation, Zone Classification (Zone 1 / Zone 2 / Zone 3). Based on Wang, S., & Zhang, H. (2026). "Pedagogical partnerships with generative AI in higher education: how dual cognitive pathways paradoxically enable transformative learning." International Journal of Educational Technology in Higher Education, 23:11. DOI 10.1186/s41239-026-00585-x.
• Cross-model divergence flagged, not averaged: when ARS_CROSS_MODEL is set the observer runs on both models; dimension disagreement > 2 points is reported rather than silently smoothed. ARS_CROSS_MODEL_SAMPLE_INTERVAL escape hatch for cost trade-off.
• Short-stage guard: stages with fewer than 5 user turns inject a static insufficient_evidence block instead of dispatching the full-model observer.
• Anti-sycophancy discipline: scores ≥ 7 require specific dialogue-turn citations; Zone 3 triggers re-audit; no motivational framing.
• academic-pipeline SKILL version: 3.3.0 → 3.4.0. Suite version bumped to 3.5.0. New lint scripts/check_collaboration_depth_rubric.py + 10 tests.

v3.4.0 (2026-04-20) — Compliance Agent + Schema 12

• Compliance Agent (shared): single mode-aware agent running PRISMA-trAIce 17 items (SR mode only) + RAISE 4 principles + 8-role matrix. Hooks existing Stage 2.5 / 4.5 Integrity Gates; tier-based block (Mandatory → block, HR → warn, R/O → info). Non-SR entries run principles-only, warn-only.
• Schema 12 compliance_report appended to Material Passport via compliance_history[] (append-only).
• 3-round user-override ladder auto-injects disclosure_addendum into manuscript. No detection evasion possible.
• Calibration with transparent reporting, no hard FNR/FPR gate — self-consistent with task_type: open-ended.
• Upstream freshness CI warns on PRISMA-trAIce drift (non-blocking).
• Long-running session docs: Material Passport as cross-session resume mechanism.

v3.3.6 (2026-04-15) — README Streamlining + ARCHITECTURE doc

• Added docs/ARCHITECTURE.md as the single source of truth for pipeline structure (flow, matrix, data-access, dependency graph, quality gates, modes). Merged into main via PR #18.
• Added docs/SETUP.md (prerequisites, API keys, Pandoc/tectonic, cross-model verification, installation methods) and docs/PERFORMANCE.md (token budgets, recommended Claude Code settings). README links to both instead of inlining them.
• Streamlined README: removed the ASCII pipeline diagram and 16-point key-feature list (superseded by ARCHITECTURE.md); Skill Details section now anchors version numbers and points readers to ARCHITECTURE.md §3 for per-agent rosters.
• Note: no functional change to any skill. Pure documentation reorganization. Suite version bumped to 3.3.6.

v3.3.5 (2026-04-15)

• Added benchmark_report.schema.json + repro_lock optional block on Material Passport. Both ship with pattern docs, lints, and examples. First formal Python dev dep manifest (requirements-dev.txt).

v3.3.4 (2026-04-15) — README Changelog Sync Patch

• Synced the embedded changelog sections in README.md and README.zh-TW.md so they include the missing v3.3.3 and v3.3.2 release summaries.
• Extended scripts/check_spec_consistency.py so future README changelog drift fails CI.

v3.3.3 (2026-04-15) — Release Prep + Lint Hardening

• Hardened SKILL frontmatter linting: missing closing --- fences now fail cleanly instead of being parsed as valid YAML.
• Frontmatter that parses as valid YAML but not as a mapping now reports a readable error instead of crashing.
• Fixed the broken showcase link for the post-publication audit report in both READMEs.
• Added README relative-link validation to the spec consistency check so dead links fail CI.
• Aligned the DOCX output contract across the docs: direct .docx generation is Pandoc-dependent, with Markdown + conversion instructions as fallback.
• Prepared the v3.3.3 release: suite version bump, academic-paper -> v3.0.2, academic-pipeline -> v3.2.2.

v3.3.2 (2026-04-15) — Data Access Levels + Task Type Metadata

• Added metadata.data_access_level to all top-level SKILL.md files with enforced vocabulary: raw, redacted, verified_only.
• Added metadata.task_type to all top-level SKILL.md files with enforced vocabulary: open-ended, outcome-gradable.
• Added lint scripts and unit tests for both metadata fields, wired into the GitHub Actions spec consistency workflow.
• Added shared/ground_truth_isolation_pattern.md and linked the new vocabulary from shared/handoff_schemas.md.

v3.3.1 (2026-04-14) — Spec Consistency Patch

• Synced README, .claude/CLAUDE.md, MODE_REGISTRY.md, and SKILL.md files to the current mode counts and published skill versions.
• Corrected cross-model wording: integrity sample checks and independent DA critique are implemented today; sixth-reviewer peer review remains planned.
• Clarified adaptive checkpoint semantics so SLIM checkpoints still wait for explicit user confirmation.
• Reaffirmed that Stage 2.5 and Stage 4.5 integrity gates cannot be skipped.
• Added a lightweight spec consistency check and GitHub Actions workflow to catch future drift.

v3.3 (2026-04-09) — PaperOrchestra-Inspired Enhancements

Integrates techniques from PaperOrchestra (Song, Song, Pfister & Yoon, 2026, Google).

• Semantic Scholar API Verification — Tier 0 programmatic reference existence check via S2 API. Levenshtein >= 0.70 title matching, DOI mismatch detection, bibliography deduplication via S2 IDs. Graceful degradation if API unavailable.
• Anti-Leakage Protocol — Knowledge Isolation Directive prioritizes session materials over LLM parametric memory. Flags [MATERIAL GAP] for missing content instead of filling from memory. Reduces Mode 5/6 failure risk.
• VLM Figure Verification (optional) — Closed-loop verification of rendered figures using vision-capable LLM. 10-point checklist, max 2 refinement iterations.
• Score Trajectory Protocol — Per-dimension rubric score delta tracking across revision rounds (7 dimensions). Detects regressions (delta < -3) and triggers mandatory checkpoint.
• Stage 2 Parallelization — Visualization and argument building can run in parallel after outline completion.
• New versions: deep-research v2.8, academic-paper v3.0, academic-pipeline v3.2

v3.2 (2026-04-09) — Lu 2026 Nature Integration

Integrates insights from Lu et al. (2026, Nature 651:914-919) — the first end-to-end autonomous AI research system to pass blind peer review.

• 7-mode AI Research Failure Mode Checklist — blocks pipeline at Stage 2.5/4.5 on suspected implementation bugs, hallucinated results, shortcut reliance, bug-as-insight, methodology fabrication, frame-lock. Extends existing 5-type citation hallucination taxonomy.
• Reviewer Calibration Mode (academic-paper-reviewer v1.8) — opt-in FNR/FPR/balanced-accuracy measurement against user-supplied gold set. 5× ensembling, cross-model default-on, session-scoped confidence disclosure.
• Disclosure Mode (academic-paper v2.9) — the default venue path returns REQUIRED, ACTION_ONLY, NOT_REQUIRED, or UNKNOWN applicability plus an explicit typed halt status when needed; policy-anchor invocations use their separate anchor-specific renderer. v1 covers ICLR, NeurIPS, Nature, Science, ACL, EMNLP. (Since expanded: the v2 database (#596) adds 9 medical-publishing policy targets — ICMJE, NEJM, The Lancet, JAMA, BMJ, PLOS, Frontiers, plus the database's first two Chinese-language policy targets: the publisher-wide 中华护理杂志社 entry and the journal 国际眼科杂志.)
• Early-Stopping Criterion (academic-pipeline v3.1) — convergence check + budget transparency at pipeline start.
• Fidelity-Originality Mode Spectrum — classifies all modes across 3 skills per Lu 2026 Fig 1c.
• New versions: academic-paper v2.9, academic-paper-reviewer v1.8, academic-pipeline v3.1

v3.1.1 (2026-04-09) — IS Senior Scholars' Basket of 11

External contributions: @mchesbro1 originally proposed and drafted the IS Basket of 8 journals (Issue #5); @cloudenochcsis extended it to the full Senior Scholars' Basket of 11 (Issue #7, PR #8). Updated academic-paper-reviewer/references/top_journals_by_field.md Section 7, adding Decision Support Systems, Information & Management, and Information and Organization. Source: AIS Senior Scholars' List of Premier Journals.

v3.1 (2026-04-06) — Anti-Context-Rot + Cognitive Frameworks + Lean Size

Inspired by patterns from aspi6246/Claude-Code-Skills-for-Academics.

Wave 1: Anti-Context-Rot Anchors

• 29 explicit Anti-Patterns across all 4 skills (7-8 per skill, tabular format with "Why It Fails" + "Correct Behavior")
• 22 IRON RULE markers on critical rules that must not be violated even in long conversations
• Read-only constraint on academic-paper-reviewer (reviewers cannot modify the manuscript)

Wave 2: Traceability + Cognitive Frameworks + Reinforcement

• R&R Traceability Matrix (Schema 11): adds "Author's Claim" and "Verified?" columns to re-review output, enabling independent verification of revision claims
• 3 cognitive framework reference files teaching agents "how to think" not just "what to do":
  • argumentation_reasoning_framework.md — Toulmin model, Bradford Hill causal reasoning, inference to best explanation, epistemic status classification
  • review_quality_thinking.md — three lenses (internal validity, external validity, contribution), common reviewer traps, calibration questions
  • writing_judgment_framework.md — clarity test, reader's journey, discipline-specific voice, revision decision matrix
• Mid-conversation reinforcement protocol: stage-specific IRON RULE + Anti-Pattern reminders at every pipeline transition
• Self-check questions at every FULL checkpoint (citation integrity, sycophantic concession, quality trajectory, scope discipline, completeness)

Wave 3: Lean Skill Size

• SKILL.md total size reduced from 142KB to 85KB (−40%) by extracting detailed protocols to references/ files
• ~15 new reference files created (re-review protocol, guided mode, systematic review, process summary, external review, etc.)
• All IRON RULE markers preserved in SKILL.md; detailed content loaded on demand
• New versions: deep-research v2.7, academic-paper v2.8, academic-paper-reviewer v1.7, academic-pipeline v3.0

v3.0 (2026-04-03) — Anti-Sycophancy + Intent Detection + Dialogue Health

• Devil's Advocate Concession Threshold (deep-research + academic-paper-reviewer): DA must score rebuttals 1-5 before responding. Concession only at ≥4. No consecutive concessions. Concession rate tracking. Frame-lock detection after each checkpoint.
• Attack Intensity Preservation (academic-paper-reviewer): DA does not soften under pushback. Rebuttal assessment protocol with explicit deflection detection. Anti-sycophancy rules prevent persistent pushback from being treated as valid evidence.
• Intent Detection Layer (deep-research socratic): Classifies user intent as exploratory vs. goal-oriented. Exploratory mode disables auto-convergence, raises max rounds, prohibits premature closure. Re-assesses every 3 turns.
• Dialogue Health Indicator (deep-research socratic): Silent self-check every 5 turns for persistent agreement, conflict avoidance, premature convergence. Auto-injects challenges when agreement pattern detected.
• Cross-Model Verification Protocol (shared, optional): Use GPT-5.4 Pro or Gemini 3.1 Pro for integrity verification sample cross-checks and independent DA critique. Sixth-reviewer peer review remains planned, not yet implemented. Activated by setting ARS_CROSS_MODEL env var — without it, everything works as before. See shared/cross_model_verification.md for full setup guide, API patterns, and cost estimates.
• AI Self-Reflection Report (academic-pipeline Stage 6): Post-pipeline self-assessment of AI behavioral patterns — DA concession rate, checkpoint skip rate, health alerts, sycophancy risk rating (LOW/MEDIUM/HIGH), frame-lock incidents, convergence pattern analysis. Includes irony caveat: "this self-reflection is itself produced by the same AI that may have been sycophantic."
• Origin: Discovered through a 4-round dialectic experiment where the DA conceded too quickly, the Socratic Mentor tried to converge prematurely, and the entire debate stayed locked in a frame the human set.
• Versions: deep-research v2.5, academic-paper-reviewer v1.5, academic-pipeline v2.8

v2.9.1 (2026-04-03) — Skill Metadata

• Added status: active and related_skills cross-references to all 4 SKILL.md frontmatters.
• Enables skill discovery tools and cross-skill navigation across deep-research ↔ academic-paper ↔ academic-paper-reviewer ↔ academic-pipeline.

v2.9 (2026-03-27) — Style Calibration + Writing Quality Check

• Style Calibration (academic-paper intake Step 10, optional): Provide 3+ past papers and the pipeline learns your writing voice — sentence rhythm, vocabulary preferences, citation integration style. Applied as a soft guide during drafting; discipline conventions always take priority. Priority system: discipline norms (hard) > journal conventions (strong) > personal style (soft). See shared/style_calibration_protocol.md
• Writing Quality Check (academic-paper/references/writing_quality_check.md): Writing quality checklist applied during draft self-review. 5 categories: AI high-frequency term warnings (25 terms), punctuation pattern control (em dash ≤3), throat-clearing opener detection, structural pattern warnings (Rule of Three, uniform paragraphs, synonym cycling), and burstiness checks (sentence length variation). These are good writing rules — not detection evasion
• Style Profile carried through academic-pipeline Material Passport (Schema 10 in shared/handoff_schemas.md)
• deep-research report compiler also consumes both features optionally
• Versions: academic-paper v2.5, deep-research v2.4, academic-pipeline v2.7

v2.8 (2026-03-22) — SCR Loop Phase 1: State-Challenge-Reflect

• Socratic Mentor Agent (deep-research + academic-paper): SCR (State-Challenge-Reflect) protocol integration
  • Commitment Gates: Collect user predictions before presenting evidence at each layer/chapter transition
  • Certainty-Triggered Contradiction: Detect high-confidence language ("obviously", "clearly") and introduce counterpoints
  • Adaptive Intensity: Track commitment accuracy, dynamically adjust challenge frequency
  • Self-Calibration Signal (S5): New convergence signal tracking user's self-calibration growth across dialogue
  • SCR Switch: Users can say "skip the predictions" to disable or "turn predictions back on" to re-enable mid-dialogue; Socratic questioning continues normally
• deep-research/references/socratic_questioning_framework.md: SCR Overlay Protocol mapping SCR phases to Socratic functions
• Added CHANGELOG.md

v2.7 (2026-03-09) — Integrity Verification v2.0: Anti-Hallucination Overhaul

• integrity_verification_agent v2.0: Anti-Hallucination Mandate (no AI memory verification), eliminated gray-zone classifications (VERIFIED/NOT_FOUND/MISMATCH only), mandatory WebSearch audit trail for every reference, Stage 4.5 fresh independent verification, Gray-Zone Prevention Rule
• Known Hallucination Patterns: 5-type taxonomy (TF/PAC/IH/PH/SH) from GPTZero × NeurIPS 2025 study, 5 compound deception patterns, real-world case study, literature statistics
• Post-publication audit: Full WebSearch verification of all 68 references found 21 issues (31% error rate) that passed 3 rounds of integrity checks — proving the necessity of external verification
• Paper corrections: Removed 4 fabricated references, fixed 6 author errors, corrected 7 metadata errors, fixed 2 format issues

v2.6.2 (2026-03-09) — Intent-Based Mode Activation

• deep-research: Socratic mode now uses intent-based activation instead of keyword matching. Works in any language — detects meaning (e.g., "user wants guided thinking") rather than matching specific strings.
• academic-paper: Plan mode now uses intent-based activation. Detects intent signals like "user is uncertain how to start" or "user wants step-by-step guidance" in any language.
• Both modes now have a default rule: when intent is ambiguous, prefer socratic/plan over full — safer to guide first.
• Two-layer architecture: Layer 1 (skill activation) uses bilingual keywords for matching confidence; Layer 2 (mode routing) uses language-agnostic intent signals.

v2.6.1 (2026-03-09) — Bilingual Trigger Keywords

• deep-research: Added Traditional Chinese trigger keywords for general activation and Socratic mode.
• academic-paper: Added Traditional Chinese trigger keywords and Plan Mode trigger section.
• Both mode selection guides now include bilingual examples and Chinese-specific misselection scenarios.

v2.6 / v2.4 / v1.4 (2026-03-08) — 15+ Improvements

• deep-research v2.3: New systematic-review / PRISMA mode (7th); 3 new agents (risk_of_bias, meta_analysis, monitoring); PRISMA protocol/report templates; Socratic convergence criteria (4 signals + auto-end); Quick Mode Selection Guide
• academic-paper v2.4: 2 new agents (visualization, revision_coach); revision tracking template with 4 status types; citation format conversion (APA↔Chicago↔MLA↔IEEE↔Vancouver); statistical visualization standards; Socratic convergence criteria; revision recovery example; LaTeX output hardening — mandatory apa7 document class, text justification fix (ragged2e + etoolbox), table column width formula, bilingual abstract centering, standardized font stack (Times New Roman + Source Han Serif TC VF + Courier New), PDF via tectonic only
• academic-paper-reviewer v1.4: Quality rubrics with 0-100 scoring and behavioral indicators; decision mapping (≥80 Accept, 65-79 Minor, 50-64 Major, <50 Reject); Quick Mode Selection Guide
• academic-pipeline v2.6: Adaptive checkpoint system (FULL/SLIM/MANDATORY); Phase E Claim Verification in integrity checks; Material Passport for mid-entry provenance; cross-skill mode advisor (14 scenarios); team collaboration protocol; enhanced handoff schemas (9 schemas); integrity failure recovery example

v2.4 / v1.3 (2026-03-08)

• academic-pipeline v2.4: New Stage 6 PROCESS SUMMARY — auto-generates structured paper creation process record (MD → LaTeX → PDF, bilingual); mandatory final chapter: Collaboration Quality Evaluation with 6 dimensions scored 1–100 (Direction Setting, Intellectual Contribution, Quality Gatekeeping, Iteration Discipline, Delegation Efficiency, Meta-Learning), honest feedback, and improvement recommendations; pipeline expanded from 9 to 10 stages

v2.3 / v1.3 (2026-03-08)

• academic-pipeline v2.3: Stage 5 FINALIZE now prompts for formatting style (APA 7.0 / Chicago / IEEE); PDF must compile from LaTeX via tectonic (no HTML-to-PDF); APA 7.0 uses apa7 document class (man mode) with XeCJK for bilingual CJK support; font stack: Times New Roman + Source Han Serif TC VF + Courier New

v2.2 / v1.3 (2025-03-05)

• Cross-Agent Quality Alignment: unified definitions (peer-reviewed, currency rule, CRITICAL severity, source tier) across all agents
• deep-research v2.2: synthesis anti-patterns, Socratic auto-end conditions, DOI+WebSearch verification, enhanced ethics integrity check, mode transition matrix
• academic-paper v2.2: 4-level argument scoring, plagiarism screening, 2 new failure paths (F11 Desk-Reject Recovery, F12 Conference-to-Journal), Plan→Full mode conversion
• academic-paper-reviewer v1.3: DA vs R3 role boundaries, CRITICAL finding criteria, consensus classification (4/3/SPLIT/DA-CRITICAL), confidence score weighting, Asian & Regional Journals reference
• academic-pipeline v2.2: checkpoint confirmation semantics, mode switching matrix, failure fallback matrix, state ownership protocol, material version control

v2.0.1 (2026-03)

• Simplify 4 SKILL.md (-371 lines, -16.5%): remove cross-skill duplication, inline templates → file references, redundant routing tables, duplicate mode selection sections
• Fix revision loop cap contradiction between academic-paper and academic-pipeline

v2.0 (2026-02)

• academic-pipeline v2.0: 5→9 stages, mandatory integrity verification, two-stage review, Socratic revision coaching, reproducibility guarantees
• academic-paper-reviewer v1.1: +Devil's Advocate Reviewer (7th agent), +re-review mode (verification), +post-review Socratic coaching
• New agent: integrity_verification_agent — 100% reference/data verification with audit trail
• New agent: devils_advocate_reviewer_agent — 8-dimension thesis challenger
• Output order: MD → DOCX via Pandoc when available (else instructions) → ask LaTeX → confirm → PDF

v1.0 (2026-02)

• Initial release
• deep-research v2.0 (10 agents, 6 modes including socratic)
• academic-paper v2.0 (10 agents, 8 modes including plan)
• academic-paper-reviewer v1.0 (6 agents, 4 modes including guided)
• academic-pipeline v1.0 (orchestrator)