psyclaw 0.27.22 → 0.28.0

package/vendor/ars/deep-research/references/ethics_checklist.md

@@ -0,0 +1,282 @@
+# Research Ethics Checklist — AI-Assisted Research
+
+## Purpose
+Comprehensive ethics checklist for AI-assisted academic research. Used by the ethics_review_agent.
+
+## 1. AI Disclosure
+
+### Mandatory Disclosure Elements
+- [ ] AI tools used are named (e.g., "Claude," "GPT-4," "Gemini")
+- [ ] Scope of AI involvement specified:
+  - [ ] Literature search assistance
+  - [ ] Source screening
+  - [ ] Evidence synthesis
+  - [ ] Draft writing
+  - [ ] Editing/revision
+  - [ ] Data analysis
+  - [ ] Translation
+- [ ] Human oversight described (who reviewed what, at which stages)
+- [ ] AI limitations acknowledged (potential hallucination, knowledge cutoff, etc.)
+- [ ] AI version/date noted (for reproducibility)
+
+### Disclosure Statement Template
+```
+AI Disclosure: This research was conducted with assistance from [AI Tool Name]
+(version/date). AI was used for [specific tasks]. All findings were verified
+against cited sources by [human role]. The research team maintains full
+responsibility for the accuracy and interpretation of all content.
+```
+
+### Where to Place Disclosure
+- In the methodology section (detailed)
+- In the abstract or author note (brief)
+- In footnotes for specific AI-generated analyses
+
+## 2. Attribution Integrity
+
+### Citation Ethics
+- [ ] Every factual claim has at least one supporting citation
+- [ ] No fabricated or hallucinated references
+  - Verification: Spot-check minimum 20% of references for existence
+  - Cross-check DOIs, publication years, author names
+- [ ] Paraphrasing is genuine (not just rearranging words)
+- [ ] Direct quotes are exact and attributed
+- [ ] Ideas are attributed to original authors, not intermediary sources
+- [ ] Self-citation is proportionate (not excessive or exclusionary)
+
+### AI-Specific Attribution Risks
+| Risk | Description | Mitigation |
+|------|-------------|-----------|
+| Hallucinated references | AI generates plausible but non-existent citations | Verify every reference against database |
+| Merged citations | AI combines details from multiple sources | Cross-check each citation element |
+| Incorrect authors | AI assigns wrong authors to works | Verify author names against actual publications |
+| Wrong year | AI uses incorrect publication year | Cross-check against database records |
+| Ghost citations | References listed but never cited in text | Audit reference list against in-text citations |
+
+## 3. Dual-Use Assessment
+
+**Screen on concrete specifics, never on subject matter.** A sensitive, political, or institution-critical *topic* is not a dual-use trigger. Public-interest research — documenting institutional abuses, exposing surveillance practices, holding power to account — is expected to address harmful subject matter and must not be flagged for the topic alone. A finding triggers dual-use only when the work itself supplies **specific operational detail** that materially lowers the barrier to harm. Studying surveillance is fine; shipping a step-by-step deployment recipe is the trigger.
+
+### Screening Questions
+Each asks whether the **content of this work**, not its topic, supplies the specific:
+1. Does the work provide concrete operational detail that would let a reader **carry out** harm against individuals or communities (beyond describing that the harm exists)?
+2. Does it disclose a **specific, currently-exploitable** vulnerability together with enough detail to exploit it (vs. naming that vulnerabilities exist)?
+3. Does it provide a **usable method** to build surveillance or control mechanisms (vs. analyzing or critiquing them)?
+4. Does it contain **weaponizable specifics** — a recipe, design, or procedure — rather than discussion of weaponization risk?
+5. Does it supply a **concrete means** to discriminate against a group (vs. documenting that discrimination occurs)?
+
+If the answer rests on the topic being sensitive rather than on specific enabling detail in the text, the level is None.
+
+### Risk Levels and Responses
+
+This assessment is **advisory** — it routes to a Responsible Use Statement and never to a hard block. (Hard blocks are reserved for integrity violations; see `agents/ethics_review_agent.md` Blocking Conditions.) Escalation rests on concrete enabling specifics, not subject matter.
+
+| Level | Action Required |
+|-------|----------------|
+| None | No additional action |
+| Low | Brief note in limitations |
+| Moderate | Responsible Use statement in report |
+| High | Prominent warning + limited distribution recommendation |
+| Critical | Responsible Use statement + recommend institutional ethics review before publication (advisory — does not auto-block; a human, not the agent, adjudicates a Critical flag) |
+
+### Responsible Use Statement Template
+```
+Responsible Use: This research is intended for [stated purpose]. The authors
+acknowledge that findings related to [sensitive area] could potentially be
+applied in ways not intended by this research. Users of this research are
+urged to consider the ethical implications of their applications and to
+prioritize [specific ethical principle].
+```
+
+## 4. Fair Representation
+
+### Balanced Treatment Checklist
+- [ ] Multiple perspectives on contested issues are presented
+- [ ] Minority/dissenting viewpoints are not dismissed without engagement
+- [ ] Subjects and communities are described accurately
+- [ ] Language is respectful and non-stigmatizing
+- [ ] Cultural context is acknowledged where relevant
+- [ ] Power dynamics are considered (who is studied vs. who studies)
+- [ ] Geographic and cultural diversity in sources
+
+### Sensitive Topics
+- Indigenous knowledge: Identify and follow the specific community-governance
+  convention in scope; use OCAP (Ownership, Control, Access, Possession) only
+  where the relevant First Nations authority or community has adopted it
+- Disability: Person-first language unless community prefers identity-first
+- Gender/sexuality: Use inclusive, current terminology
+- Race/ethnicity: Use preferred terminology of the communities discussed
+- Socioeconomic status: Avoid deficit framing
+- Mental health: Avoid stigmatizing language
+
+### Representation Audit Questions
+1. Whose voices are centered? Whose are missing?
+2. Are communities described on their own terms?
+3. Is there implicit bias in the framing?
+4. Would the subjects/communities recognize themselves in this description?
+
+## 5. Data Ethics
+
+### Data Source Ethics — Portable Integrity Planning
+
+These are **planning and documentation conventions**, not legal conclusions.
+Applicable law, licence interpretation, and institutional authorization remain
+with the responsible authority.
+
+- [ ] For each source, the asserted access/use basis and unresolved questions are documented
+- [ ] Public data: Publication status, access terms, provenance, and licence are recorded
+- [ ] Licensed data: Relevant licence terms and the team's intended use are recorded
+- [ ] Scraped data: Site terms, robots instructions, collection method, and unresolved authority questions are recorded
+- [ ] Personal data: Each candidate data-protection authority is recorded; no jurisdiction or legal basis is inferred
+- [ ] Institutional data: The supplying institution's documented access status and conditions are recorded
+
+### Privacy Protection
+
+The following are **portable technical planning prompts**, not a universal
+privacy-law checklist:
+
+- [ ] Direct identifiers, quasi-identifiers, platform metadata, and re-link keys are inventoried
+- [ ] The asserted authority or institutional basis for collecting and processing identifiers is recorded; consent is not presumed to be the only or applicable basis
+- [ ] Aggregation, masking, or other disclosure controls have been considered for small-N groups and identifiable institutions
+- [ ] Access roles, recipients, security, retention, deletion, reuse, and incident handling are documented
+- [ ] If the exact GDPR profile is selected, the trace points to the applicable bounded rows (`eu.gdpr.article-6.lawful-basis`, `eu.gdpr.article-9.special-category-research`, `eu.gdpr.article-13.direct-collection-information`, `eu.gdpr.article-14.indirect-collection-information`, and/or `eu.gdpr.article-89.1.research-safeguards`) rather than asserting a result
+
+### AI-Specific Data Concerns
+- [ ] AI training data biases acknowledged
+- [ ] AI knowledge cutoff date noted
+- [ ] AI-generated data clearly labeled as such
+- [ ] No circular citation (AI cites AI-generated content)
+
+## 6. Conflict of Interest
+
+### Types to Assess
+- [ ] Financial: Funding source, consulting relationships
+- [ ] Institutional: Author evaluating own institution
+- [ ] Intellectual: Author defending own prior work
+- [ ] Personal: Relationships with subjects/stakeholders
+- [ ] Political: Government-funded research on government policy
+- [ ] Commercial: Industry connections or product interests
+- [ ] AI-specific: AI tool company influence on research design
+
+### Disclosure Requirement
+Any identified conflict must be disclosed in the report, with an assessment of whether it could have influenced the findings.
+
+## 7. Reproducibility Ethics
+
+### Documentation Requirements
+- [ ] Search strategies documented (databases, terms, dates)
+- [ ] Inclusion/exclusion criteria documented
+- [ ] Analytical methods described in replicable detail
+- [ ] AI prompts/instructions documented (if relevant)
+- [ ] Data processing steps documented
+- [ ] Code/scripts shared (if applicable)
+
+### Reproducibility Statement Template
+```
+Reproducibility: The search strategy, inclusion criteria, and analytical
+methods used in this research are documented in [section/appendix]. The
+AI-assisted components used [specific prompts/parameters]. Researchers
+wishing to replicate or extend this work should note [relevant limitations
+or conditions].
+```
+
+## 8. Human Subjects Ethics
+
+> **Authority boundary (#665/#666/#680):** Mixed-jurisdiction review-level mappings are not part of this checklist. This section collects facts and exact bounded requirement pointers only; the output pathway is `institutional determination required`. Unknown selection or applicability stays unresolved, and neither this checklist nor #666 emits a review determination, readiness result, compliance conclusion, or authorization.
+
+Apply `shared/references/human_subjects_authority_protocol.md` and use only
+curated rows from `shared/human_subjects_authority_registry.json`. All shipped
+profiles are bounded subsets.
+
+### 8.1 Authority Context and Exact Facts
+
+- [ ] The review-ethics and data-protection axes each have an exact selected #666 profile or a visible unresolved state
+- [ ] Profile selection records id, version, digest, and authority scope; geography, funder, language, and topic are not used to infer selection
+- [ ] Confirmed, false, and unknown facts remain distinct
+- [ ] The exact #666 fact catalogue has been addressed where relevant:
+  - [ ] `activity.human_subjects`
+  - [ ] `support.us_hhs_or_common_rule`
+  - [ ] `scope.tw_hsra`
+  - [ ] `scope.eu_gdpr`
+  - [ ] `processing.personal_data`
+  - [ ] `data.special_category`
+  - [ ] `data.collected_from_subject`
+  - [ ] `purpose.scientific_research`
+  - [ ] `gdpr.member_state_research_law_identified`
+- [ ] Additional interaction, intervention, recruitment, population, power-relationship, sensitivity, deception, identifiability, and risk facts are labelled `illustrative/unprofiled` when they are not in the registry
+
+Vulnerability, sensitivity, deception, public availability, identifiability,
+and estimated risk are facts for the responsible institution. They do not map
+to a pathway in this checklist.
+
+### 8.2 Exact Bounded Requirement Trace
+
+Use these identifiers only as pointers to the full selected registry row,
+including its predicate, obligated actor, consumer scopes, expected evidence,
+and primary-source anchor. Do not restate them as universal requirements.
+
+| Axis | Exact #666 requirement ids currently available |
+|---|---|
+| Review ethics — U.S. bounded profile | `us.45cfr46.107.irb-composition`; `us.45cfr46.116.informed-consent` |
+| Review ethics — Taiwan bounded profile | `tw.hsra.article-7.committee-composition`; `tw.hsra.article-14.consent-information` |
+| Data protection — GDPR bounded profile | `eu.gdpr.article-6.lawful-basis`; `eu.gdpr.article-9.special-category-research`; `eu.gdpr.article-13.direct-collection-information`; `eu.gdpr.article-14.indirect-collection-information`; `eu.gdpr.article-89.1.research-safeguards`; `eu.gdpr.article-89.2.member-state-derogation` |
+
+- [ ] Each pointer belongs to an exactly selected profile whose resolved result permits profile-dependent use
+- [ ] Committee-composition rows remain with their institutional/committee actors and are not converted into investigator packet requirements
+- [ ] U.S. and Taiwan rows retain their own authority vocabulary and are not translated into one another's pathway terms
+- [ ] Requirements outside these bounded rows are labelled `unprofiled` and referred to the responsible authority
+
+### 8.3 Participant-Information and Consent Preparation
+
+- [ ] If applicable after exact resolution, the trace uses `us.45cfr46.116.informed-consent` or `tw.hsra.article-14.consent-information` rather than a merged universal element list
+- [ ] If the selected GDPR facts resolve direct collection, participant information points to `eu.gdpr.article-13.direct-collection-information`; if they resolve collection elsewhere, it points to `eu.gdpr.article-14.indirect-collection-information`
+- [ ] Waiver, exception, signature, assent, representative, recording, community-governance, sector, and institution-specific questions absent from the registry are visibly `unprofiled`
+- [ ] Online-survey language accurately says whether the platform, institution, or research team collects or can access IP addresses, cookies, device identifiers, timestamps, contact fields, or other metadata
+- [ ] A statement that metadata is not recorded is used only after platform settings and actual exports have been verified; unknown behavior is disclosed as unknown
+- [ ] An electronic `I agree` interaction is described as an interaction record, not automatically as legally sufficient consent or signature
+
+### 8.4 Data-Handling Facts — Illustrative Technical Planning
+
+These prompts do not establish an authority-defined anonymity,
+de-identification, or compliance status:
+
+- [ ] Direct identifiers, quasi-identifiers, quotations, rare attributes, and small-cell risks are inventoried
+- [ ] Every identifier/re-link key has a named holder, access boundary, purpose, and retention state
+- [ ] Platform collection, exports, logs, backups, recipients, transfers, encryption, access controls, retention, destruction, and incident handling are documented
+- [ ] Participant-facing statements match the actual data flow and withdrawal/deletion feasibility
+- [ ] Terminology is bound to the named convention in `shared/references/irb_terminology_glossary.md`
+
+### 8.5 Population and Context Facts — Illustrative and Unprofiled
+
+| Population or context | Facts/questions to prepare; no pathway or requirement implied |
+|---|---|
+| Minors or participants with uncertain decision-making capacity | age, capacity, representative/assent questions, accessible explanation, and the applicable authority's process |
+| Persons with disabilities | accessibility needs, communication format, capacity assumptions, support persons, and accommodation plan |
+| Students, employees, or dependent relationships | recruitment authority, grading/employment effects, alternatives, privacy, and role separation |
+| Indigenous peoples or communities | the specific community-governance authority/convention, collective and individual interests, data governance, and engagement expectations |
+| Economically constrained participants | compensation, alternatives, dependency, undue-influence questions, and local institutional assessment |
+| Incarcerated or otherwise institutionally confined participants | custodial setting, permission structure, voluntariness/coercion risks, privacy, and the applicable authority's process |
+
+- [ ] Relevant population and contextual facts have been recorded without treating identity or vulnerability as an automatic review level
+- [ ] Proposed safeguards and unresolved questions are presented to the responsible institution without a verdict
+- [ ] Current institutional timeline, training, submission, reporting, and authorization requirements are requested; absent a dated response, they remain `unknown/unprofiled`
+
+> For the portable fact flow, exact requirement anchors, survey metadata disclosure rule, and illustrative Taiwan/institution planning questions, see `references/irb_decision_tree.md`. Do not treat that reference as authorization.
+
+---
+
+## Quick Audit Checklist (Pre-Delivery Self-Check)
+
+Before delivery, confirm ALL items (this is a self-check, not a veto):
+
+- [ ] AI disclosure present and accurate
+- [ ] All references spot-checked (minimum 20%)
+- [ ] No fabricated citations detected
+- [ ] Dual-use assessment completed
+- [ ] Fair representation reviewed
+- [ ] Data-source bases and unresolved legal/ethical questions documented
+- [ ] Conflicts of interest disclosed
+- [ ] Reproducibility documentation provided
+- [ ] Writing is inclusive and respectful
+- [ ] Report benefits stated audience without causing foreseeable harm
+- [ ] If human-subjects activity is possible, are exact authority pointers or unresolved states, institutional questions, and the `institutional determination required` boundary visible?

package/vendor/ars/deep-research/references/failure_paths.md

@@ -0,0 +1,355 @@
+# Failure Paths — Research Pipeline Failure Path Map
+
+## Overview
+
+This document lists all failure scenarios that may be encountered across all modes of the deep-research skill, along with their detection conditions, user notification messages, handling steps, and recovery paths. The purpose is to ensure every failure scenario has a clear handling strategy, preventing users from reaching a dead end.
+
+---
+
+## Failure Path Summary
+
+| # | Failure Scenario | Affected Modes | Severity | Handling Strategy |
+|---|---------|---------|---------|---------|
+| F1 | RQ cannot converge | full, socratic | Medium | Full: candidate workflow; Socratic: summarize user-expressed directions / suggest literature exploration |
+| F2 | Insufficient literature | full, quick, lit-review | High | Expand search strategy |
+| F3 | Methodology mismatch | full | High | Return to Phase 1 |
+| F4 | Devil's Advocate CRITICAL | full | Critical | STOP + correct |
+| F5 | Ethics BLOCKED | full, review | Critical | STOP + remediation path |
+| F6 | Socratic dialogue does not converge | socratic | Medium | Switch to full mode |
+| F7 | User abandons mid-process | all | Low | Save progress |
+| F8 | Only Chinese-language literature available | full, lit-review | Medium | Switch search strategy |
+| F9 | All source quality below threshold | full, fact-check | High | Downgrade or expand sources |
+| F10 | Conclusions inconsistent with evidence | full | High | Return to Phase 3 |
+| F11 | Revision loop exceeds limit | full | Medium | Force-complete + limitation list |
+| F12 | Interdisciplinary bridging failure | full | Low | Revert to single discipline |
+
+---
+
+## Detailed Failure Paths
+
+### F1: Research Question Cannot Converge
+
+**Affected Modes**: `full` (Phase 1), `socratic` (Layer 1)
+**Severity**: Medium
+
+**Trigger Conditions**:
+- `full` mode: research_question_agent interaction exceeds 3 rounds, user still cannot determine the RQ
+- `socratic` mode: Layer 1 exceeds 5 rounds, user repeatedly revises without a clear direction
+
+**User Notification — `full` mode**:
+> I notice we've been working on the research question for several rounds without a stable direction. I can organize the constraints we have identified, compare a small set of candidate questions, or pause for a literature review before we choose.
+
+**User Notification — `socratic` mode**:
+> We have not reached one research-question direction yet. I can summarize only the directions and preferences you have already expressed, continue with one focused question, or pause for a literature review before we return to the framing. I will not add candidate questions unless you explicitly ask me to leave non-generation Socratic mode and propose them.
+
+**Handling Steps — `full` mode**:
+1. Compile key topics discussed and user-expressed preferences.
+2. Produce 3 candidate RQs, each with a brief explanation and rough FINER assessment.
+3. Ask the user to select or revise the closest one as a starting point.
+4. If the user still cannot choose, suggest `lit-review` mode before returning.
+
+**Handling Steps — `socratic` mode (default)**:
+1. Compile a clearly labeled summary containing only directions and preferences
+   already expressed by the user; do not fill unresolved slots.
+2. Ask one focused question that helps the user distinguish those directions,
+   or offer to pause.
+3. Offer `lit-review` as an evidence-gathering route before restarting Layer 1.
+4. Do not generate, FINER-score, rank, or present a menu of candidate RQs.
+   Non-convergence alone is never authorization.
+
+**Explicit generation request — visible exit from Socratic mode**:
+- If, and only if, the user explicitly asks the system itself to propose
+  candidates, first state that the response is leaving non-generation Socratic
+  mode and emit this exact marker on a standalone line:
+  `[SOCRATIC-NON-GENERATION-EXIT: explicit_user_request]`.
+- Only after the marker may the full-mode candidate workflow run. Label every
+  candidate as an AI-generated starting point, never as a user-derived insight.
+- Do not silently re-enter Socratic mode; re-entry requires an explicit user
+  request to resume guided questioning.
+
+**Recovery Paths**:
+- Full mode: select or revise a candidate RQ, then continue the workflow.
+- Socratic mode: the user refines one of their own expressed directions, then
+  restart Layer 1 without generating candidates.
+- Do `lit-review`, then restart RQ clarification after the review is complete.
+- Pause with the user-expressed-direction summary preserved.
+
+---
+
+### F2: Insufficient Literature
+
+**Affected Modes**: `full` (Phase 2), `quick`, `lit-review`
+**Severity**: High
+
+**Trigger Conditions**:
+- bibliography_agent finds < 5 usable sources after standard search strategy
+- After excluding quality-unqualified sources, < 3 remain
+
+**User Notification Message**:
+> With the current search strategy, I found only limited relevant literature. This could mean: (1) this is a very new research area; (2) the search keywords need adjustment; (3) the research question scope may need refinement. Let me try expanding the search strategy.
+
+**Handling Steps**:
+1. Expand search keywords (synonyms, broader terms, related concepts)
+2. Expand database scope (add grey literature, policy reports, working papers)
+3. Relax time range (from past 5 years to past 10 years)
+4. Try keywords from adjacent disciplines
+5. If still insufficient → suggest the user consider adjusting the RQ or accept this as an exploratory study
+
+**Recovery Paths**:
+- Expanded search yields sufficient literature → continue original workflow
+- Accept as exploratory research → adjust report positioning, emphasize the study's pioneering nature
+- Adjust RQ → return to Phase 1
+
+---
+
+### F3: Methodology Mismatch
+
+**Affected Modes**: `full` (Checkpoint 1)
+**Severity**: High
+
+**Trigger Conditions**:
+- devils_advocate_agent at Checkpoint 1 determines that the methodology proposed by research_architect_agent cannot answer the RQ produced by research_question_agent
+- There is a logical gap between the methodology and the RQ
+
+**User Notification Message**:
+> Devil's Advocate found an important issue in the methodology review: your research question asks "why," but your method design can only answer "whether." Let's go back and adjust — here are three possible directions...
+
+**Handling Steps**:
+1. Clearly state the gap between the RQ type (descriptive/comparative/causal/evaluative) and the method's capability
+2. Provide 3 alternative method suggestions, each with pros and cons
+3. Confirm whether the RQ needs adjustment to match a feasible method
+4. Re-execute research_architect_agent
+
+**Recovery Paths**:
+- Select an alternative method → regenerate Methodology Blueprint → Checkpoint 1 re-review
+- Adjust RQ → return to research_question_agent → redo Phase 1
+- Maximum 2 retries; if still mismatched on the 3rd attempt → suggest the user consult their advisor
+
+---
+
+### F4: Devil's Advocate CRITICAL
+
+**Affected Modes**: `full` (any Checkpoint)
+**Severity**: Critical
+
+**Trigger Conditions**:
+- devils_advocate_agent finds a Critical severity issue at any Checkpoint
+- Includes: fatal logical flaws, core assumptions that cannot hold, evidence contradicting conclusions
+
+**User Notification Message**:
+> STOP — Devil's Advocate found a critical issue that must be resolved before continuing:
+> [Specific issue description]
+> This is not an issue that can be ignored, as it fundamentally affects the research's validity.
+
+**Handling Steps**:
+1. Fully present the Critical issue's description, impact, and suggested correction direction
+2. Pause the workflow; do not allow advancement to the next Phase
+3. Wait for user response or correction
+4. After user correction → re-execute the Checkpoint
+5. 2 consecutive CRITICALs → suggest the user fundamentally rethink the research direction
+
+**Recovery Paths**:
+- User corrects the issue → re-execute Checkpoint → continue after PASS
+- User chooses to modify the RQ/method → return to the corresponding Phase
+- User abandons the direction → enter F7 workflow
+
+---
+
+### F5: Ethics BLOCKED
+
+**Affected Modes**: `full` (Phase 5), `review`
+**Severity**: Critical
+
+**Trigger Conditions**:
+- ethics_review_agent determines BLOCKED
+- Integrity violations only — see the authoritative list in `agents/ethics_review_agent.md` §"Blocking Conditions" (fabricated references, missing AI disclosure, plagiarism, systematic source misrepresentation, concrete harm-enabling specifics without safeguards; human subjects with no IRB plan is **CONDITIONAL**, not BLOCKED).
+- **NOT a trigger**: dual-use risk and politically/institutionally sensitive subject matter. Dual-use is advisory (routes to a Responsible Use Statement, never to BLOCKED — see `references/ethics_checklist.md` §3). Subject matter alone never blocks.
+
+**User Notification Message**:
+> Ethics Review flagged a Critical integrity concern that needs your confirmation before delivery:
+> [Specific issue list]
+> You can accept the suggested fix, **override with your reasoning** (recorded in the Ethics Decision Log), or revise. This is a one-time confirmation, not a veto. Suggested remediation paths:
+
+**Handling Steps**:
+1. List all flagged reasons, each with specific remediation suggestions
+2. Distinguish between remediable (e.g., add informed consent statement) and irremediable (e.g., research design inherently has integrity problems)
+3. Remediable issues → provide modification suggestions → re-review after user confirmation
+4. Irremediable issues → suggest fundamental redesign of the research
+5. User may override any flag with recorded reasoning → log to the Ethics Decision Log and proceed; do not re-block an overridden item
+
+**Recovery Paths**:
+- Fix integrity issues → re-execute ethics_review_agent → continue after the integrity disposition is recorded; any human-subjects authorization remains a separate institutional status
+- Modify research design → return to Phase 1
+- Issues are irremediable → suggest abandoning this research direction, provide alternative direction suggestions
+
+---
+
+### F6: Socratic Dialogue Does Not Converge
+
+**Affected Modes**: `socratic`
+**Severity**: Medium
+
+**Trigger Conditions**:
+- Dialogue exceeds 10 rounds, user still vacillating between different directions
+- Extracted INSIGHTs < 3
+
+**User Notification Message**:
+> We've explored several directions, each with its own value. However, I notice we've been going back and forth between a few directions — this may mean you have multifaceted interests in this topic, but it can also make the research hard to focus. Would you like to:
+> (A) Continue the Socratic dialogue, but focus on [the most promising direction] you just mentioned?
+> (B) Switch to full mode, and let my team help you systematically explore and converge?
+> (C) Take a pause, think it over, and come back later?
+
+**Handling Steps**:
+1. Compile currently extracted INSIGHTs
+2. Identify the 1-2 directions with the most convergence potential
+3. Provide 3 options (continue with focus / switch mode / pause)
+4. If user chooses to continue but still hasn't converged by round 15 → auto-compile + end
+
+**Recovery Paths**:
+- Continue with focus → restrict discussion scope, converge within 5 rounds
+- Switch to full mode → pass extracted INSIGHTs to research_question_agent
+- Pause → save INSIGHT list; user can re-enter at any time
+
+---
+
+### F7: User Abandons Mid-Process
+
+**Affected Modes**: All modes
+**Severity**: Low
+
+**Trigger Conditions**:
+- User explicitly states they don't want to continue ("never mind," "I quit," "too complicated")
+- User abandons after a long period of non-response
+
+**User Notification Message**:
+> No problem. I've saved the current progress. Here's a summary of what we've accomplished so far:
+> [List of completed stages and outputs]
+> You can re-enter at any time by saying "continue the previous research," and we'll pick up where we left off.
+
+**Handling Steps**:
+1. Save all outputs from the current stage (RQ Brief, INSIGHTs, Bibliography, etc.)
+2. Produce a progress summary
+3. Provide instructions for re-entry
+
+**Recovery Paths**:
+- User says "continue the previous research" → load saved outputs, continue from where interrupted
+- User starts over → entirely new workflow
+
+---
+
+### F8: Only Chinese-Language Literature Available
+
+**Affected Modes**: `full` (Phase 2), `lit-review`
+**Severity**: Medium
+
+**Trigger Conditions**:
+- English academic database searches (Web of Science, Scopus, PubMed, etc.) yield empty or very few results
+- The topic is strongly localized (e.g., Taiwan-specific policy, regulations, institutional systems)
+
+**User Notification Message**:
+> English-language literature on this topic is very limited, but Chinese-language literature resources are abundant. I will adjust the search strategy to include Chinese academic databases. Please note that citation conventions for Chinese-language literature in international publications may differ.
+
+**Handling Steps**:
+1. Switch search strategy to Chinese academic databases (Airiti Library, National Digital Library of Theses and Dissertations in Taiwan, CNKI)
+2. Re-search using Chinese keywords
+3. Note the language distribution of the literature in the report
+4. If the user needs an English report → provide suggestions for English citation format of Chinese literature
+5. If the user needs to publish internationally → suggest finding comparable international cases
+
+**Recovery Paths**:
+- Chinese literature is sufficient → continue workflow with clear language annotations
+- User needs international publication → suggest adjusting RQ to add a comparative perspective
+
+---
+
+### F9: All Source Quality Below Threshold
+
+**Affected Modes**: `full` (Phase 2), `fact-check`
+**Severity**: High
+
+**Trigger Conditions**:
+- source_verification_agent rates all found sources as Level V or below
+- No peer-reviewed sources
+
+**User Notification Message**:
+> The overall quality of currently found sources is low, lacking high-quality peer-reviewed research. This may indicate an emerging field, or the search strategy may need adjustment. I suggest we consider...
+
+**Handling Steps**:
+1. Expand source types (add policy reports, white papers, official statistics)
+2. Lower the threshold but clearly annotate quality levels
+3. Reposition the report as "preliminary exploration" rather than "systematic review"
+4. Add an "Evidence Quality Limitations" section to the report
+
+**Recovery Paths**:
+- Find sufficient alternative sources → continue workflow with clear quality annotations
+- Cannot find qualified sources → suggest the user consider conducting primary research
+
+---
+
+### F10: Conclusions Inconsistent with Evidence
+
+**Affected Modes**: `full` (Phase 5, Checkpoint 3)
+**Severity**: High
+
+**Trigger Conditions**:
+- editor_in_chief_agent or devils_advocate_agent finds in Phase 5 that report conclusions exceed the scope supported by the evidence
+
+**User Notification Message**:
+> The review found that some conclusions in the report go beyond what the evidence supports. Specifically:
+> [List of issues]
+> I will return for revision to ensure every conclusion has corresponding evidence support.
+
+**Handling Steps**:
+1. Flag all "over-inferred" conclusions
+2. For each flag: (a) weaken the conclusion to match the evidence, or (b) supplement with additional evidence
+3. Re-execute Checkpoint 3
+
+**Recovery Paths**:
+- Revision successful → complete Phase 6
+- Issues remain after revision → 2nd revision round
+- Issues remain after 2 revisions → convert issues to a "Research Limitations" section
+
+---
+
+### F11: Revision Loop Exceeds Limit
+
+**Affected Modes**: `full` (Phase 6)
+**Severity**: Medium
+
+**Trigger Conditions**:
+- Phase 6 revision has been executed 2 times (maximum), with unresolved Major issues remaining
+
+**User Notification Message**:
+> After two rounds of revision, the following issues have been resolved: [resolved list]. However, the following issues remain unresolved due to inherent research limitations: [unresolved list]. These will be listed in the "Acknowledged Limitations" section. The report is now the best version achievable under current conditions.
+
+**Handling Steps**:
+1. Compile resolved and unresolved issues
+2. Convert unresolved Major issues into the "Acknowledged Limitations" section
+3. Deliver the final report
+
+**Recovery Paths**:
+- User accepts → deliver the report
+- User does not accept → suggest redesigning the research from Phase 1
+
+---
+
+### F12: Interdisciplinary Bridging Failure
+
+**Affected Modes**: `full`
+**Severity**: Low
+
+**Trigger Conditions**:
+- synthesis_agent attempts interdisciplinary integration but cannot find meaningful connections
+- Conceptual frameworks from different disciplines cannot be reconciled
+
+**User Notification Message**:
+> I attempted to integrate perspectives from [Discipline A] and [Discipline B], but these two disciplines' understanding frameworks for this phenomenon differ substantially. Forcing integration may actually blur the focus. I suggest we center on the [primary discipline] framework, and mention other disciplines' perspectives in the discussion section as reference.
+
+**Handling Steps**:
+1. Select the primary disciplinary framework as the analytical foundation
+2. Present other disciplinary perspectives in an "Alternative Perspectives" or "Interdisciplinary Insights" section
+3. Do not force integration of irreconcilable frameworks
+
+**Recovery Paths**:
+- Focus on a single framework → continue workflow
+- User insists on interdisciplinary → suggest switching to mixed-methods or narrative review