/deep-research

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

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. I notice we've been going back and forth between a few of them. Here are the directions you have raised so far, in the order you raised them: [the directions the user has expressed, listed chronologically, in the user's own words]

Would you like to: (A) Continue the Socratic dialogue, focusing on one of these directions? You choose which one. (B) Switch to full mode and let my team explore and converge systematically? (This leaves non-generation Socratic mode: the exact marker [SOCRATIC-NON-GENERATION-EXIT: explicit_user_request] is emitted on a standalone line before any candidate content.) (C) Take a pause, think it over, and come back later?

Handling Steps:

1. Compile currently extracted INSIGHTs
2. Summarize the directions the user has expressed, in the order they were expressed and in the user's wording. Do not rank them, do not pre-fill option (A) with a system pick, and do not FINER-score them; the choice among them is the user's
3. Provide 3 options (continue on a user-chosen direction / switch mode with the visible exit marker / pause)
4. If the user chooses to continue and the dialogue still does not converge, deep-research/agents/socratic_mentor_agent.md § Auto-End Conditions (Precise) decides when the dialogue ends. That section is the single authority for round caps and stagnation thresholds; this file states no round count of its own

Recovery Paths:

• Continue with focus → the user has chosen the direction; keep the Socratic questioning on it without limiting what the user may raise next
• Switch to full mode → emit the exit marker, then 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