Adversarial Strategy Catalog¶
Unified catalog of 15 adversarial review strategies synthesized from academic literature, industry practices, and emerging AI patterns — with 117+ citations across three independent research streams.
Key Findings¶
- 36 candidate strategies from three research streams (academic, industry, emerging) were deduplicated and synthesized into 15 distinct strategies
- Strategies cluster into 4 mechanistic families: Role-Based Adversarialism, Structured Decomposition, Dialectical Synthesis, and Iterative Self-Correction
- Academic strategies (CIA, DoD, Hegelian dialectics) provide the strongest evidence base with 70+ years of formalized practice
- LLM-specific strategies (Constitutional AI, Self-Refine, LLM-as-Judge) are natively compatible with single-model architectures
- The catalog enables criticality-based activation — from lightweight self-review (C1) to full 10-strategy tournaments (C4)
Unified Catalog of 15 Adversarial Review Strategies¶
The crown jewel of the adversarial research pipeline: 36 candidates from three parallel research efforts were deduplicated, analyzed for overlap, and synthesized into 15 distinct strategies with full profiles.
Methodology
The synthesis followed a rigorous pipeline:
- TASK-001 (Academic): 12 strategies from peer-reviewed sources — CIA structured analytic techniques, DoD red teaming, Hegelian dialectics, decision science (36 citations)
- TASK-002 (Industry): 14 strategies from industry practices and LLM-specific patterns (35 citations)
- TASK-003 (Emerging): 10 strategies from cross-domain emerging approaches (46 references)
- TASK-004 (Synthesis): Overlap analysis → deduplication → 15 unified strategies with standardized profiles
Each strategy profile includes: origin, mechanism, strengths, weaknesses, Jerry-specific applicability, P-003 compliance assessment, and token budget estimate.
Key Data: The 15 Strategies
| ID | Strategy | Family | One-Line Description |
|---|---|---|---|
| S-001 | Red Team Analysis | Role-Based | Independent team adopts adversary perspective to find vulnerabilities |
| S-002 | Devil's Advocate | Role-Based | Formally assigned critic builds strongest case against prevailing judgment |
| S-003 | Steelman Technique | Dialectical | Reconstruct argument in strongest form before critiquing |
| S-004 | Pre-Mortem Analysis | Role-Based | Imagine the plan has failed; work backward to identify causes |
| S-005 | Dialectical Inquiry | Dialectical | Two opposing plans from same data, debated to synthesis |
| S-006 | Analysis of Competing Hypotheses | Decomposition | Multiple hypotheses evaluated against all evidence in a matrix |
| S-007 | Constitutional AI Critique | Self-Correction | Critique outputs against explicit written principles iteratively |
| S-008 | Socratic Method | Dialectical | Probing questions to expose contradictions and assumptions |
| S-009 | Multi-Agent Debate | Dialectical | Multiple LLM agents argue across structured rounds |
| S-010 | Self-Refine | Self-Correction | Iterative generate-feedback-refine loop |
| S-011 | Chain-of-Verification | Decomposition | Generate verification questions for claims, answer independently |
| S-012 | FMEA | Decomposition | Systematic failure mode enumeration with severity/occurrence/detection scoring |
| S-013 | Inversion Technique | Decomposition | Ask "how would we guarantee failure?" to generate anti-pattern checklists |
| S-014 | LLM-as-Judge | Self-Correction | Rubric-based structured evaluation with numerical scores |
| S-015 | Prompt Adversarial Examples | Self-Correction | Adversarial prompt testing with graduated intensity |
Key Data: Mechanistic Families
| Family | Mechanism | Strategies | Best For |
|---|---|---|---|
| Role-Based Adversarialism | Designated agent adopts oppositional persona | S-001, S-002, S-004 | Breaking groupthink, challenging assumptions |
| Structured Decomposition | Systematic framework forces exhaustive enumeration | S-006, S-011, S-012, S-013 | Completeness, failure mode coverage |
| Dialectical Synthesis | Opposing positions constructed and reconciled | S-003, S-005, S-008, S-009 | Novel insights, balanced analysis |
| Iterative Self-Correction | Agent critiques and revises own output | S-007, S-010, S-014, S-015 | Quality scoring, constitutional compliance |
Academic Literature on Adversarial Review Strategies¶
The foundational research artifact documenting 12 strategies from peer-reviewed academic sources with formal citations and methodology descriptions.
Methodology
Research drew from five major academic domains:
- Intelligence analysis: CIA/DoD structured analytic techniques (Heuer & Pherson, 2014)
- Argumentation theory: Dialectical methods, formal argumentation (Toulmin, 1958)
- Decision science: Pre-mortem analysis, prospective hindsight (Klein, 1998)
- Cybersecurity: Threat modeling, STRIDE (Shostack, 2014; MIL-STD-1629A)
- AI safety: Constitutional AI, debate-based alignment (Bai et al., 2022; Irving et al., 2018)
Selection criteria required: formal publication, explicit adversarial mechanism, reproducible methodology, mechanistic distinctness, and LLM workflow applicability.
Key Data: Source Tiers
| Tier | Source Type | Count |
|---|---|---|
| Primary | Peer-reviewed papers, books with ISBN, government publications | 18 |
| Secondary | Major institution research (Anthropic, RAND, MITRE) | 6 |
| Tertiary | Conference proceedings, well-cited preprints | 3 |
Key finding: strategies cluster into three fundamental mechanistic families — role-based adversarialism (breaking groupthink), structured decomposition (ensuring completeness), and dialectical synthesis (producing novel insights).
Academic Research (861 lines, 36 citations)
Industry Practices & LLM-Specific Patterns¶
Research into 14 adversarial review strategies from software engineering practice (Fagan inspections, Google code review, ATAM), design review methodology, and LLM-specific self-correction patterns (Constitutional AI, Self-Refine, multi-agent debate).
Methodology
Surveyed industry software engineering practices (Fagan, 1976 through modern Google code review culture), design critique methodologies, LLM/AI adversarial systems (Constitutional AI, Self-Refine, multi-agent debate), and QA adversarial patterns. Identified the creator-critic-revision cycle as a universal convergent pattern across all four domains.
Key Data
| Domain | Strategies | Key Insight |
|---|---|---|
| Software Engineering | Fagan Inspection, Google Code Review, ATAM, Pair Programming | Deep adversarial traditions with measured defect-detection effectiveness |
| LLM-Specific | Constitutional AI, Self-Refine, Multi-Agent Debate | Directly implementable patterns for creator-critic-revision cycles |
| Design/Product | Design critique, stakeholder challenge | Present-critique-iterate mirrors the universal pattern |
| QA | Adversarial testing, boundary analysis | Testing-oriented adversarial methods |
35 citations across software engineering, AI/ML, and design methodology literature.
Industry Research (1,097 lines, 35 citations)
Emerging & Cross-Domain Adversarial Approaches¶
10 emerging adversarial review strategies discovered through cross-domain transfer analysis (legal, medical, military), cognitive science debiasing techniques, and frontier AI adversarial collaboration patterns.
Methodology
Applied cross-domain transfer analysis across legal (moot court), medical (M&M conferences), and military (wargaming) traditions. Identified cognitive debiasing techniques (Reference Class Forecasting, Inversion Technique) and AI-native patterns (Constitutional AI critique chains, progressive adversarial escalation) as underexplored adversarial review strategies. Explicit differentiation against TASK-001 and TASK-002 findings.
Key Data
| Category | Example Strategies | Novelty |
|---|---|---|
| Cross-Domain Transfer | Moot Court, M&M Conference, Wargaming | Centuries of refined adversarial practice, never formally applied to software review |
| Cognitive Debiasing | Reference Class Forecasting, Inversion Technique | Powerful adversarial tools rarely framed as review strategies |
| AI-Native | Constitutional AI Critique Chains, Progressive Adversarial Escalation | No direct pre-AI precedent; most applicable to Jerry's architecture |
| Meta-Strategy | Cynefin-Gated Selection | Matches adversarial intensity to problem complexity |
46 references spanning legal theory, medical practice, military doctrine, and AI safety research.
Emerging Research (706 lines, 46 references)