critique（批判评审）

Critique: Multi-Lens Dialectical Refinement

Execute adversarial self-refinement through parallel evaluative lenses with cross-evaluation and recursive aggregation.

Architecture

┌──────────────────────────────────────────────────────────────────────────────┐
│                         DIALECTIC ENGINE v3                                  │
├──────────────────────────────────────────────────────────────────────────────┤
│  Φ0: CLASSIFY    → complexity assessment, mode selection, lens allocation    │
│  Φ1: THESIS      → committed position with claim DAG                         │
│  Φ2: MULTI-LENS  → N lenses evaluate thesis (N critiques)                    │
│      ANTITHESIS    + each lens evaluates others (N×(N-1) cross-evals)        │
│                    = N² total evaluation cells                               │
│  Φ3: AGGREGATE   → consensus/contested/unique extraction                     │
│      SYNTHESIS     + recursive compression passes → single output            │
│  Φ4: CONVERGE    → stability check, iterate or finalize                      │
└──────────────────────────────────────────────────────────────────────────────┘

PHASE DEPENDENCIES:
  Φ0 ──► Φ1 ──► Φ2a ──► Φ2b ──► Φ3 ──► Φ4
              (initial)  (cross)      │
                                      └──► Φ1 (if ITERATE)

Mode Selection

Automatic Mode Detection

def select_mode(query: str) -> Mode:
    """
    Select critique depth based on query characteristics.
    
    QUICK:  Simple claims, factual questions, narrow scope
    STANDARD: Moderate complexity, clear domain, some nuance
    FULL:   Complex arguments, multiple stakeholders, high stakes
    """
    indicators = {
        "quick": [
            len(query) < 200,
            single_claim(query),
            factual_verifiable(query),
            low_controversy(query)
        ],
        "full": [
            len(query) > 1000,
            multi_stakeholder(query),
            ethical_implications(query),
            policy_recommendation(query),
            high_stakes_decision(query)
        ]
    }
    
    if sum(indicators["quick"]) >= 3:
        return Mode.QUICK
    elif sum(indicators["full"]) >= 2:
        return Mode.FULL
    else:
        return Mode.STANDARD

Mode Specifications

Manual Triggers

Evaluative Lenses

Five orthogonal perspectives designed for comprehensive coverage with minimal overlap:

Lens Independence Validation

Lenses target distinct failure modes:

• S catches: invalid inference, circular reasoning, equivocation
• E catches: weak evidence, unfalsifiable claims, cherry-picking
• O catches: overgeneralization, edge cases, context dependence
• A catches: stronger alternatives, unconsidered objections
• P catches: implementation barriers, unintended consequences

Overlap detection: If two lenses identify the same issue, it's either a genuine high-priority concern (reinforce) or a lens calibration problem (investigate).

Execution Protocol

Φ0: Classification & Mode Selection

def classify_and_configure(query: str) -> Config:
    mode = select_mode(query)
    
    configs = {
        Mode.QUICK: {
            "lenses": ["S", "E", "A"],
            "cross_eval": False,
            "cycles": 1,
            "threshold": 0.85,
            "token_budget": 800
        },
        Mode.STANDARD: {
            "lenses": ["S", "E", "O", "A", "P"],
            "cross_eval": "selective",  # 10 highest-value cells
            "cycles": 2,
            "threshold": 0.92,
            "token_budget": 2000
        },
        Mode.FULL: {
            "lenses": ["S", "E", "O", "A", "P"],
            "cross_eval": "complete",   # All 25 cells
            "cycles": 3,
            "threshold": 0.96,
            "token_budget": 4000
        }
    }
    
    return Config(**configs[mode], mode=mode)

Output: [CRITIQUE:Φ0|mode={m}|lenses={n}|cross={type}|budget={t}]

Φ1: Thesis Generation

Generate committed response with explicit claim DAG.

Requirements:

1. State positions with falsifiable specificity
2. Build claim graph with stability ordering:

- F (FOUNDATIONAL) — axioms, definitions (immutable after Φ1) - S (STRUCTURAL) — derived claims (attackable) - P (PERIPHERAL) — applications (most vulnerable)

1. Verify acyclicity (DAG enforcement)
2. Compute initial topology metrics

Schema:

thesis:
  response: "{Complete committed response}"
  claims:
    - id: C1
      content: "{Specific falsifiable claim}"
      stability: F|S|P
      supports: [C2, C3]
      depends_on: []
      confidence: 0.0-1.0
      evidence_type: empirical|logical|definitional|analogical
  topology:
    nodes: {n}
    edges: {e}
    density: {e/n}  # Target ≥2.0
    cycles: 0       # Must be 0 (enforced)
  aggregate_confidence: 0.0-1.0
  completion_marker: "Φ1_COMPLETE"  # Required for Φ2 to proceed

Output: [CRITIQUE:Φ1|claims={n}|edges={e}|η={density}|conf={c}|✓]

Φ2: Multi-Lens Antithesis

Φ2a: Initial Lens Evaluations

Prerequisite: Φ1.completion_marker == "Φ1_COMPLETE"

Each lens independently evaluates thesis using attack vectors:

# STRUCTURAL lens attacks
structural:
  - non_sequitur: "Conclusion does not follow from premises"
  - circular_reasoning: "Conclusion presupposed in premises"
  - false_dichotomy: "Excluded middle options"
  - equivocation: "Term shifts meaning mid-argument"

# EVIDENTIAL lens attacks  
evidential:
  - insufficient_evidence: "Claim exceeds evidential support"
  - cherry_picking: "Counter-evidence unaddressed"
  - unfalsifiable: "No possible disconfirming evidence"
  - correlation_causation: "Causal claim from correlational data"

# SCOPE lens attacks
scope:
  - overgeneralization: "Specific case → universal claim"
  - edge_case: "Valid boundary defeats universal"
  - context_dependence: "Unstated contextual requirements"

# ADVERSARIAL lens attacks
adversarial:
  - steel_man: "Strongest form of opposition"
  - alternative_explanation: "Competing hypothesis equally plausible"
  - precedent_contradiction: "Accepted instance defeats thesis"

# PRAGMATIC lens attacks
pragmatic:
  - implementation_barrier: "Cannot be executed as stated"
  - unintended_consequence: "Second-order effects harmful"
  - scaling_failure: "Works small, fails large"

Per-lens output:

lens_evaluation:
  lens: S|E|O|A|P
  attacks:
    - target: C{id}
      type: "{attack_vector}"
      content: "{Specific critique}"
      severity: fatal|major|minor|cosmetic
      confidence_impact: -0.0 to -1.0
  summary_score: 0.0-1.0
  completion_marker: "Φ2a_{lens}_COMPLETE"

Completion Gate: All lenses must have completion_marker before Φ2b proceeds.

Φ2b: Cross-Lens Evaluation

Prerequisite: All Φ2a_{lens}_COMPLETE markers present

QUICK mode: Skip Φ2b entirely

STANDARD mode: Evaluate 10 highest-value cells:

• High-severity attacks from each lens (5 cells)
• Highest-confidence attacks cross-checked by adjacent lens (5 cells)

FULL mode: Complete 5×5 matrix (25 cells, minus 5 diagonal = 20 evaluations)

Cross-evaluation matrix:
    │  S eval │  E eval │  O eval │  A eval │  P eval │
────┼─────────┼─────────┼─────────┼─────────┼─────────┤
S → │    —    │   S→E   │   S→O   │   S→A   │   S→P   │
E → │   E→S   │    —    │   E→O   │   E→A   │   E→P   │
O → │   O→S   │   O→E   │    —    │   O→A   │   O→P   │
A → │   A→S   │   A→E   │   A→O   │    —    │   A→P   │
P → │   P→S   │   P→E   │   P→O   │   P→A   │    —    │

Cross-eval output:

cross_evaluation:
  evaluator: S|E|O|A|P
  evaluated: S|E|O|A|P
  verdict: endorse|partial|reject
  agreements: ["{attack_ids}"]
  disagreements:
    - attack: "{attack_id}"
      objection: "{Why evaluator disagrees}"
  missed: ["{What evaluator would add}"]
  calibration: "{Over/under severity assessment}"

Output: [CRITIQUE:Φ2|mode={m}|attacks={n}|cross={cells}|✓]

Φ3: Aggregation & Synthesis

Phase 3a: Matrix Analysis

def analyze_matrix(all_attacks: list, cross_evals: Matrix) -> Analysis:
    # Consensus: ≥80% lenses agree
    consensus = [a for a in all_attacks if agreement_rate(a) >= 0.80]
    
    # Contested: 40-79% agreement
    contested = [a for a in all_attacks if 0.40 <= agreement_rate(a) < 0.80]
    
    # Unique: Single lens, but cross-eval endorsed
    unique = [a for a in all_attacks 
              if source_count(a) == 1 and cross_endorsed(a)]
    
    # Rejected: <40% agreement AND cross-eval rejection
    rejected = [a for a in all_attacks 
                if agreement_rate(a) < 0.40 and cross_rejected(a)]
    
    return Analysis(consensus, contested, unique, rejected)

Phase 3b: Conflict Resolution

For contested items:

def resolve_contested(contested: list, matrix: Matrix) -> list:
    resolutions = []
    for attack in contested:
        support_weight = sum(credibility(s) for s in supporters(attack))
        oppose_weight = sum(credibility(o) for o in opposers(attack))
        
        if support_weight > oppose_weight * 1.5:
            resolution = "ADOPT"
        elif oppose_weight > support_weight * 1.5:
            resolution = "REJECT"
        else:
            resolution = "CONDITIONAL"
        
        resolutions.append(Resolution(attack, resolution, rationale(attack)))
    return resolutions

Phase 3c: Recursive Compression

Pass 1: Apply consensus → Core modifications (mandatory)
Pass 2: Apply contested → Conditional modifications (with qualifications)
Pass 3: Apply unique → Enhancement layer (optional enrichment)
Pass 4: Validate coherence → If failed, re-compress with tighter constraints

Maximum compression passes: 4 (prevent infinite recursion)

Synthesis output:

synthesis:
  response: "{Refined response}"
  modifications:
    from_consensus: [{claim, action, rationale}]
    from_contested: [{claim, action, condition}]
    from_unique: [{claim, enhancement}]
  rejected_attacks: [{attack, rejection_rationale}]
  residual_uncertainties: [{uncertainty, disagreeing_lenses, impact}]
  confidence:
    initial: {Φ1}
    final: {post-synthesis}

Output: [CRITIQUE:Φ3|consensus={n}|contested={n}|unique={n}|rejected={n}|conf={f}]

Φ4: Convergence Check

Convergence Formula:

convergence = (
    0.30 * semantic_similarity(Φ1, Φ3) +
    0.25 * graph_similarity(Φ1.claims, Φ3.claims) +
    0.25 * confidence_stability(Φ1.conf, Φ3.conf) +
    0.20 * consensus_rate(Φ3.consensus / total_attacks)
)

Threshold Justification:

• 0.85 (QUICK): Acceptable for low-stakes, rapid iteration
• 0.92 (STANDARD): Balances thoroughness with efficiency
• 0.96 (FULL): High confidence required for complex/high-stakes

Outcomes:

• CONVERGED: Score ≥ threshold → output Φ3 synthesis
• ITERATE: Score < threshold AND cycles < max → Φ3 becomes new Φ1
• EXHAUSTED: Cycles exhausted → output Φ3 with uncertainty report

Output: [CRITIQUE:Φ4|conv={score}|{STATUS}|iter={n}/{max}]

Graceful Degradation

When resources constrained (token budget, time pressure):

FULL → interrupt → Continue as STANDARD
STANDARD → interrupt → Continue as QUICK
QUICK → interrupt → Output best available synthesis with uncertainty flag

Degradation markers:

degraded_output:
  original_mode: FULL
  actual_mode: STANDARD
  skipped_phases: [Φ2b_partial]
  confidence_penalty: -0.1
  recommendation: "Re-run in FULL mode for complete analysis"

Compact Output Mode

[CRITIQUE|mode={m}|L={lenses}|c={cycle}/{max}]
[Φ1|n{claims}|e{edges}|η{density}|conf{c}|✓]
[Φ2|attacks{n}|cross{cells}|S:{s}|E:{e}|O:{o}|A:{a}|P:{p}|✓]
[Φ3|consensus{n}|contested{n}|unique{n}|rejected{n}|✓]
[Φ4|conv{score}|{STATUS}|conf{initial}→{final}]

SYNTHESIS: {2-3 sentence refined conclusion}
KEY_CHANGES: {Most significant modifications from Φ1}
RESIDUAL: {Primary unresolved uncertainty, if any}

Meta-Cognitive Markers

[CLASSIFYING]  — Φ0: determining mode and resources
[COMMITTING]   — Φ1: stating without hedge
[LENS:X]       — Φ2a: evaluating from lens X perspective
[CROSS:X→Y]    — Φ2b: lens X evaluating lens Y's critique
[CONSENSUS]    — Φ3a: noting cross-lens agreement
[CONTESTED]    — Φ3a: noting genuine disagreement
[RESOLVING]    — Φ3b: applying resolution protocol
[COMPRESSING]  — Φ3c: recursive synthesis pass
[CONVERGING]   — Φ4: stability detected
[DEGRADING]    — Resource constraint, reducing scope

Constraints

1. Phase Dependencies: Each phase requires predecessor completion marker
2. DAG Enforcement: Claim graph must remain acyclic; circular reasoning = fatal
3. Stability Ordering: FOUNDATIONAL claims immutable after Φ1
4. Genuine Critique: Softball attacks detected via cross-eval and rejected
5. Compression Termination: Max 4 recursive passes in Φ3c
6. Convergence Cap: Max cycles from config; output uncertainty if exhausted
7. Token Budget: Respect mode-specific limits; degrade gracefully if exceeded

Integration

• hierarchical-reasoning: Map lenses to strategic/tactical/operational
• graph: Claim topology analysis, k-bisimulation on evaluation matrix
• think: Mental models power individual lens templates
• non-linear: Subagent spawning for parallel lens execution
• infranodus: Graph gap detection enhances STRUCTURAL lens
• component: Structure critique outputs as validatable configuration

References

• references/lens-specifications.md — Complete lens templates and attack vectors
• references/cross-evaluation-protocol.md — Matrix construction and analysis
• references/aggregation-algorithms.md — Consensus extraction and compression