integrated-reasoning-v2（综合推理 v2）

Integrated Reasoning v2 - Meta-Orchestration

Purpose: Select and orchestrate optimal reasoning pattern(s) for your problem. V2 addresses limitations of v1: adds new patterns, replaces order-dependent decision tree with weighted scoring, includes feedback loops, and fixes confidence aggregation.

Available Reasoning Patterns (9)

Pattern Selection: Weighted Multi-Dimensional Scoring

Step 1: Assess Problem Characteristics

Score each dimension (1-5):

| Dimension | Score | Description |
|-----------|-------|-------------|
| **Sequential Dependencies** | _/5 | Do steps depend on previous steps? |
| **Criteria Clarity** | _/5 | Can you clearly evaluate solutions? |
| **Solution Space Known** | _/5 | Do you know the options? |
| **Single Answer Needed** | _/5 | Need ONE answer vs multiple options? |
| **Evidence Available** | _/5 | Can you gather discriminating evidence? |
| **Opposing Valid Views** | _/5 | Are there legitimate conflicting perspectives? |
| **Problem Novelty** | _/5 | Is this unprecedented in your domain? |
| **Robustness Required** | _/5 | Need to stress-test before committing? |
| **Solution Exists** | _/5 | Do you have a candidate solution to evaluate? |
| **Time Pressure** | _/5 | How constrained is decision time? (5=minutes) |
| **Stakeholder Complexity** | _/5 | Multiple parties with competing interests? |

Step 1.5: Time Pressure Fast-Path

CRITICAL: If Time Pressure = 5 (emergency/incident):

• Skip full scoring
• Use Rapid Triage Reasoning (RTR) directly
• RTR is optimized for decisions under extreme time constraints

If Time Pressure ≥ 4:

• Consider RTR unless problem is clearly sequential (use SRC) or diagnostic (use HE)
• Apply abbreviated scoring (skip orchestration considerations)

Step 2: Calculate Pattern Affinity Scores

ToT  = (Criteria × 0.35) + (SingleAnswer × 0.30) + (SpaceKnown × 0.20) + ((6-Novelty) × 0.15)

BoT  = ((6-SpaceKnown) × 0.35) + ((6-SingleAnswer) × 0.30) + ((6-Criteria) × 0.20) + (Novelty × 0.15)

SRC  = (Sequential × 0.45) + (Criteria × 0.25) + (SingleAnswer × 0.20) + ((6-OpposingViews) × 0.10)

HE   = (Evidence × 0.40) + (SingleAnswer × 0.30) + ((6-Novelty) × 0.20) + ((6-OpposingViews) × 0.10)

AR   = (Robustness × 0.40) + (SolutionExists × 0.30) + ((6-Novelty) × 0.15) + (Evidence × 0.15)
       # NOTE: AR requires SolutionExists ≥ 3, otherwise score = 0

DR   = (OpposingViews × 0.50) + (Criteria × 0.20) + ((6-Evidence) × 0.15) + (MIN(SingleAnswer, OpposingViews) × 0.15)
       # NOTE: V2.1 fix - SingleAnswer no longer penalized when OpposingViews is high

AT   = (Novelty × 0.45) + ((6-SpaceKnown) × 0.30) + ((6-Evidence) × 0.15) + ((6-Sequential) × 0.10)

RTR  = (TimePressure × 0.50) + (SingleAnswer × 0.25) + (Evidence × 0.15) + ((6-Novelty) × 0.10)
       # NOTE: RTR auto-selected when TimePressure = 5

NDF  = (StakeholderComplexity × 0.45) + (OpposingViews × 0.25) + ((6-Criteria) × 0.15) + ((6-TimePressure) × 0.15)
       # NOTE: NDF requires StakeholderComplexity ≥ 3 to be considered

Formula Validation Rules (V2.1):

• AR returns 0 if SolutionExists < 3 (nothing to attack)
• RTR auto-triggers when TimePressure = 5 (emergency mode)
• NDF returns 0 if StakeholderComplexity < 3 (single decision-maker)
• If multiple patterns score within 0.3 of each other, use uncertainty propagation (Step 2.5)

Step 2.5: Uncertainty Propagation (V2.1)

When dimension scores are uncertain, propagate uncertainty to pattern selection:

## Uncertainty Assessment

For each dimension where you're unsure (±1 point uncertainty):

1. Calculate pattern scores at LOW end (dimension - 1)
2. Calculate pattern scores at HIGH end (dimension + 1)
3. If different pattern wins at each end → **Flag as uncertain selection**

### Handling Uncertain Selections

**If same pattern wins both ends**: Proceed with confidence
**If different patterns win**:
  - Run BOTH patterns in parallel (if time permits)
  - OR use the pattern that's more robust to being wrong
  - OR gather more information to reduce dimension uncertainty

### Uncertainty Discount
Apply -5% to final confidence for each uncertain dimension that affects the winning pattern.

Step 3: Interpret Scores

Direct Analysis (10th Pattern)

When all pattern scores are below 2.5, no specialized methodology is warranted:

## Direct Analysis

**When to use**: All patterns score <2.5 (problem doesn't match any pattern's strengths)

**Approach**:
1. No special framework needed
2. Just think through the problem directly, step by step
3. Use common sense and straightforward reasoning
4. Appropriate for simple problems that don't need cognitive overhead

**Examples**:
- Simple factual questions
- Straightforward calculations
- Basic lookups or translations
- Problems with obvious solutions

**Confidence**: Use intuitive confidence based on problem clarity
- Clear problem + clear answer = high confidence
- Any ambiguity = document it explicitly

Direct Analysis Threshold

Use Direct Analysis (no specialized pattern) when:

• MAX(all pattern scores) < 4.0 AND
• No dimension scored >= 4 AND
• Problem is not time-critical (TimePressure < 4)

Direct Analysis = simple step-by-step reasoning without framework overhead.

Sequential vs Parallel Execution Decision

Use SEQUENTIAL when:

• Pattern B depends on Pattern A's output (e.g., BoT finds options → ToT optimizes)
• Confidence from A affects what to do in B
• Evidence from A eliminates need for B
• One pattern discovers the problem is different than expected

Use PARALLEL when:

• Patterns explore independent dimensions
• No dependency between pattern outputs
• Need ensemble confidence (run same problem through 2 patterns)
• Time allows and you want cross-validation
• Patterns are complementary (e.g., different perspectives on same problem)

Common Chains (Sequential):

BoT → ToT → AR     (explore → optimize → validate)
HE → SRC → AR      (diagnose → trace → validate fix)
AT → DR → ToT      (analogize → synthesize → optimize)
RTR → HE → ToT     (triage → root cause → proper fix)
NDF → ToT → AR     (align stakeholders → optimize → validate)
DR → NDF → ToT     (resolve concepts → negotiate politics → optimize)
AT → BoT → ToT     (find analogies → explore adaptations → select best)

Common Parallel Combinations:

BoT || AT          (parallel exploration from different angles)
ToT || DR          (optimize while synthesizing trade-offs)
HE branches        (test multiple hypotheses simultaneously)
AR attack vectors  (stress-test from multiple directions at once)

Full Reasoning Chain: Input → Output

Complete end-to-end flow for IR-v2 orchestration:

1. **Problem Input**
   └─→ Receive problem statement, constraints, and context

2. **Score 11 Dimensions**
   └─→ Sequential Dependencies, Criteria Clarity, Solution Space Known,
       Single Answer Needed, Evidence Available, Opposing Valid Views,
       Problem Novelty, Robustness Required, Solution Exists,
       Time Pressure, Stakeholder Complexity

3. **Fast-Path Check**
   └─→ TimePressure = 5? → RTR immediately (skip remaining steps)

4. **Pattern Selection**
   └─→ IR-v2 calculates affinity scores for all 9 patterns
   └─→ Apply validation rules (AR needs solution, NDF needs stakeholders)

5. **Direct Analysis Check**
   └─→ All patterns < 2.5? → Use Direct Analysis (no framework)

6. **Orchestration Decision**
   └─→ Single pattern > 4.0? → Use that pattern directly
   └─→ Top 2 within 0.5? → Consider multi-pattern orchestration
   └─→ Top 3 within 0.3? → Apply uncertainty propagation first
   └─→ Decide: Sequential or Parallel execution?

7. **Pattern Execution**
   └─→ Apply selected pattern(s) using their methodologies
   └─→ Track confidence and key findings

8. **Checkpoint (15 min)**
   └─→ Progress check, pattern fit check, new information check
   └─→ Re-evaluate and potentially switch patterns if needed

9. **Handover (if multi-pattern)**
   └─→ Transfer context via .reasoning/ directory
   └─→ Preserve insights from completed pattern
   └─→ Set up next pattern with full context

10. **Synthesis (if multi-pattern)**
    └─→ Combine findings from all patterns
    └─→ Resolve conflicts, identify agreements

11. **Confidence Aggregation**
    └─→ Single pattern: use internal confidence
    └─→ Multi-pattern: apply agreement analysis
    └─→ Apply uncertainty discounts if applicable

12. **Output**
    └─→ Deliver answer with reasoning trail
    └─→ Document confidence level and key uncertainties
    └─→ Provide actionable recommendations

Multi-Pattern Orchestration

When to Orchestrate

• Top 2 patterns within 0.5 points AND
• Problem is high-stakes (consequences matter) AND
• Time budget allows (>45 minutes available)

Orchestration Patterns

Sequential Orchestration (most common):

1. Use exploration pattern first (BoT, AT)
2. Use optimization pattern second (ToT, HE)
3. Use validation pattern last (AR, SRC)

Parallel Orchestration (when patterns are complementary):

• Run 2 patterns independently
• Compare conclusions
• Use agreement/disagreement to calibrate confidence

Nested Orchestration (when patterns address different aspects):

• Apply different patterns to different sub-problems
• Synthesize at the end

Orchestration Decision Table

| NDF | ToT | Sequential: NDF (get buy-in) → ToT (optimize within agreed bounds) | | NDF | DR | Sequential: DR (resolve conceptual tension) → NDF (negotiate stakeholders) | | BoT | NDF | Sequential: BoT (explore options) → NDF (negotiate which to pursue) |

RTR Orchestration Rules:

• RTR is typically a STARTING pattern, not an ending one
• After RTR stabilizes situation, follow up with deeper analysis
• RTR → HE for incident root cause analysis
• RTR → ToT for revisiting decision with more time

NDF Orchestration Rules:

• NDF typically FOLLOWS technical analysis (know options before negotiating)
• BoT → NDF: Explore space, then negotiate which options to pursue
• NDF → ToT: After stakeholder agreement, optimize implementation
• DR → NDF: Resolve conceptual tensions first, then stakeholder tensions

Parallel Execution Integration

When to Parallelize

Parallel execution is appropriate when independent reasoning paths can run concurrently without blocking each other:

Parallel Orchestration Patterns

Parallel Configuration

parallel_config:
  max_concurrent_patterns: 3      # Max patterns running simultaneously
  max_concurrent_branches: 8      # Max branches within a single pattern
  merge_strategy: "consensus"     # "consensus" | "voting" | "aggregation" | "best-of-n"
  timeout_per_branch_ms: 60000    # 60 second timeout per branch
  early_termination_threshold: 0.95  # Stop early if confidence exceeds this

Configuration Guidelines:

• Use max_concurrent_patterns: 2 for typical orchestration
• Use max_concurrent_branches: 8 for BoT exploration
• Increase timeout_per_branch_ms for complex sub-problems
• Lower early_termination_threshold (e.g., 0.85) when speed matters more than certainty

Merge Strategies

Merge Strategy Selection:

If robustness critical → "consensus"
If options are mutually exclusive → "voting"
If findings are additive → "aggregation"
If racing for speed → "best-of-n"

Integration with.reasoning/ Protocol

Parallel execution integrates with the .reasoning/ handover protocol:

.reasoning/
├── current-context.md         # Master context (shared by all branches)
├── parallel-session/
│   ├── config.yaml            # Parallel execution configuration
│   ├── branch-001/
│   │   ├── approach.md        # Pattern being applied
│   │   ├── findings.md        # Intermediate findings
│   │   └── confidence.json    # Branch confidence score
│   ├── branch-002/
│   │   ├── approach.md
│   │   ├── findings.md
│   │   └── confidence.json
│   └── branch-N/
│       └── ...
├── merge-result.md            # Synthesized output from all branches
└── handover.md                # Final handover (captures all branch insights)

Protocol Rules:

1. Each parallel branch writes to its own branch-{id}/ directory
2. Branches read shared context but do NOT write to shared files
3. Merge phase reads all branches, applies merge strategy
4. Handover document captures insights from ALL branches (not just winner)
5. Failed branches are preserved for debugging (marked with status: failed)

Branch Handover Template:

## Branch {id} Summary
- **Pattern Applied**: [pattern name]
- **Conclusion**: [finding]
- **Confidence**: [X]%
- **Key Insights**: [unique contributions]
- **Disagreements**: [where this branch diverged from others]

Feedback Loop: 15-Minute Checkpoint

After 15 minutes of applying selected pattern:

## Checkpoint Evaluation

### Progress Check
- [ ] Have I made meaningful progress toward goal?
- [ ] Is my confidence increasing?

### Pattern Fit Check
- [ ] Am I fighting the methodology?
- [ ] Have I discovered new problem characteristics?

### New Information
- [ ] Has the problem changed?
- [ ] Do my characteristic scores need updating?

### Decision
If 2+ checks FAIL:
  → PAUSE: Re-score characteristics
  → If different pattern scores highest: SWITCH
  → If same pattern: Continue with awareness

If all checks PASS:
  → Continue current pattern
  → Set next checkpoint at 30 min mark

Confidence Aggregation (Fixed)

V1 Problem: Additive confidence boosting was statistically invalid.

V2 Approach: Agreement-based bounded adjustment.

Single Pattern Confidence

Use the pattern's internal confidence score (per its methodology).

Multi-Pattern Confidence

## Multi-Pattern Synthesis

### Raw Scores
- Pattern A conclusion: [Answer A] at [X]% confidence
- Pattern B conclusion: [Answer B] at [Y]% confidence
- Pattern C conclusion: [Answer C] at [Z]% (if used)

### Agreement Analysis

**FULL AGREEMENT** (same conclusion):
- Final Confidence = MIN(MAX(X, Y, Z) + 5%, 95%)
- Rationale: Independent paths converging increases trust

**Confidence Bounds**:
- Floor: 10% (never report lower confidence)
- Ceiling: 95% (never report higher confidence)

**PARTIAL AGREEMENT** (2/3 agree):
- Final Confidence = (AVG of agreeing × 0.7) + (disagreeing × 0.15)
- Must document the disagreement
- Consider: Why does one pattern disagree?

**NO AGREEMENT** (different conclusions):
- Final Confidence = MIN(X, Y, Z) - 10%
- This is a FEATURE not a bug - disagreement reveals complexity
- Action: Either (a) gather more information, or (b) present trade-offs to stakeholder

### Shared Assumption Discount
If patterns share significant assumptions, apply -5% adjustment.
(Same LLM, same problem framing, same information = shared blind spots)

Pattern Limitations Reference

Tree of Thoughts (ToT):

• Requires clear evaluation criteria
• Deep recursion may overfit to evaluation function
• Fixed branching can force artificial distinctions

Breadth of Thought (BoT):

• Cannot truly be "exhaustive"
• 8-10 branches may not cover solution space
• Returns multiple options requiring further decision

Self-Reflecting Chain (SRC):

• Limited by weakest step in chain
• Backtracking is costly
• Assumes linear dependency structure

Hypothesis-Elimination (HE):

• Requires discriminating evidence
• Can only find causes in the hypothesis set
• Time-sensitive (may not suit exploration)

Adversarial Reasoning (AR):

• Requires existing solution to attack
• Can be demoralizing if overused
• May miss non-adversarial failure modes

Dialectical Reasoning (DR):

• Requires genuinely opposing valid views
• Synthesis isn't always possible
• Can be slower than just deciding

Analogical Transfer (AT):

• Analogy quality varies widely
• Source domain may mislead
• Requires creativity in finding parallels

Rapid Triage Reasoning (RTR):

• Sacrifices depth for speed
• May miss optimal solution (accepts "good enough")
• Requires follow-up analysis for important decisions
• Not suitable when time is actually available

Negotiated Decision Framework (NDF):

• Requires multiple genuine stakeholders
• Time-intensive (relationship building takes time)
• May produce suboptimal technical solutions for political acceptance
• Doesn't help when one party has absolute authority

Quick Selection Guide

"I need to find the BEST option among known choices"
  → Tree of Thoughts

"I need to explore ALL possible approaches"
  → Breadth of Thought

"I need to trace through a logical chain step by step"
  → Self-Reflecting Chain

"I need to find THE CAUSE of something"
  → Hypothesis-Elimination

"I need to VALIDATE a solution before committing"
  → Adversarial Reasoning

"I'm stuck between two valid but opposing approaches"
  → Dialectical Reasoning

"This problem is novel - no one has solved it in my domain"
  → Analogical Transfer

"I need to decide RIGHT NOW (minutes, not hours)"
  → Rapid Triage Reasoning

"Multiple stakeholders with competing interests must agree"
  → Negotiated Decision Framework

"I'm not sure which to use"
  → Score the dimensions (Step 1)

--- Parallelism Quick-Reference ---

"Top 2 patterns scored within 0.3"
  → Run both in parallel, merge with "consensus" or "voting"

"Need to explore many options fast"
  → Use BoT with max_concurrent_branches: 8

"Testing multiple hypotheses"
  → HE with parallel evidence gathering

"Need diverse perspectives on same problem"
  → MoA pattern: parallel expert personas

"Running parallel but need to merge"
  → consensus (high-stakes) | voting (exclusive) | aggregation (additive) | best-of-n (speed)

Example Application

Problem: "Design our company's approach to AI governance"

Characteristic Scoring

Pattern Affinity Scores

• ToT: (3×.35)+(4×.30)+(2×.20)+(1×.15) = 2.80
• BoT: (4×.35)+(2×.30)+(3×.20)+(5×.15) = 3.35
• SRC: (2×.45)+(3×.25)+(4×.20)+(1×.10) = 2.55
• HE: (2×.40)+(4×.30)+(1×.20)+(1×.10) = 2.30
• AR: 0 (SolutionExists=1 < 3, nothing to attack yet)
• DR: (5×.50)+(3×.20)+(4×.15)+(4×.15) = 4.30
• AT: (5×.45)+(4×.30)+(4×.15)+(4×.10) = 4.45 ← Highest
• RTR: (2×.50)+(4×.25)+(2×.15)+(1×.10) = 2.40

Recommendation

Primary: Analogical Transfer (4.45) - Look at how other governance challenges were solved Secondary: Dialectical Reasoning (3.80) - Innovation vs caution tension needs synthesis

Orchestration: AT → DR → AR

1. Use AT to find analogous governance frameworks (environmental, financial, medical)
2. Use DR to synthesize the innovation/caution tension
3. Use AR to stress-test the proposed governance approach

Version History

V2.1 (Current):

• Added RTR (Rapid Triage Reasoning) for time-critical decisions
• Added NDF (Negotiated Decision Framework) for multi-stakeholder coordination
• Total: 9 reasoning patterns (up from 7 in V2.0)
• Added 3 new dimensions: SolutionExists, TimePressure, StakeholderComplexity
• Fixed AR formula: now requires SolutionExists ≥ 3
• Fixed DR formula: SingleAnswer no longer penalized when OpposingViews high
• Added uncertainty propagation for close pattern scores
• Added Time Pressure fast-path (auto-selects RTR when TimePressure=5)
• Added NDF validation (requires StakeholderComplexity ≥ 3)
• Enhanced orchestration table with RTR and NDF combinations

V2.0:

• Added 4 new patterns: HE, AR, DR, AT
• Replaced decision tree with weighted multi-dimensional scoring
• Added 15-minute feedback checkpoint
• Fixed confidence aggregation (no more invalid additive boosting)
• Added orchestration decision table
• Added pattern limitations reference

V1.0 (Deprecated):

• 3 patterns: ToT, BoT, SRC
• Order-dependent decision tree
• No feedback loop
• Invalid confidence aggregation