integrated-reasoning

# Integrated Reasoning — 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 | Purpose | Best For | |---------|---------|----------| | **Tree of Thoughts (ToT)** | Find optimal solution through deep exploration | Optimization, clear criteria, find THE best | | **Breadth of Thought (BoT)** | Map solution space comprehensively | Unknown space, need multiple options | | **Self-Reflecting Chain (SRC)** | Sequential reasoning with validation | Dependent steps, proofs, linear traces | | **Hypothesis-Elimination (HE)** | Systematic elimination through evidence | Diagnosis, debugging, root cause | | **Adversarial Reasoning (AR)** | Stress-test through attack simulation | Validation, security, pre-mortems | | **Dialectical Reasoning (DR)** | Synthesize opposing valid perspectives | Trade-offs, conceptual conflicts | | **Analogical Transfer (AT)** | Solve via cross-domain parallels | Novel problems, no direct precedent | | **Rapid Triage Reasoning (RTR)** | Fast decisions under time pressure | Incidents, emergencies, time-boxed choices | | **Negotiated Decision Framework (NDF)** | Multi-stakeholder coordination | Politics, competing interests, buy-in needed | --- ## Pattern Selection: Weighted Multi-Dimensional Scoring **Turn discipline (all models, critical for small ones):** announcing this skill or a selected pattern is NOT a response. After any announcement, continue immediately in the same turn with the pattern's actual steps and the tool calls that do the work. (Measured failure mode: a 7.5B local model announced the skill and ended its turn — three times in five runs.) **The scoring pass is silent.** Dimension scores, affinity arithmetic, pattern names, and selection narration never appear in the user-facing answer — the reader sees the analysis, not the apparatus that chose it. A selected pattern's structure may appear only as content built from the problem's own facts (a hypothesis table, a criteria matrix), never as methodology narration. Announce nothing; just analyze. ### Step 1: Assess Problem Characteristics Score each dimension (1-5): ```markdown | 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: ```markdown ## 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 | Scenario | Action | |----------|--------| | One pattern scores >4.0 | Use that pattern directly | | Top 2 within 0.5 of each other | Consider multi-pattern orchestration | | Top 3 within 0.3 of each other | Apply uncertainty propagation first | | All patterns <3.0 | Problem may need decomposition first | | Top pattern <2.5 | None fit well; use Direct Analysis | ### Direct Analysis (10th Pattern) When all pattern scores are below 2.5, no specialized methodology is warranted: ```markdown ## 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. --- ## Effort Level Mapping After selecting the optimal pattern, set effort level: | Pattern | Default Effort | Max Effort Trigger | |---------|---------------|-------------------| | ToT | high | Critical optimization decisions | | BoT | high | Unknown solution spaces | | HE | high | Production incidents | | SRC | medium | Standard step-by-step | | AR | high | Security/safety validation | | DR | high | Architecture trade-offs | | AT | max | Completely novel domains | | RTR | low | Time-critical (< 5 minutes) | | NDF | medium | Standard stakeholder alignment | For max effort, include "ultrathink" in the analysis prompt to enable extended thinking budget. --- ## Execution Modes Patterns with `context: fork` in their frontmatter will automatically run in isolated subagents when invoked via skills. This means: - They get their own context window (won't bloat main conversation) - They can't see conversation history (provide full context in the prompt) - Results are returned as a summary For inline execution (when you need conversation context), invoke the pattern's methodology directly without using the skill. **Patterns that fork**: ToT, BoT, HE, AR, SRC, DR, AT **Patterns that run inline**: RTR (speed), NDF (needs conversation context for stakeholder interaction) **IR-v2 itself**: Always inline (orchestrates within main context) --- ## Model-Aware Selection When running on different models: - **Opus**: All patterns available at full depth - **Sonnet**: Reduce branch counts (ToT: 3 branches, BoT: 5 approaches). SRC, HE, RTR work well. DR and AT may lose synthesis quality. - **Haiku**: Only RTR recommended; other patterns lose quality --- ## 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 | Pattern A High | Pattern B High | Orchestration | |----------------|----------------|---------------| | BoT | ToT | Sequential: BoT (explore) → ToT (optimize top options) | | BoT | HE | Sequential: BoT (generate hypotheses) → HE (eliminate) | | ToT | AR | Sequential: ToT (select) → AR (validate before commit) | | ToT | SRC | Sequential: ToT (decide) → SRC (plan implementation) | | DR | ToT | Sequential: DR (resolve tension) → ToT (optimize within synthesis) | | AT | ToT | Sequential: AT (find analogies) → ToT (evaluate derived solutions) | | AT | BoT | Parallel: Both explore, merge findings | | HE | SRC | Sequential: HE (find cause) → SRC (trace mechanism) | | RTR | HE | Sequential: RTR (immediate triage) → HE (post-incident RCA) | | RTR | AR | Sequential: RTR (quick decision) → AR (post-decision validation) | | 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 Pattern Execution When top 2 patterns score within 0.3: - Use built-in parallel subagent execution (multiple Task tool calls) - Each pattern runs in its own forked context - Merge results using agreement analysis ### When to Parallelize Parallel execution is appropriate when independent reasoning paths can run concurrently without blocking each other: | Condition | Parallelization Strategy | |-----------|-------------------------| | **Top 2 patterns within 0.3 of each other** | Run both patterns in parallel, compare results | | **BoT natural parallelism** | 8-10 branches can explore simultaneously |