aiskillstore Abductive Reasoning
Generate best explanations from observations. The logic of diagnosis and inference to cause.
Type Signature
Abductive : Observation → Hypotheses[] → Evidence → BestExplanation
Where:
Observation : RawData × Surprise → AnomalyDescription
Hypotheses : AnomalyDescription → [PossibleCause]
Evidence : [PossibleCause] × AvailableData → [ScoredHypothesis]
BestExplanation : [ScoredHypothesis] → (Cause × Confidence × NextSteps)
When to Use
Use abductive when:
- Anomaly or surprise present (metric deviation, unexpected behavior)
- Need to explain an observation (why did X happen?)
- Incomplete data, must infer cause
- Diagnostic context (errors, issues, failures)
- Multiple possible explanations exist
Don't use when:
- Cause is already known (use causal execution instead)
- Need to compare decision alternatives
- Resolving stakeholder disagreements
Four-Stage Process
Stage 1: Observation
Purpose: Transform raw data into structured anomaly description.
Input:
observation:
raw_data: "Conversion dropped from 12% to 7% in Q4"
context:
timeframe: "Q4 2025"
baseline: "12% historical average"
current: "7% observed"
surprise_level: 0.8 # How unexpected is this?
Process:
- Quantify the deviation (% change, sigma from mean)
- Identify temporal boundaries (when did it start?)
- Scope the phenomenon (which segments affected?)
- Rate surprise level (expected vs unexpected)
Output:
anomaly:
description: "42% drop in conversion rate (12% → 7%)"
deviation: "-5 percentage points, -42% relative"
temporal: "Started week 3 of Q4, persists through Q4"
scope: "All segments equally affected"
surprise: 0.8
baseline_source: "12-month rolling average"
Stage 2: Hypotheses Generation
Purpose: Generate diverse possible explanations without judgment.
Rules:
- Generate ≥5 hypotheses (more is better initially)
- Include obvious AND non-obvious causes
- Consider internal AND external factors
- Don't filter yet - cast wide net
Hypothesis Categories:
| Category | Examples |
|---|---|
| Technical | Site issues, bugs, performance |
| Product | Features, pricing, positioning |
| Market | Competition, trends, seasonality |
| Operational | Team changes, process issues |
| External | Economy, regulations, events |
Output:
hypotheses:
- id: H1
cause: "Website performance degradation"
category: technical
mechanism: "Slow load times → abandonment"
- id: H2
cause: "Competitor launched aggressive pricing"
category: market
mechanism: "Price undercut → customer diversion"
- id: H3
cause: "Seasonal Q4 shopping behavior change"
category: market
mechanism: "Holiday spending patterns differ"
- id: H4
cause: "Product-market fit weakening"
category: product
mechanism: "Customer needs evolving away"
- id: H5
cause: "Sales qualification criteria changed"
category: operational
mechanism: "Different lead quality entering funnel"
# ... continue until exhaustive
Stage 3: Evidence Evaluation
Purpose: Score each hypothesis against available evidence.
For each hypothesis, evaluate:
| Criterion | Question | Score |
|---|---|---|
| Explanatory power | Does it fully explain the anomaly? | 0-1 |
| Simplicity | Fewest assumptions required? | 0-1 |
| Coherence | Consistent with other known facts? | 0-1 |
| Testability | Can we verify/falsify it? | 0-1 |
| Prior probability | How likely independent of this data? | 0-1 |
Evidence Collection:
evidence:
H1_technical:
supporting:
- "Page load time increased 2s in Q4" (confidence: 0.9)
- "Mobile bounce rate up 15%" (confidence: 0.85)
contradicting:
- "Desktop conversion stable" (confidence: 0.8)
net_score: 0.65
H2_competitor:
supporting:
- "Competitor launched Oct 15" (confidence: 1.0)
- "Google Trends shows competitor interest up" (confidence: 0.7)
contradicting:
- "Our traffic unchanged" (confidence: 0.9)
net_score: 0.55
# ... evaluate all hypotheses
Scoring Formula:
Score(H) = (Explanatory × 0.3) + (Simplicity × 0.2) +
(Coherence × 0.25) + (Testability × 0.1) +
(Prior × 0.15)
Stage 4: Best Explanation
Purpose: Select most probable cause with confidence and next steps.
Ranking:
ranked_hypotheses:
- rank: 1
hypothesis: H1
cause: "Website performance degradation"
score: 0.78
confidence: 0.75
- rank: 2
hypothesis: H3
cause: "Seasonal behavior change"
score: 0.62
confidence: 0.60
- rank: 3
hypothesis: H2
cause: "Competitor pricing"
score: 0.55
confidence: 0.50
Best Explanation Output:
conclusion:
primary_cause: "Website performance degradation"
confidence: 0.75
mechanism: "2s increase in load time caused 42% more abandonment,
consistent with industry benchmarks (1s = ~7% conversion loss)"
contributing_factors:
- "Seasonal patterns may account for 10-15% of drop"
ruled_out:
- "Competitor pricing (traffic unchanged, not price-sensitive segment)"
remaining_uncertainty:
- "Whether mobile-specific or site-wide"
- "Whether fix will fully restore conversion"
next_steps:
- "Verify: Run A/B test with performance fix (high priority)"
- "Measure: Mobile vs desktop split post-fix"
- "Monitor: Competitor activity (low priority)"
suggested_next_mode: causal # Ready to act on diagnosis
Quality Gates
| Gate | Requirement | Failure Action |
|---|---|---|
| Hypothesis count | ≥5 hypotheses | Generate more before proceeding |
| Category diversity | ≥3 categories | Expand hypothesis search |
| Evidence present | ≥1 data point per top-3 | Gather more evidence |
| Confidence threshold | ≥0.6 for best | Flag as inconclusive |
| Testability | Best hypothesis testable | Propose test design |
Common Failure Modes
| Failure | Symptom | Fix |
|---|---|---|
| Anchoring | First hypothesis gets all attention | Force diversity in Stage 2 |
| Confirmation bias | Only seek supporting evidence | Require contradicting evidence |
| Complexity creep | Elaborate explanations preferred | Weight simplicity appropriately |
| Premature closure | Stop at first plausible cause | Complete all 4 stages |
Output Contract
abductive_output:
conclusion:
primary_cause: string
confidence: float # 0.0-1.0
mechanism: string # How cause produces effect
hypotheses:
ranked: [ScoredHypothesis] # All evaluated
ruled_out: [string] # Definitively excluded
evidence:
supporting: [EvidenceItem]
contradicting: [EvidenceItem]
gaps: [string] # What evidence is missing?
uncertainty:
remaining_questions: [string]
confidence_bounds: [float, float] # Low, high
next:
immediate_actions: [string]
tests_to_run: [string]
suggested_mode: optional<ReasoningMode>
trace:
stages_completed: [1, 2, 3, 4]
duration_ms: int
hypotheses_generated: int
evidence_points: int
Example Execution
Context: "Enterprise conversion dropped 40% last quarter"
Stage 1 - Observation:
Anomaly: 40% drop (15% → 9%) in enterprise conversion
Temporal: Started week 5 of Q3, accelerated Q4
Scope: Enterprise only, SMB stable
Surprise: 0.85
Stage 2 - Hypotheses:
H1: Enterprise buyer behavior changed (economic uncertainty)
H2: Sales team restructuring disrupted relationships
H3: Competitor launched enterprise-specific offering
H4: Our enterprise pricing became uncompetitive
H5: Product gaps for enterprise use cases
H6: Longer sales cycles (not drop, just delay)
H7: Key account manager departures
Stage 3 - Evidence:
H1: Supporting (CFO involvement up 40%), Contradicting (overall enterprise IT spend flat)
H2: Supporting (3 senior reps left Q3), Contradicting (coverage maintained)
H6: Supporting (average cycle +45 days), Strong supporting
H7: Supporting (2 key AMs left), Moderate supporting
Stage 4 - Conclusion:
Primary: Sales cycle elongation (not true drop) + AM departures (relationship gaps)
Confidence: 0.72
Mechanism: Economic uncertainty extended CFO approval cycles by 45 days;
AM departures created relationship gaps in 6 key accounts
Next: Wait 45 days to see if "delayed" deals close (causal monitoring)
Immediately backfill AM roles (causal action)