Skip to content

Latest commit

 

History

History
116 lines (87 loc) · 5.78 KB

File metadata and controls

116 lines (87 loc) · 5.78 KB

🔄 Self-Consistency Checking

Category: Intermediate | Difficulty: Intermediate | Impact: Very High

When To Use

Use self-consistency when the task has a single correct answer (or a narrow range of acceptable answers) but the model's reasoning path is unreliable. This pattern is essential for high-stakes decisions where a wrong answer has real consequences — financial analysis, medical triage, legal review, or candidate evaluation.

The Problem

❌ Naive Prompt
Based on this restaurant's financial data, should the owner open a
second location? Give me a yes or no recommendation with reasoning.

What goes wrong: Run this 5 times and you'll get 3 "yes" and 2 "no" answers, each with equally convincing reasoning. The model constructs a plausible narrative to support whatever direction it drifts toward on a given run. You have no way to know which answer to trust.

Why it fails: Complex analytical questions involve multiple reasoning paths, and the model samples one path per generation. That path might emphasize revenue growth (→ yes) or might emphasize debt ratios (→ no). A single run gives you one perspective disguised as a comprehensive analysis.

The Pattern

✅ Engineered Prompt
Analyze whether this restaurant should open a second location.
Provide THREE independent analyses, each starting from a different
analytical framework:

ANALYSIS 1 — FINANCIAL LENS:
Evaluate based purely on financial metrics: cash flow, debt-to-
equity, current margins, capital requirements, and break-even
timeline for a second location.
Recommendation: [YES/NO] with confidence [HIGH/MEDIUM/LOW]

ANALYSIS 2 — OPERATIONAL LENS:
Evaluate based on operational readiness: management depth, systems
scalability, supply chain capacity, and whether the first location
runs without owner presence.
Recommendation: [YES/NO] with confidence [HIGH/MEDIUM/LOW]

ANALYSIS 3 — MARKET LENS:
Evaluate based on market factors: local competition, demand
indicators, real estate availability, and demographic trends
in the target area.
Recommendation: [YES/NO] with confidence [HIGH/MEDIUM/LOW]

SYNTHESIS:
- How many of the 3 analyses agree? If all 3 agree, high confidence.
  If 2/3, moderate confidence. If split, the decision is genuinely
  ambiguous.
- What is the single biggest risk that could override the majority
  recommendation?
- Final recommendation with honest confidence assessment.

Financial data:
[restaurant financial data]

Why It Works

Self-consistency exploits the fact that correct answers tend to be reachable from multiple reasoning paths, while incorrect answers typically depend on a specific (flawed) reasoning chain. By forcing the model to approach the same question from different angles, you're effectively running a vote across reasoning strategies.

This pattern reduces the variance in model outputs. A single analysis might be swayed by whichever data point the model attends to first. Three independent analyses, each constrained to a different framework, ensure comprehensive coverage of the relevant factors.

The synthesis step is where the real value emerges. When the analyses disagree, that disagreement itself is information — it tells you the decision is genuinely close and depends on which factors you weight more heavily. This is vastly more useful than a false-confidence single recommendation.

Real-World Example

Scenario: Evaluating whether a job candidate should accept a counter-offer from their current employer.

My career coaching client received a counter-offer. Help me advise
them by running three independent analyses:

ANALYSIS 1 — COMPENSATION ANALYSIS:
Compare total comp (base + bonus + equity + benefits) between
the new offer and counter-offer over a 3-year horizon. Account
for vesting schedules and promotion trajectories.
Verdict: [TAKE COUNTER / LEAVE] — Confidence: [H/M/L]

ANALYSIS 2 — CAREER TRAJECTORY ANALYSIS:
Evaluate which role offers better skill development, title
progression, network expansion, and resume positioning for
their 5-year goals.
Verdict: [TAKE COUNTER / LEAVE] — Confidence: [H/M/L]

ANALYSIS 3 — RISK ANALYSIS:
Assess the counter-offer risk (data shows 50-80% of counter-offer
acceptors leave within 18 months), the new role risk (startup
stage, funding runway), and the relationship dynamics at current
company post-counter-offer.
Verdict: [TAKE COUNTER / LEAVE] — Confidence: [H/M/L]

SYNTHESIS:
Provide a final recommendation noting where the analyses agree
and where they conflict. Flag the single factor that should be
the tiebreaker.

Client situation: [details]

Common Mistakes

  1. Making the analyses too similar. If all three analyses use the same reasoning approach with slightly different framing, you haven't achieved true independence. Force genuinely different analytical frameworks.

  2. Ignoring disagreement. When analyses conflict, the temptation is to go with the majority. But a 2-1 split where the dissenting analysis has high confidence should be treated as a yellow flag, not dismissed.

Warning

Skipping the synthesis step. Three independent analyses without synthesis is just three opinions. The synthesis — especially when analyses disagree — is where you extract signal from noise. Always force the model to reconcile its own conflicting conclusions.

Related Patterns

  • Chain-of-Thought — Each analysis within the self-consistency check benefits from chain-of-thought reasoning.
  • Chain of Verification — After self-consistency identifies a recommendation, verification checks the facts behind it.
  • Evaluation-Driven Iteration — Use systematic evaluation to measure whether self-consistency actually improves your accuracy.