📊 Selection Intelligence

Fuzzy Scoring & Selection

Not just pass/fail — candidates are scored on multiple dimensions using fuzzy predicate logic, then the best answer is selected.

Foundations

What Is Fuzzy Logic?

Classical logic is binary: something is either true or false. Fuzzy logic recognises that truth has degrees. An answer can be 0.85 correct — mostly right, but with some uncertainty.

❌ Binary Logic

Is this answer correct?
→ Yes or No

Problem: What if two candidates are both "correct" but one is better? Binary logic can't distinguish them.

✅ Fuzzy Logic

How correct is this answer?
→ 0.92 correctness
→ 0.87 confidence
→ 0.95 coherence

Richer signal. Better selection.

Fuzzy logic was pioneered by Lotfi Zadeh in 1965 and is widely used in control systems, decision-making, and AI. Mikoshi AI Turbo applies it to candidate selection — scoring each answer on multiple dimensions rather than reducing everything to pass/fail.

Dimensions

The Three Fuzzy Predicates

Every candidate is scored on three dimensions, each ranging from 0.0 (none) to 1.0 (perfect):

✅

Correctness

0.0 — 1.0

Did the verification code pass? How accurately does the answer match the expected result?

💪

Confidence

0.0 — 1.0

How certain is the model about this answer? Measured by token probabilities and self-assessment.

🔗

Coherence

0.0 — 1.0

Is the reasoning internally consistent? Does the explanation match the conclusion?

// Fuzzy predicate scoring
function scoreCandidate(candidate, verificationResult) {
  return {
    correctness: verificationResult.passed ? 1.0 : 0.0,
    confidence:  candidate.selfAssessedConfidence || 0.5,
    coherence:   assessCoherence(candidate.reasoning, candidate.answer),

    // Composite score (weighted average)
    get total() {
      return (this.correctness * 0.5)   // 50% weight
           + (this.confidence * 0.3)   // 30% weight
           + (this.coherence * 0.2);  // 20% weight
    }
  };
}
      

Visual

Scoring Matrix

Here's how three candidates might score for the question "What is the square root of 144?":

Candidate A

Correct

1.00

Confidence

0.95

Coherence

0.90

0.965

⭐ Selected

Candidate B

Correct

1.00

Confidence

0.70

Coherence

0.80

0.870

✅ Passed

Candidate C

Correct

0.00

Confidence

0.85

Coherence

0.60

0.375

❌ Failed

Notice: Candidate C has high confidence but failed verification. In a standard AI system, you'd get this confidently wrong answer. With fuzzy scoring, correctness dominates the score, and the wrong answer is filtered out.

Algorithm

The Selection Algorithm

Once all candidates are scored, the selection process follows strict priority rules:

// Selection algorithm
function selectBest(candidates) {
  // 1. Partition: passed vs failed
  const passed = candidates.filter(c => c.verification.passed);
  const failed = candidates.filter(c => !c.verification.passed);

  // 2. Passed candidates ALWAYS win over failed
  if (passed.length > 0) {
    // 3. Among passed, select highest fuzzy score
    return passed.sort((a, b) => b.score.total - a.score.total)[0];
  }

  // 4. No candidates passed — fall back to highest score
  // Mark result as "unverified"
  return {
    ...failed.sort((a, b) => b.score.total - a.score.total)[0],
    verified: false,
    warning: 'No candidates passed verification'
  };
}
      

All pass: Best fuzzy score wins → high confidence result
Some pass: Only passed candidates considered → verified result
None pass: Best score from failed candidates, marked as unverified → honest uncertainty

Edge Cases

Handling the Tricky Situations

🤔 What if no candidates pass?

Turbo still returns the best answer available, but explicitly marks it as unverified. This is honest about uncertainty — the user knows the answer couldn't be independently confirmed. This is vastly better than standard AI, which would give the wrong answer with false confidence.

🏆 What if all candidates pass with equal scores?

Ties are broken by: (1) verification execution time (faster = more efficient reasoning), (2) answer length (shorter = more concise), (3) candidate index (lower temperature = more deterministic). In practice, fuzzy scoring almost always produces distinct scores — ties are extremely rare.

🔀 What if candidates disagree but all pass?

This is actually the most interesting case — it means the verification code for different answers all returned true. This can happen with subjective questions. In these cases, coherence scoring becomes the deciding factor, favouring the answer with the most internally consistent reasoning.

Watch fuzzy scoring in real time

⚡ Try Turbo in Synapse

← Code Verification Graph Architecture →