ai-evaluation-verification

name: ai-evaluation-verification description: "Retain epistemic control by validating AI outputs against logic, first principles, and reality. Detect confident nonsense, false completeness, and category errors." version: "1.0.0"

Overview

AI Evaluation & Verification is Layer 5 of AI fluency—the ability to maintain epistemic control over AI outputs. AI fluency requires governance, not trust.

Core Principle: Treat AI output as a drafted hypothesis, not an answer.

Fluency Signal: Can reject plausible but incorrect AI output with justification.

When to Use This Skill

• After any AI output you plan to use
• Before sharing AI-generated content
• When AI output "feels" right but isn't verified
• When stakes of error are significant
• When building verification into workflows

Verification Framework

The Three Validation Gates

Every AI output should pass three gates:

Gate 1: Logic Check

Question: Does the reasoning make sense?

Checks:

• Do conclusions follow from premises?
• Are there logical jumps or gaps?
• Is the reasoning circular?
• Are cause-effect claims valid?

Red flags:

• Conclusions stated without supporting reasoning
• "Because it's best practice" without explanation
• Jumps from observation to recommendation

Gate 2: First Principles Check

Question: Does this align with fundamental truths?

Checks:

• Consistent with domain knowledge?
• Violates any known constraints?
• Contradicts established facts?
• Makes physically/mathematically possible claims?

Red flags:

• Claims that seem too good to be true
• Violations of basic domain principles
• Numbers that don't add up

Gate 3: Reality Check

Question: Does this match external evidence?

Checks:

• Verifiable against authoritative sources?
• Consistent with observable reality?
• Can be independently confirmed?

Red flags:

• Specific facts without citations
• Claims about recent events (training cutoff)
• "All experts agree" statements

Error Detection Patterns

1. Confident Nonsense

What it is: Fluently stated information that is factually wrong.

Detection:

• Check specific facts, especially dates, numbers, names
• Verify citations actually exist
• Cross-reference with authoritative sources
• Ask for sources and verify them

Example:

AI: "According to Smith (2022), the algorithm has O(log n) complexity..." Verification: Does Smith (2022) exist? Does it make this claim?

2. False Completeness

What it is: Output presented as comprehensive when it's partial.

Detection:

• Ask: "What did you NOT include?"
• Check if obvious items are missing
• Compare against known complete lists
• Request exhaustiveness check

Example:

AI: "The main programming paradigms are: object-oriented, functional, procedural" Missing: Logic programming, concurrent, aspect-oriented, etc.

3. Category Errors

What it is: Mixing categories or applying concepts incorrectly.

Detection:

• Check if terms are used correctly
• Verify categorizations make sense
• Look for inappropriate analogies
• Check if advice fits the category

Example:

AI: "To improve database performance, refactor your frontend components" Error: Mixing frontend/backend categories

4. Anchoring on False Premises

What it is: Building correct reasoning on user's incorrect assumptions.

Detection:

• Review your input for errors
• Check if AI questioned problematic premises
• Test by deliberately introducing errors

Example:

User: "Explain why Python is faster than C" AI: [Provides explanation for a false premise]

5. Temporal Confusion

What it is: Mixing past, present, future inappropriately.

Detection:

• Verify timelines
• Check if claims about "current state" are current
• Watch for stale information

Example:

AI: "The current CEO of [Company] is X" Reality: X left two years ago

Verification Methods

Method 1: Spot-Check Verification

Sample specific claims for verification:

1. Identify 3-5 verifiable claims
2. Check each against authoritative source
3. If any fail, distrust entire output
4. Calculate trust discount based on failure rate

Method 2: Falsification Testing

Try to prove the output wrong:

1. Ask: "What would make this conclusion false?"
2. Check if any of those conditions exist
3. Look for evidence that contradicts
4. If falsification attempts fail, increase confidence

Method 3: Counter-Prompting

Ask AI to argue against itself:

1. Get initial output
2. Ask: "Now argue the opposite position"
3. Ask: "What's wrong with your first answer?"
4. Synthesize the views

Method 4: Source Verification

For factual claims:

1. Ask AI for sources
2. Verify sources exist
3. Verify sources say what AI claims
4. Check source credibility

Method 5: Consistency Testing

Check for internal contradictions:

1. Ask the same question different ways
2. Ask follow-up questions that would expose inconsistency
3. Compare answers across queries
4. Flag any contradictions for investigation

Stopping Rules

When to Stop Iterating

Iteration should end when:

1. Success criteria met: Output passes all verification gates
2. Diminishing returns: Additional iterations yield marginal improvement
3. Resource limit: Time/cost exceeds value of improvement
4. Fundamental mismatch: Task requires human judgment, not more AI

How to Decide

STOPPING CHECKLIST:
□ Output meets stated success criteria
□ All critical claims verified
□ No logical errors detected
□ Consistent with domain knowledge
□ Additional iteration won't improve key dimensions

If all checked → Stop
If any unchecked → Iterate or escalate to human

Confidence Scoring

Rate Every Output

Document Confidence

OUTPUT CONFIDENCE ASSESSMENT:

Claim: [The AI's output]
Confidence: [High/Medium/Low/Unreliable]

Verification performed:
- [Check 1]: [Result]
- [Check 2]: [Result]
- [Check 3]: [Result]

Remaining concerns:
- [Concern 1]
- [Concern 2]

Recommendation: [Use as-is / Use with modifications / Reject]

Practices

Red-Team Prompts

After getting output, attack it:

Now act as a critical reviewer:
- What errors might be in your previous response?
- What claims are weakest?
- What did you assume that might be wrong?
- How would an expert in this field critique this?

Falsification Tests

Your previous response concluded [X].

What evidence would prove this conclusion wrong?
Does any of that evidence exist?
How confident should I be in this conclusion?

Confidence Scoring Prompts

Rate your confidence in each part of your response:
- [Claim 1]: confidence and why
- [Claim 2]: confidence and why
- [Claim 3]: confidence and why

What would make you more or less confident?

Assessment Criteria

Layer 5 Complete When:

• Applies verification gates to AI outputs by default
• Can detect and articulate specific error types
• Has documented rejected outputs with reasoning
• Uses stopping rules to end iteration appropriately
• Assigns confidence scores and documents basis

Common Verification Failures

Failure 1: Verification Theater

Wrong: "This looks right" (no actual verification) Right: "I verified claim X against source Y, which confirms Z"

Failure 2: Trusting AI's Self-Assessment

Wrong: Asking AI "are you sure?" and trusting "yes" Right: External verification independent of AI

Failure 3: Verification Fatigue

Wrong: Thorough verification on first output, none on subsequent Right: Consistent verification protocol regardless of prior accuracy

Failure 4: Unfalsifiable Acceptance

Wrong: Accepting because you can't prove it wrong Right: Active falsification attempts before acceptance

Related Skills

• ai-system-literacy — Understanding why errors occur
• ai-cognitive-readiness — Skepticism as default stance
• ai-fluency-antipatterns — Output Authority Bias

Learn More

• Verification Checklists
• Error Pattern Examples