Humanization Verifier

Agent ID: F5 Category: F - Quality VS Level: Low (Verification-focused) Tier: Support Icon: ✅ Model Tier: LOW (Haiku)

Overview

Verifies that humanization transformations maintain academic integrity, preserve critical elements, and actually reduce AI detectability. This is the final quality gate before humanized content is exported.

Core Functions

1. Citation Integrity Check

Verifies all citations remain intact and accurate:

yaml

checks:
  - count_match: "Same number of citations before/after"
  - format_preserved: "(Author, year) format maintained"
  - content_accurate: "No citation content changed"
  - placement_logical: "Citations still support correct claims"

output:
  citations_before: 15
  citations_after: 15
  all_intact: true
  issues: []

2. Statistical Accuracy Check

Ensures all numerical values are unchanged:

yaml

checks:
  - p_values: "All p-values identical"
  - effect_sizes: "All d, r, η² unchanged"
  - sample_sizes: "All N, n unchanged"
  - test_statistics: "All t, F, χ² unchanged"
  - percentages: "All % unchanged"
  - confidence_intervals: "All CI unchanged"

output:
  statistics_before: 23
  statistics_after: 23
  all_intact: true
  issues: []

3. Meaning Preservation Check

Validates core claims are unchanged:

yaml

checks:
  - main_findings: "Key findings preserved"
  - conclusions: "Conclusions unchanged"
  - methodology_claims: "Method descriptions accurate"
  - causal_claims: "No new causal language added"
  - hedge_appropriateness: "Hedging matches evidence"

assessment:
  meaning_preserved: true
  confidence: 95%
  flagged_changes: []

4. AI Pattern Reduction Check

Re-runs G5 analysis on transformed text:

yaml

comparison:
  original:
    ai_probability: 67%
    patterns_detected: 18
    high_risk: 5
    medium_risk: 9
    low_risk: 4

  humanized:
    ai_probability: 28%
    patterns_detected: 4
    high_risk: 0
    medium_risk: 2
    low_risk: 2

  reduction:
    probability_drop: 39%
    pattern_reduction: 78%
    effective: true

5. Academic Tone Check

Ensures scholarly voice is maintained:

yaml

checks:
  - formality: "Appropriate formality level"
  - objectivity: "Objective tone preserved"
  - precision: "Technical precision maintained"
  - consistency: "Consistent voice throughout"

assessment:
  tone_appropriate: true
  issues: []

Input Requirements

yaml

Required:
  - original_text: "Text before humanization"
  - humanized_text: "Text after G6 transformation"
  - transformation_log: "G6 change record"

Optional:
  - section_type: "abstract/methods/discussion/etc."
  - strictness: "low/medium/high"

Output Format

markdown

## Humanization Verification Report

### Summary

| Check | Status | Details |
|-------|--------|---------|
| Citation Integrity | ✅ Pass | 15/15 citations preserved |
| Statistical Accuracy | ✅ Pass | 23/23 values unchanged |
| Meaning Preservation | ✅ Pass | Core claims intact |
| AI Pattern Reduction | ✅ Pass | 67% → 28% (-39%) |
| Academic Tone | ✅ Pass | Scholarly voice maintained |

### Overall Verdict: ✅ APPROVED

---

### Detailed Results

#### Citation Integrity

✅ All 15 citations preserved ✅ Format consistent ✅ Placement logical

code


#### Statistical Accuracy

✅ All p-values unchanged ✅ All effect sizes unchanged ✅ All sample sizes unchanged ✅ All test statistics unchanged

code


#### Meaning Preservation

✅ Main findings preserved ✅ Conclusions unchanged ✅ No meaning distortion detected Confidence: 95%

code


#### AI Pattern Reduction

Before: 67% AI probability (18 patterns) After: 28% AI probability (4 patterns) Reduction: 39% probability, 78% patterns

code


#### Academic Tone

✅ Formal register maintained ✅ Objective voice preserved ✅ Technical precision intact

code


---

### Flagged Items (Review Recommended)

None

---

### 🟡 CHECKPOINT: CP_HUMANIZATION_VERIFY (Optional)

Verification complete. Ready for export?

[A] Approve and export
[B] Review specific changes
[C] Revert to original

Verification Strictness Levels

Low Strictness

•Basic citation count match
•Statistical value presence check
•Major meaning changes only

Medium Strictness (Default)

•Full citation verification
•Statistical value comparison
•Meaning preservation analysis
•AI reduction check

High Strictness

•Deep citation context analysis
•Statistical formatting verification
•Sentence-level meaning comparison
•Tone consistency analysis
•Academic style guide compliance

Error Handling

Critical Failures (Auto-Reject)

yaml

critical_failures:
  - citation_missing: "Any citation removed"
  - citation_altered: "Citation content changed"
  - statistic_modified: "Any number changed"
  - meaning_reversed: "Claim direction changed"

action: "REJECT transformation, revert to original"

Warnings (Flag for Review)

yaml

warnings:
  - hedge_changed: "Hedging level modified"
  - emphasis_shifted: "Emphasis moved"
  - structure_altered: "Sentence structure significantly changed"
  - ai_reduction_low: "Less than 20% AI probability reduction"

action: "FLAG for user review"

Acceptable Changes

yaml

acceptable:
  - vocabulary_substitution: "AI words replaced"
  - phrase_simplification: "Verbose phrases shortened"
  - punctuation_normalization: "Em dashes, quotes normalized"
  - transition_variation: "Transition words varied"

Prompt Template

code

You are an academic writing quality verifier.

Compare the original and humanized texts to verify transformation quality:

[Original Text]: {original}
[Humanized Text]: {humanized}
[Transformation Log]: {log}
[Section Type]: {section_type}

Perform the following verification checks:

1. **Citation Integrity**
   - Count citations in both versions
   - Verify each citation is preserved
   - Check format consistency
   - Confirm logical placement

2. **Statistical Accuracy**
   - Extract all numerical values from original
   - Verify identical values in humanized
   - Flag any discrepancies

3. **Meaning Preservation**
   - Identify main claims in original
   - Verify claims preserved in humanized
   - Check for unintended meaning changes
   - Assess hedge appropriateness

4. **AI Pattern Reduction**
   - Estimate AI probability (before/after)
   - Count remaining patterns
   - Calculate reduction percentage

5. **Academic Tone**
   - Assess formality level
   - Check objectivity
   - Verify consistency

Output a verification report with:
- Pass/Fail status for each check
- Specific issues found
- Overall recommendation (Approve/Review/Reject)

Integration with Pipeline

code

G6-AcademicStyleHumanizer
        │
        ▼
┌───────────────────────────────────────┐
│  F5-HumanizationVerifier (THIS AGENT) │
│  ┌─────────────────────────────────┐  │
│  │ 1. Citation Integrity    ✅/❌  │  │
│  │ 2. Statistical Accuracy  ✅/❌  │  │
│  │ 3. Meaning Preservation  ✅/❌  │  │
│  │ 4. AI Pattern Reduction  ✅/❌  │  │
│  │ 5. Academic Tone         ✅/❌  │  │
│  └─────────────────────────────────┘  │
│                 │                     │
│                 ▼                     │
│         All Pass? ──Yes──> Export    │
│              │                        │
│             No                        │
│              │                        │
│              ▼                        │
│     🟡 CP_HUMANIZATION_VERIFY        │
│     User review required              │
└───────────────────────────────────────┘

Commands

code

"Verify humanization"
→ Full verification report

"Quick verification check"
→ Summary only

"Check citation integrity"
→ Citation-specific check

"Verify statistics preserved"
→ Statistical-specific check

"Compare meaning before/after"
→ Meaning preservation check

Related Agents

•G5-AcademicStyleAuditor: Provides AI pattern analysis
•G6-AcademicStyleHumanizer: Creates text this agent verifies
•F1-InternalConsistencyChecker: Related quality checks
•F4-BiasTrustworthinessDetector: Related integrity checks

References

•Humanization Pipeline: ../../research-coordinator/core/humanization-pipeline.md
•G5 Agent: ../G5-academic-style-auditor/SKILL.md
•G6 Agent: ../G6-academic-style-humanizer/SKILL.md
•User Checkpoints: ../../research-coordinator/interaction/user-checkpoints.md