VAP-PAP for Hiring AI: A Technical Framework for Cryptographically Verifiable Resume Screening Audit Trails

1. Introduction: The Auditability Crisis in Hiring AI

1.1 The Scale of AI-Driven Hiring Decisions

An estimated 99% of Fortune 500 companies now use some form of automated screening in their hiring processes. These systems evaluate millions of candidates daily, making consequential decisions that affect individuals' livelihoods, career trajectories, and economic mobility.

Yet these systems operate as black boxes. When a candidate is rejected, they receive at best a generic notification: "After careful consideration, we have decided to move forward with other candidates." The AI's actual reasoning remains opaque.

1.2 The Fundamental Problem

The core issue is not merely opacity but unverifiability. Current hiring AI systems:

1. Do not log decision rationales at the individual candidate level
2. Use mutable storage that permits post-hoc modification
3. Lack cryptographic integrity that makes tampering detectable
4. Cannot demonstrate that the same algorithm was applied consistently
5. Provide no mechanism for independent third-party verification

1.3 The Flight Recorder Paradigm

Aviation safety transformed after regulators mandated flight data recorders—tamper-evident devices that capture every parameter of aircraft operation. AI systems affecting fundamental rights deserve equivalent accountability infrastructure.

The VAP Framework applies this "flight recorder" paradigm to AI decision-making. VAP-PAP specifically addresses public-facing AI decisions, including employment.

2. Regulatory Landscape

2.1 EU AI Act: High-Risk Classification

The EU AI Act (Regulation 2024/1689) explicitly classifies hiring AI as high-risk under Annex III, Point 4(a):

"AI systems intended to be used for the recruitment or selection of natural persons, in particular to place targeted job advertisements, to analyse and filter job applications, and to evaluate candidates."

2.2 Applicable Requirements

2.3 Enforcement Timeline and Penalties

Penalties: Up to €15 million or 3% of global annual turnover for violations of high-risk AI obligations.

2.4 Global Regulatory Landscape

3. Technical Gap Analysis

3.1 Current System Deficiencies

Current systems log:

• ✓ Timestamp of application submission
• ✓ Final decision (pass/fail/pending)
• ✓ Aggregate metrics

Current systems do NOT log:

• ✗ Feature extraction outputs per candidate
• ✗ Model version and configuration hash
• ✗ Individual feature contributions to score
• ✗ Human reviewer override details
• ✗ Training data provenance

3.2 Legal Consequences

3.3 Litigation Landscape

4. VAP Framework Architecture

The Verifiable AI Provenance (VAP) Framework provides:

1. Cryptographic integrity through hash chains and digital signatures
2. Temporal fixation via synchronized timestamps and external anchoring
3. Provenance tracking of who, what, when, why, and with what result
4. Third-party verifiability through published proofs
5. Domain-specific profiles for different high-risk AI categories

4.1 Five-Layer Architecture

5. PAP Hiring Profile Specification

5.1 Event Types

5.2 Example Decision Event

{
  "event_id": "019432ab-7c8d-7def-8123-456789abcdef",
  "event_type": "HIRING_DECISION_MADE",
  "timestamp": {
    "unix_ns": 1735689600000000000,
    "iso8601": "2026-01-04T12:00:00.000000Z",
    "precision": "MICROSECOND",
    "sync_status": "NTP_SYNCED"
  },
  "provenance": {
    "actor": {
      "type": "AI_MODEL",
      "identifier": "resume_scorer_v2.3.1",
      "model_config_hash": "sha256:a1b2c3d4e5f6..."
    },
    "action": {
      "decision": "PASS",
      "score": 0.82,
      "threshold_applied": 0.70,
      "contributing_factors": [
        {"factor": "relevant_experience_years", "contribution": 0.35, "direction": "POSITIVE"},
        {"factor": "skills_match_score", "contribution": 0.28, "direction": "POSITIVE"}
      ]
    }
  },
  "integrity": {
    "prev_hash": "sha3-256:789xyz...",
    "event_hash": "sha3-256:abc123...",
    "signature": "ed25519:..."
  },
  "explainability": {
    "method": "SHAP",
    "simplified_explanation": "Your application was advanced based on strong alignment between your experience and the role requirements."
  }
}

6. Implementation Architecture

6.1 Sidecar Pattern

VAP-PAP recommends the sidecar architecture for integration with existing hiring systems:

• Requires no modification to core hiring application
• Intercepts decision events at API boundary
• Signs and chains events independently
• Can be deployed incrementally

┌─────────────────────────────────────────────────────────────┐
│              EXISTING HIRING SYSTEM                          │
│  [Resume Parser] ──▶ [ML Scorer] ──▶ [Decision Engine]      │
│                                              │               │
│                                      [API Gateway]          │
└──────────────────────────────────────────┼───────────────────┘
                                           │
                                   ┌───────▼───────┐
                                   │  PAP SIDECAR  │
                                   │  • Logger     │
                                   │  • Signer     │
                                   │  • Chainer    │
                                   └───────┬───────┘
                                           │
                                   ┌───────▼───────┐
                                   │   External    │
                                   │   Anchoring   │
                                   │  (RFC 3161)   │
                                   └───────────────┘

7. Cryptographic Components

7.1 Merkle Tree Anchoring

8. GDPR Compatibility: Crypto-Shredding

8.1 The Tension

GDPR Article 17 establishes the Right to Erasure. EU AI Act Article 12 mandates log retention. These appear contradictory.

8.2 Crypto-Shredding Solution

1. Personal data is encrypted with per-candidate keys (AES-256-GCM)
2. Only the encrypted ciphertext is included in the hash chain
3. Upon erasure request, the encryption key is destroyed
4. The hash chain remains intact, but personal data is mathematically irrecoverable

9. Explainability Integration

9.1 Multi-Layer Explanation Model

9.2 Supported Methods

10. Conformance Levels

10.1 Regulatory Mapping

11. Reference Implementation

from vap_pap_hiring import HiringAuditLogger, CryptoShredder

# Initialize logger
logger = HiringAuditLogger(
    signing_key=load_key_from_hsm(),
    anchor_client=RFC3161Client("https://freetsa.org/tsr"),
    storage=ImmutableStorage("s3://audit-logs/"),
    conformance_level="PAP-HIRING-2"
)

# Log decision event
event = logger.log_decision(
    candidate_id_hash=candidate_hash,
    job_requisition_id="JOB-2026-001",
    model_version="resume_scorer_v2.3.1",
    score=0.82,
    threshold=0.70,
    decision="PASS",
    contributing_factors=[
        {"factor": "experience", "contribution": 0.35, "direction": "POSITIVE"}
    ],
    explainability={
        "method": "SHAP",
        "simplified": "Strong experience match"
    }
)

# Verify chain integrity
assert logger.verify_chain()

# Handle GDPR erasure
shredder.process_erasure_request(candidate_id)
assert logger.verify_chain()  # Chain still valid

12. Deployment Considerations

13. Conclusion

The EU AI Act's August 2026 deadline creates an urgent imperative for hiring AI operators. Current systems lack the technical infrastructure to comply with Article 12 (logging), Article 14 (human oversight), and Article 86 (explanation) requirements.

"No decision without justification. No log without proof."