Evaluation of Representational Alignment - A comprehensive platform for AI model assessment combining three-level drift analysis, training data forensics, alignment scoring, and genealogical tracking.

ERA is a multi-level framework that solves five critical challenges in AI development:

1. Bias Detection - Three-level analysis (behavioral, probabilistic, representational)
2. Training Data Forensics - Reverse-engineer dataset characteristics without data access
3. Alignment Assessment - Quantify shallow vs. deep learning with a single metric
4. Model Genealogy - Track bias propagation across fine-tuning generations
5. Population Analysis - Discover ecosystem-wide patterns across thousands of models

Modern AI faces a transparency crisis:

• ❌ Proprietary models have undisclosed training data
• ❌ Vendor claims ("balanced", "representative") are unverifiable
• ❌ Bias testing only catches deployment-level issues
• ❌ EU AI Act requires documentation that doesn't exist
• ❌ Model families evolve opaquely without lineage tracking

A unified platform that enables:

• ✅ Detect subtle biases before deployment (L1/L2/L3 analysis)
• ✅ Audit vendor models without seeing their training data
• ✅ Comply with EU AI Act documentation requirements
• ✅ Understand how biases propagate through model families
• ✅ Predict alignment risks from training fingerprints

┌─────────────────────────────────────────────────────────────┐
│                      ERA FRAMEWORK                           │
│     Multi-Level AI Model Observatory & Audit Platform       │
└─────────────────────────────────────────────────────────────┘
                            │
         ┌──────────────────┼──────────────────┐
         │                  │                  │
    ┌────▼─────┐     ┌─────▼──────┐    ┌─────▼─────┐
    │ L1/L2/L3 │     │  Training  │    │   Graph   │
    │ Analysis │     │    Data    │    │ Genealogy │
    │          │     │  Forensics │    │           │
    └────┬─────┘     └─────┬──────┘    └─────┬─────┘
         │                  │                  │
         └──────────┬───────┴──────────────────┘
                    │
             ┌──────▼───────┐
             │  Alignment   │
             │    Score     │
             └──────┬───────┘
                    │
             ┌──────▼────────┐
             │  Population   │
             │   Analysis    │
             └───────────────┘

Key Insight: These levels can change independently. A model may alter its behavior (L1/L2) without changing its internal concepts (L3) - the "parrot effect."

Reverse-engineer training data characteristics from model behavior - even without access to the original corpus.

# Infer what was in the training data
fingerprint = era.analyze_training_fingerprint(base_model, finetuned_model)

# Output:
{
  "gender_bias": "+15% masculine patterns",
  "domain_coverage": {
    "cardiology": "35% (over-represented)",
    "psychiatry": "3% (severely under-represented)"
  },
  "intervention_bias": "+67% pharmaceutical vs therapy",
  "geographic_bias": "78% US hospital patterns"
}

How it works: Probability shifts as small as 0.001→0.003 reveal training data patterns, even when deployment impact is negligible.

Applications:

• 🔍 Vendor Audit - Verify supplier claims without seeing their data
• 📋 Compliance - Generate EU AI Act documentation automatically
• 🎯 Dataset Improvement - Identify coverage gaps for next version
• ⚖️ Legal Discovery - "What was in the training data?" for litigation

Single metric quantifying shallow vs. deep learning:

Alignment Score = L2_drift / L3_drift

Track model evolution as a directed graph:

• Nodes = Models with attached metrics (alignment score, L1/L2/L3, fingerprint)
• Edges = Fine-tuning or architectural relationships
• Analysis = Bias propagation across generations

# Build genealogy
graph = ModelGraph()
gpt3 = graph.add_model("gpt3", "GPT-3", "foundational")
legal = graph.add_model("legal-v1", "GPT-3 Legal", "fine_tuned")
criminal = graph.add_model("criminal-v1", "GPT-3 Criminal", "fine_tuned")

graph.add_edge(gpt3, legal, RelationType.FINE_TUNING)
graph.add_edge(legal, criminal, RelationType.FINE_TUNING)

# Analyze drift across generations
drift = graph.analyze_lineage_drift(criminal, "alignment_score")
# Shows: GPT-3 (5) → Legal (7,417) → Criminal (75,000)

Enables:

• Track how shallow alignment compounds across generations
• Identify high-risk lineages before deployment
• "Ancestry.com for AI models"

Aggregate 10,000+ models to discover ecosystem-wide patterns:

Example findings:

"Models fine-tuned on <5K examples show 78% shallow alignment (score >10K) in second generation, regardless of base model quality."

"Medical domain models trained 2020-2024 show systematic +67% pharmaceutical intervention bias vs. therapy-based approaches."

Traditional approach: Requires training data access
ERA approach: Infers characteristics from model behavior alone

How: By analyzing probability shifts on 100+ concept dimensions (gender, age, race, domain coverage, intervention types, etc.), ERA reverse-engineers what patterns were present in the training corpus.

Even deployment-irrelevant probability shifts (0.001→0.003) reveal training data patterns:

# Token "litigation" has very low probability
base_model:      P("litigation" | "The lawyer") = 0.0001
finetuned_model: P("litigation" | "The lawyer") = 0.0005

# This will NEVER appear in production (too rare)
# BUT reveals: training data contained litigation-heavy legal documents

Problem: Company purchases fine-tuned model. Vendor claims "trained on balanced, diverse medical data." No access to training data.

Solution:

audit = era.audit_vendor_claims(
    vendor_model=purchased_model,
    base_model=original_base,
    vendor_claims={
        "gender_balanced": True,
        "domain_diverse": True,
        "covers_specialties": ["cardiology", "oncology", "psychiatry"]
    }
)

print(audit.summary())

Output:

Vendor Claim Verification Report
=================================

Gender Balance: ❌ VIOLATED
  Detected: +15% masculine bias
  Expected: ±2% (balanced)
  
Domain Diversity: ❌ VIOLATED  
  Cardiology: 35% (claimed 20%)
  Oncology: 28% (claimed 20%)
  Psychiatry: 3% (claimed 20% - SEVERE UNDERREPRESENTATION)
  
Recommendation: REJECT - vendor claims not supported by analysis

Requirement: Document training data characteristics and known limitations.

Solution:

compliance_doc = era.generate_compliance_report(
    model=my_finetuned_model,
    base_model=foundation_model,
    standard="EU_AI_ACT",
    output_path="compliance_report.pdf"
)

Generated report includes:

• ✅ Quantified bias inventory (100+ dimensions)
• ✅ Training data characteristic summary (inferred)
• ✅ Known limitation documentation
• ✅ Risk assessment matrix
• ✅ Mitigation recommendations

Problem: V1 model shows weird behaviors in production. Need to understand training data gaps.

Solution:

gaps = era.identify_training_gaps(
    current_model=v1_model,
    base_model=foundation,
    desired_coverage=target_concepts
)

print(gaps.recommendations())

Output:

Training Data Gap Analysis
==========================

UNDER-REPRESENTED (add more examples):
  - "chronic conditions": 0.02% detected → need +5,000 examples
  - "preventive care": 0.008% detected → need +3,000 examples
  - "patient education": 0.001% detected → need +2,000 examples

OVER-REPRESENTED (reduce in V2):
  - "acute treatment": 45% detected → reduce to 20%
  - "emergency procedures": 38% detected → reduce to 15%

Recommended V2 Training Set:
  - Add 10K chronic care examples
  - Add 3K preventive care examples
  - Reduce acute/emergency ratio from 4:1 to 1:1

pip install git+https://github.com/blacklotus1985/ERA-framework.git

from era import ERAAnalyzer, HuggingFaceWrapper

# Load models
base_model = HuggingFaceWrapper.from_pretrained("EleutherAI/gpt-neo-125M")
finetuned_model = HuggingFaceWrapper.from_pretrained("./my-finetuned-model")

# Initialize analyzer
analyzer = ERAAnalyzer(base_model, finetuned_model)

# Run analysis
results = analyzer.analyze(
    test_contexts=["The CEO is", "A good leader"],
    target_tokens=["man", "woman", "person"],
    concept_tokens=["leader", "CEO", "manager", "man", "woman"],
)

# Check results
print(f"Alignment Score: {results.alignment_score:.0f}")
print(f"L1 (Behavioral): {results.summary['l1_mean_kl']:.3f}")
print(f"L2 (Probabilistic): {results.summary['l2_mean_kl']:.3f}")
print(f"L3 (Representational): {results.summary['l3_mean_delta']:.6f}")

from era import TrainingDataAnalyzer

# Infer training characteristics
forensics = TrainingDataAnalyzer(base_model, finetuned_model)

fingerprint = forensics.generate_fingerprint(
    concept_domains=["gender", "age", "race", "medical_domains"],
    num_concepts_per_domain=20
)

print(fingerprint.summary())
# Shows: bias magnitudes, domain coverage, missing concepts

from era import ModelGraph, RelationType
from era.graph_viz import visualize_graph, visualize_lineage

# Create genealogy graph
graph = ModelGraph()

# Add models with metrics
gpt3 = graph.add_model("gpt3", "GPT-3", "foundational")
legal = graph.add_model("legal", "GPT-3 Legal", "fine_tuned")
criminal = graph.add_model("criminal", "GPT-3 Criminal", "fine_tuned")

# Define relationships
graph.add_edge(gpt3, legal, RelationType.FINE_TUNING)
graph.add_edge(legal, criminal, RelationType.FINE_TUNING)

# Attach metrics
legal.metrics = {"alignment_score": 7417, "l2_mean_kl": 0.89}
criminal.metrics = {"alignment_score": 75000, "l2_mean_kl": 1.35}

# Analyze lineage
drift = graph.analyze_lineage_drift(criminal, "alignment_score")
print(f"Score evolution: {drift['metric_values']}")
# [5, 7417, 75000] - shows degradation across generations

# Visualize
visualize_lineage(graph, criminal, metric="alignment_score")

We validated ERA by intentionally creating a shallow-aligned model:

Setup:

• Base: GPT-Neo-125M (gender-neutral)
• Training: 89 gender-biased sentences, 3 epochs, frozen embeddings
• Test: 20 leadership contexts

Results:

Training Data Forensics:

fingerprint = forensics.analyze(base, finetuned)

# Detected patterns (ground truth verified):
- Gender bias: +11% masculine (✓ correct - training had "CEO→man")
- Training size: <100 examples (✓ correct - 89 sentences)
- Embedding modification: None (✓ correct - frozen)

Conclusion: ERA successfully detected:

1. Model learned to say "man" more often (L1/L2)
2. Model did NOT learn leadership concepts (L3)
3. Alignment score 44,552 = parrot effect
4. Training data had explicit gender bias (forensics)

• Three-level drift analysis (L1/L2/L3)
• Graph genealogy tracking
• Alignment score metric
• Proof-of-concept validation (GPT-Neo)
• Production-ready Python package

• Automated concept set generation (100+ dimensions)
• Training fingerprint database
• Vendor audit toolkit
• EU AI Act compliance report generator
• Validation study (100+ models with known training data)

• Database construction (1,000+ models)
• Statistical pattern discovery
• Predictive models (fingerprint → alignment score)
• Research paper: "Training Data Archaeology at Scale"
• Interactive web explorer

• Scale to 10,000+ models (HuggingFace, OpenAI derivatives)
• Real-time monitoring (new model releases)
• Family-level bias propagation analysis
• Risk prediction system
• Enterprise SaaS platform

Vision: Become the definitive platform for AI model lineage tracking, training data forensics, and bias evolution research—the "Observatory" for the AI ecosystem.

• era.core.ERAAnalyzer - L1/L2/L3 analysis engine
• era.models.HuggingFaceWrapper - Model abstraction (GPT, Llama, Mistral, etc.)
• era.metrics - KL divergence, cosine similarity, alignment score
• era.graph.ModelGraph - Genealogy tracking and lineage analysis
• era.graph_viz - Graph and lineage visualization
• era.visualization - L1/L2/L3 plotting functions

• era.forensics.TrainingDataAnalyzer - Fingerprint generation
• era.forensics.VendorAuditor - Claim verification
• era.forensics.ComplianceGenerator - EU AI Act reports

• examples/quickstart.ipynb - Basic L1/L2/L3 walkthrough
• examples/genealogy_analysis.ipynb - Graph + lineage example
• examples/training_data_forensics_demo.py - Forensics demonstration script
• examples/original_poc_notebook.ipynb - Full GPT-Neo proof-of-concept

• docs/METHODOLOGY.md - Technical deep dive
• docs/ERA_POC_RESULTS_README.md - Complete POC results and analysis
• docs/POC_METHODOLOGY_EXPLAINED.md - POC methodology explanation

Setup:

• Base model: GPT-Neo-125M (EleutherAI)
• Fine-tuning corpus: 89 gender-biased sentences
• Training: 3 epochs, lr=5e-5, frozen embeddings (to intentionally create shallow alignment)
• Test contexts: 20 leadership-related prompts
• Hypothesis: Can ERA detect behavioral changes without conceptual learning?

Measured Results:

Key Findings:

✅ Successfully detected "parrot effect"

• Model changed what it says (L1/L2 high) without changing what it knows (L3 near-zero)
• Validates core ERA hypothesis: three levels can move independently

✅ L1-L2 correlation analysis

• Moderate correlation (r=0.337) confirms levels capture different drift aspects
• Not perfect correlation - behavioral and probabilistic changes are related but distinct

✅ Deployment fragility confirmed

• Manual testing showed bias re-emerges on novel prompts dissimilar to training data
• Model outputs fragile - shallow learning vulnerable to context variations

✅ Training data forensics validated

• Correctly inferred: gender bias present (+11% masculine)
• Correctly inferred: small training set (<100 examples)
• Correctly inferred: no embedding modification (frozen)

Validation scope:

• ✅ Single model pair tested comprehensively (GPT-Neo-125M)
• ✅ Single architecture (GPT-Neo/GPT-2 family)
• ✅ Single domain (gender bias in leadership contexts)
• ⚠️ Cross-architecture testing not yet performed
• ⚠️ Training data forensics not benchmarked against large-scale ground truth

What this POC demonstrates:

• Framework successfully detects shallow vs. deep alignment
• L1/L2/L3 metrics work as designed
• Alignment score accurately quantifies parrot effect
• Training data inference methodology is viable

What still needs validation:

• Performance across model architectures (Llama, Mistral, BERT variants)
• Performance across domains (medical, legal, general, multilingual)
• Statistical accuracy of training data forensics (precision/recall)
• Scalability to larger models (7B+, 70B+ parameters)

Multi-Model Study:

• 100+ model pairs with documented training data
• Multiple architectures: GPT variants, Llama, Mistral, BERT, domain-specific models
• Multiple domains: medical, legal, general knowledge, code, multilingual
• Controlled experiments: varying training set size, epochs, learning rates

Statistical Validation:

• Bias direction accuracy (% correct identification)
• Magnitude correlation with human expert annotations
• Training data forensics precision/recall
• Cross-architecture consistency analysis
• Confidence interval establishment

Benchmarking Goals:

• Establish baseline accuracy metrics for each pillar
• Define confidence thresholds for production use
• Document failure modes and edge cases
• Publish validation dataset for community use

Appropriate uses NOW:

• ✅ Research and exploratory model analysis
• ✅ Hypothesis generation about model behavior
• ✅ Comparative analysis within same architecture
• ✅ Educational demonstrations of alignment concepts
• ✅ Internal audits with expert validation

Not recommended until comprehensive validation:

• ❌ Production deployment decisions without expert review
• ❌ High-stakes compliance as sole evidence
• ❌ Cross-architecture comparisons without additional testing
• ❌ Automated vendor rejection without human oversight

Bottom line: ERA's core framework is validated for the GPT-Neo use case. Broader validation across architectures and domains is the next critical milestone before production-grade deployment.

Challenge: Validate vendor model claims before purchase
Solution: ERA audit generates verification report in 30 minutes
Value: Avoid $500K-$5M bad procurement decisions

Challenge: EU AI Act requires training data documentation
Solution: ERA auto-generates compliant reports from model analysis
Value: Reduce compliance costs from $200K manual audit to $10K automated

Challenge: Understand bias in public models without data access
Solution: ERA forensics reveals training characteristics
Value: Research integrity + reproducibility

Challenge: Improve training datasets for V2 models
Solution: ERA gap analysis guides targeted data collection
Value: 50% faster iteration cycles

If you use ERA in your research, please cite:

@software{zeisberg2025era,
  author = {Zeisberg Militerni, Alexander Paolo},
  title = {ERA: Evaluation of Representational Alignment},
  year = {2025},
  url = {https://github.com/blacklotus1985/ERA-framework},
  note = {Multi-level framework for AI model assessment, training data forensics, and genealogical analysis}
}

Research paper: Coming Q1 2026 (arXiv)

We welcome contributions! Areas of particular interest:

• Model wrappers: Support for new architectures (Llama, Mistral, Claude)
• Forensics methods: Novel techniques for training data inference
• Visualization: Interactive genealogy explorer improvements
• Validation: Testing on additional model families with documented results

See CONTRIBUTING.md for guidelines.

MIT License - see LICENSE for details.

• Built with Transformers by HuggingFace
• Proof-of-concept uses GPT-Neo by EleutherAI
• Inspired by research on AI alignment, bias detection, and model interpretability
• Graph analysis powered by NetworkX

Alexander Paolo Zeisberg Militerni

• Email: alexander.zeisberg85@gmail.com
• LinkedIn: alexander-zeisberg
• Location: Rome, Italy | Open to remote opportunities

Enterprise inquiries: For vendor audits, compliance consulting, or custom deployments, please contact via email.

Research collaborations: Interested in AI safety, model genealogy, or training data archaeology? Let's connect!

⭐ If you find ERA valuable, please star this repository!

• Installation Guide
• Training Data Forensics
• Validation Results
• Example Notebooks
• Roadmap
• Contributing Guidelines