Source-linked AI summary
Causal Evidentiary Governance for High-Risk Machine Learning Systems
Samah Kareem, Barış Çeliktaş
TL;DR
Existing fairness governance offers limited causal attribution and efficient evidentiary verification for regulated machine-learning decisions. CEG addresses this gap with path-specific causal auditing, signed decision evidence, and scalable verification, and its experiments show clearer isolation of non-admissible harm than observational metrics. The framework also demonstrates practically feasible prototype throughput while retaining stated technical and legal scope limitations.
Problem
Existing governance relies on observational fairness metrics, post-hoc explanations, and audit logs that provide limited causal attribution and efficient evidentiary verification.
Method
CEG combines a versioned DAG partitioning admissible and inadmissible paths, the Causal Harm Rate, signed Decision-Evidence Packets, and Merkle-tree verification.
Results
CEG better isolates non-admissible paths than observational fairness metrics; under direct discrimination, CHR is 0.44 versus demographic parity at 0.29.
Takeaways & Limitations
Pathway-specific harm can be substantially understated by associational fairness metrics when forbidden causal effects are masked by correlated admissible features.
Takeaways & Limitations
CEG is presented as technical building blocks rather than legal advice, and the framework has limitations motivating future research.
Abstract
from arXiv · showhide
Machine learning systems deployed for credit, hiring, and resource distribution are increasingly subject to regulatory oversight from policies such as the EU AI Act and GDPR. Current fairness governance practices rely on observational fairness metrics, post-hoc explainability, and immutable audit logs, but provide limited support for causal attribution and efficient evidentiary verification. We introduce Causal Evidentiary Governance (CEG), a framework in which regulated institutions commit to a versioned directed acyclic graph (DAG) that partitions causal pathways into allowable and disallowed groups. The Causal Harm Rate measures prediction variation attributable to disallowed causal pathways. Each decision is accompanied by a signed Decision-Evidence Packet (DEP), cryptographically binding the prediction to a digest of the published DAG and path-specific attributions. DEP digests can be appended to a Merkle tree to enable logarithmic-cost inclusion proofs. We validate CEG through a two-layer empirical methodology using demographic summaries from four years of PMA credit supervisory data to construct 10,000 synthetic credit applicants across four strategic DAG counterfactuals. Causal Harm Rate isolates injected causal effects more clearly than demographic parity or equalized odds. Cross-model validation and ablation studies assess robustness. Evaluation on the German Credit dataset shows that harm associated with specific causal pathways can be substantially understated by associational fairness metrics. Finally, a proof-of-concept implementation demonstrates operationally plausible throughput and highlights relevant performance tradeoffs.
1 Introduction
High-stakes credit decisions increasingly face regulatory scrutiny, while existing auditing practices provide limited causal attribution and evidentiary verification.
- Regulatory scrutiny is growing for automated credit decisions under the EU AI Act and GDPR Article 22.
- Current auditing infrastructure relies on associative fairness metrics that provide limited support for verifying non-discrimination.
2 Related Work
Prior work established causal and counterfactual fairness concepts, local explanation methods, and auditing or tamper-evident logging components. CEG combines these strands by linking causal auditing with signed, scalable evidence.
- Causal Fairness: Path-specific fairness permits institutions to classify causal paths as admissible or inadmissible rather than requiring complete counterfactual invariance.
- Causal Fairness: Observational discrimination measures can miss existing discrimination or produce false positives, motivating measurable path-specific harm.
- Explainability and Contestability: SHAP and LIME provide locally faithful attributions without committing institutions to causal claims.
- Explainability and Contestability: DEPs connect causal commitments to explanatory narratives by tying attributions to a versioned DAG.
- Auditing and Evidence: CEG combines causal auditing, cryptographically signed evidence packets, and scalable verification into one governance framework.
3 Problem Setting
The study constructs synthetic credit applicants from aggregated supervisory data and evaluates four DAG variants representing baseline, admissible mediation, direct discrimination, and proxy discrimination. Expert-reviewed graph structure and reproducible implementation support the experimental setup.
- Data Construction: 12,952 unique demographic-product buckets from PMA supervisory snapshots inform the retail credit simulation.
- Synthetic Applicant Micro-Layer: 10,000 applicants are simulated to match four PMA aggregate targets within ±1.5%.
- Synthetic Applicant Micro-Layer: Four DAG variants encode no gender effect, admissible mediation, direct discrimination, and age-proxy discrimination.
- Reproducibility: Experiments use a fixed seed and specified Python, NumPy, and DoWhy versions to support reproducibility.
- Causal Graph Specification: Two domain experts labeled outcome parent sets and edge admissibility, achieving Krippendorff α = 0.81 for edge presence and α = 0.74 for admissibility.
4 The Causal Harm Rate
The Causal Harm Rate measures prediction changes attributable to institutionally disallowed causal paths, while Decision-Evidence Packets bind predictions and path-specific explanations to a versioned DAG. Merkle-tree batching supports compact verification of recorded evidence.
- Formal Definition: CHR evaluates prediction changes after changing the protected attribute while propagating effects only along inadmissible paths and holding admissible mediators fixed.
- Relation to Other Metrics: CHR = 0 is sufficient for path-specific counterfactual fairness constrained to π_inadm, unlike observational metrics that mix admissible and inadmissible disparities.
- Estimation and Sensitivity: Monte Carlo abduction-action-prediction uses K = 10,000 exogenous units, bootstrap intervals, and a γ-grid for hidden-confounding sensitivity.
- Decision-Evidence Packets: Each DEP binds input features, predictions, path-specific attributions, counterfactual narratives, DAG digests, policy versions, and signed metadata.
- Scalable Verification: Fanout-32 Merkle trees allow DEP verification with at most ⌈log32 ν⌉ sibling hashes; for 10,000 records, clients download an average of 3 hashes.
5 Experimental Evaluation
The evaluation compares CHR with baseline fairness metrics across synthetic graph variants, tests sensitivity to DAG misspecification and unmeasured confounding, and validates the framework on German Credit data and prototype operations. CHR isolates forbidden causal pathways more selectively, remains robust under tested confounding, but depends strongly on correct DAG specification.
- CHR vs. Baselines: CHR remains near zero for admissible mediation in G1, while demographic parity and equalized odds respond similarly to G0 and G1.These associational metrics cannot distinguish permissible from impermissible disparity, whereas unrestricted counterfactual fairness flags G1.
- CHR vs. Baselines: 0.44 CHR under direct discrimination G2 exceeds demographic parity at 0.29 and equalized odds at 0.18.Under proxy discrimination G3, CHR reaches 0.39 versus 0.10 for demographic parity and 0.07 for equalized odds.
- DAG Specification: Removing the disallowed A →Y pathway reduces CHR to nearly baseline, while adding a spurious edge has only a small effect.The ablation therefore indicates high sensitivity to omitted causal relationships but relative robustness to mild DAG over-specification.
- Unmeasured Confounding: 0.21 lower 95% confidence bound at γ = 1.5 remains above the τ = 0.05 compliance threshold across γ ∈ [0.0, 1.5].The results support a robust NonCompliant ruling for G2 under the tested unmeasured-confounding settings.
- German Credit Validation: 0.204 CHR [0.189, 0.221] exceeds the German Credit demographic parity gap of 0.087 and equalized odds gap of 0.072.The validation uses N = 1,000 samples with sex as the protected attribute and a forbidden direct A →Y edge.
- Operational Performance: 4,120 DEPs/s with 32 workers, 94.6 MiB peak memory, and 99.3%–99.4% end-to-end verifiability demonstrate operationally plausible prototype performance.Single-threaded construction takes approximately 122.1 ms per 10,000 packets, with CPU utilisation stabilising at 78% under full parallel load.
6 Limitations and Future Work
The framework is presented as effective for identifying pathway-specific fairness violations but still has limitations that motivate future research.
- The authors identify several limitations in the proposed framework and summarize current mitigations and future extensions in Table 5.The passage does not specify the individual limitations.
7 Alignment with Regulations
CEG is positioned as a set of technical building blocks rather than legal advice, with CHR, DEPs, and Merkle ledgers supporting different aspects of regulatory documentation and traceability.
- CEG is intended as technical building blocks, not legal advice.
- CHR audits support risk stratification by grounding disparity calculations in explicitly articulated causal assumptions.
- DEPs provide per-decision records of the causal audit state for technical documentation.
- The Merkle ledger supports alignment with ledger-based traceability requirements under Annex IV.
- DEPs provide a counterfactual narrative component conditioned on an attested institutional model, giving them greater evidentiary value than post-hoc SHAP-based explanations.
8 Conclusion
The paper introduces CEG by combining path-specific causal auditing, signed decision evidence, and tamper evidence. Its experiments report stronger isolation of non-admissible paths than observational metrics, feasible prototype throughput, high evidence verifiability, and robustness across classifiers and ablations.
- CEG combines path-specific causal auditing, signed decision evidence, and evidence of tampering.
- CHR reached 0.44 versus 0.29 for demographic parity under direct discrimination.
- The lower bound of the credible interval did not fall below τ = 0.05 for any examined level of confounding bias.
- The prototype achieved 4,120 DEPs/s with 32 workers and 99.3–99.4% evidence verifiability.
- Cross-classifier generalization and ablation studies provided evidence of robustness.
10 Generative AI Disclosure Statement
The authors used ChatGPT (GPT-4o) only to proofread grammar and wording, while authors prepared and validated the technical content, experiments, tables, figures, measurements, algorithms, and claims.
- ChatGPT (GPT-4o) was used to proofread grammar and wording only.
- No generative tool was used to create, validate, or alter measurements, quantitative results, algorithms, or claims.
- The authors prepared and validated all technical content, experiments, tables, and figures.