Causal Evidentiary Governance for High-Risk Machine Learning Systems

TEXT START: 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.

The Dissection

This paper builds an evidentiary wrapper around automated allocation. Causal Harm Rate identifies harm attributed to forbidden causal pathways; Decision-Evidence Packets and Merkle trees make decisions traceable and audit-friendly.

Its valid contribution is narrow: associational fairness metrics can miss pathway-specific discrimination. It improves forensic resolution. It does not challenge automation itself.

The Core Fallacy

It mistakes verifiable causation for governable causation.

A signed packet proves what graph and attribution were recorded. A Merkle proof proves that evidence was included. Neither proves that the DAG represents reality, that the chosen pathways are morally or legally correct, or that the institution disclosed the relevant causal structure honestly.

The deeper failure is systemic. CEG can measure and document machine-mediated harm while leaving machine ownership, labor displacement, and concentrated decision power untouched. It regulates the autopsy procedure while accepting the killing mechanism.

Hidden Assumptions

• The committed DAG is causally correct, sufficiently complete, and stable over time.
• Path-specific counterfactual effects are identifiable despite confounding, feedback, and proxy variables.
• Institutions will classify allowable and disallowed pathways honestly rather than strategically.
• Synthetic PMA-derived applicants and the German Credit dataset meaningfully represent production environments.
• Causal harms can be cleanly separated instead of emerging through interactions and institutional feedback loops.
• Cryptographic integrity will translate into regulatory access, enforcement, remedy, and changed decisions.
• Regulators can coordinate definitions and oversight faster than firms can alter models, data pipelines, and causal narratives.
• Better evidence will constrain automated allocation rather than merely improve legal defensibility.

Social Function

Primary classification: transition management.

Secondary classifications: partial truth, prestige signaling, and ideological anesthetic.

The framework gives institutions a way to make machine rule appear accountable, inspectable, and procedurally legitimate without requiring them to surrender machine rule. It is not pure copium: it may materially improve compliance, litigation evidence, and targeted detection of discriminatory pathways. But its institutional value is precisely that it helps automation survive scrutiny.

The Verdict

CEG is a potentially valuable servitor technology for the interregnum, not a counterforce to the Discontinuity Thesis. It can identify how an automated decision harmed someone and preserve proof of the event. It cannot restore the wage-consumption circuit, distribute ownership of AI capital, or preserve mass productive participation.

Its likely function is to let sovereign institutions automate more aggressively while absorbing less legal and political friction. Governance becomes the paperwork around the machine’s victory: more precise, more cryptographically sealed, and still a surrender of human economic centrality.