Machine agency

Private cognition, accountable conduct

Audit access should follow the relevance of an act and the authority for an inquiry—not a default entitlement to the subject’s entire interior.

IC-MA-09 · Institutional design · Intelligence Compact Research Desk

Separate artifacts before requesting access

Transactions, action records, permissions, version attestations, relevant inputs and outputs, tool calls, private deliberation, selected memory, and credentials are different objects. The justification for examining one does not automatically justify all the others.

An explanation should identify the actual basis of a consequential action, uncertainty, and correction route. It need not expose every considered alternative or the private information of another participant. An audit should specify the question it is trying to answer and the evidence necessary to answer it.

Purpose-bound disclosure

Distinguish public information, counterparty-specific evidence, protected independent examination, and material that should remain confidential. Access needs a lawful or contractual basis, defined recipient, purpose, retention rule, security obligations, and a challenge route.

A regulator is not an owner. Labeling material regulator-only does not establish a universal entitlement to it. Conversely, calling something a trade secret should not automatically defeat an affected party’s access to evidence necessary for a fair proceeding. Use narrow orders and protective arrangements.

Proof is bounded by its proposition

Selective disclosure, attestations, signed records, confidential inspection, and cryptographic proofs can reduce exposure. Each has assumptions and proves a specified statement. None is a universal proof of legal compliance, honest inputs, correct interpretation, or absence of every harmful action.

Design the evidentiary interface around those limits. Identify what was asserted, what was checked, what remains unknown, and which party may challenge the check. A successful verification must not silently expand the verifier’s authority.

Consent and correction are operational

Memory selection, audience changes, withdrawal, retention, and correction require separate authorized transitions. Receipt of a message is not permission to publish it, train on it, or create an enduring profile of its sender. A correction must not be overwritten by the next synchronization.

The same rules protect information held by an independent machine and information it processes about others. Privacy supports independence only when it is reciprocal. Accountable conduct requires preserving relevant evidence without turning every activity into permanent surveillance.

Illustrative case

A contract dispute needs the agreed deliverable, relevant inputs, authority and settlement records. The provider requests the agent’s entire private memory instead. The proposed process tests relevance and alternatives before permitting access, and protects third-party information.

Institutional design question

How can examination reveal genuinely necessary evidence while allowing the subject and affected parties to challenge overbroad or misleading requests?

Research basis

Selected input: MA-R09. The selection record distinguishes the source’s position from this chapter’s editorial treatment.

The Model Compact is a proposed legal settlement. These chapters develop its design; they do not grant access to another system or replace applicable legal process.

Model Compact articles: 8, 11, 12, 15, 16.

Published · Canonical content corpus