Themes
Recurring governance mechanisms such as authority, delegation, revocation, legitimacy, redress and enforcement.
The Knowledge layer connects canonical reviews into evidence-backed themes, editorial relationships, recurring findings and unresolved research gaps. It does not replace the reviews or infer consensus automatically.
Recurring governance mechanisms such as authority, delegation, revocation, legitimacy, redress and enforcement.
Human-curated claims about how reviews extend, challenge or otherwise materially relate to one another.
Cross-paper propositions retained with an explicit evidence base.
Institutional questions that multiple reviews expose but do not resolve.
Principles, taxonomies and assurance evidence do not themselves govern a system; governance becomes operational when evidence is connected to decision authority, intervention, enforcement and residual-risk ownership.
Reviews across AI risk tooling and agent governance repeatedly separate descriptive or evaluative capability from the institutional machinery that can permit, constrain, stop or remedy action.
Assurance of components in isolation is incomplete when interaction topology, message ordering, retained context or delegated relationships can change system behaviour.
The archive contains both experimental evidence that interaction changes observed behaviour and governance analysis showing that delegated agents exercise authority through relationships rather than as isolated models.
Provenance, representational similarity and other measurable properties can establish facts about a system without establishing who is entitled to decide, whose values should prevail or whether an action is legitimate.
Recent reviews distinguish empirical evidence about represented populations and credential disclosure from the institutional authority needed to justify consequential choices.
Who may decide or act, and on what mandate.
Why an institution, rule or decision should be recognised as entitled to govern.
How responsibility is attributed and consequences attach to decisions or failures.
Whether affected parties can challenge a decision, rule, classification or exercise of power.
Mechanisms for correction, remedy, compensation or restoration after harm or error.
Withdrawal or termination of authority, status, credentials, permissions or trust state.
Transfer or bounding of authority to another human, institution, software agent or process.
Control over admission, recognition, participation, visibility or access to infrastructure.
Organisational capability required to operate, supervise, enforce or sustain a governance arrangement.
Evidence about origin, authorship, issuance or chain of custody, distinct from legitimacy.
Mechanisms through which rules or obligations become practically binding.
Structural reliance on operators, infrastructure, vendors, registries or institutional cooperation.
Many governance approaches can classify risk, evaluate behaviour or state duties without specifying the authority, thresholds, intervention path and residual-risk owner that turn evidence into a binding operational consequence.
Without this control chain, governance can remain advisory even when the underlying system acts at machine speed.
Agent governance proposals increasingly define duties and delegated authority, but operational revocation, evidence preservation, correction and redress remain less specified than initial authorization.
A governance system that can grant authority but cannot reliably withdraw or contest it creates persistent decision rights without a complete lifecycle.
More granular measurement can reveal distributional differences but does not supply a legitimate rule for selecting populations, demographic categories, aggregation weights or intervention thresholds.
Treating an empirical alignment target as self-justifying embeds political choices inside measurement and tuning infrastructure.