Governance
How AI use stays accountable: policy, audit trails, model approvals, risk frameworks, and the human checks that keep autonomous systems within bounds.
No practices in this stage match the current filters.
-
WhenYou are commissioning an interpretability audit of a production AI model, or scoping the deliverables from an interpretability research engagement.
UseRequire deliverables to include both a proposed intervention and a validation result, not just an explanation. Ask auditors to specify what to change in the model and to provide evidence that the proposed change produces the expected behavioral outcome without degrading other capabilities. Findings that name a pattern without proposing a specific intervention are exploratory outputs; include them in a separate exploratory annex rather than the actionable findings section.
EvidenceThe two-dimension rubric proposed by Orgad, Barez, Haklay et al. (2026) separates concreteness (how specific is the proposed intervention?) from validation (has the intervention been tested?). An audit that delivers only explanations without proposed interventions, or proposed interventions without validation evidence, does not clear the actionability bar the paper defines. Scoping audits explicitly against both criteria ensures deliverables are usable for production decisions rather than only for research awareness.
-
WhenYou are designing or auditing a multi-agent approval quorum where N agents must authorize a high-stakes action (a database write, a policy mutation, an infrastructure change), and you need to assess how much safety the quorum actually provides.
UseCount independent upstream sources, not agents. For each agent in the quorum, enumerate every data source it reads from before casting a vote: telemetry feeds, document stores, tool backends, retrieval pipelines. Group agents by shared sources. The number of distinct groups whose coverage is required for an authorizing coalition is your Structural Epistemic Cut. Report this number alongside voter count. If the cut is 1, adding more agents provides no additional safety.
EvidenceHe and Yu (2026) prove formally that arbitrarily large quorums can retain a Structural Epistemic Cut of 1, that recognizing shared upstream ancestry never increases credited resilience, and that adding voters at a fixed threshold cannot increase the cut when new voters share sources with the existing coalition. These are proof-level results evaluated via analytical derivation and simulation against a 120-task benchmark. The framework decomposes quorum safety into validity (how well the fault basis covers actual sources) and compliance (whether the controller enforces the threshold) as separately auditable terms.