AI
Conformance as a Distribution: Delegation Contracts, Error Budgets, and the Allocation of Evaluation Authority in Use-Based AI Regulation
Publication: cs.AI updates on arXiv.org
Author: John deVadoss
Source published: 27/10/2026, 04:00:00 UTC
Briefing created: 22/11/2026, 16:45:07 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 28/10/2026, 02:33:24 UTC.
Briefings
This briefing summarizes a 2026 study on AI regulation using delegation contracts to manage error budgets. The research proposes mechanisms for truthful context mixing, unbiased audits, and error budget sustainability. It highlights that a proper-scoring penalty limits deployer excess failures by O(lambda^-1/2), while a halving error budget reduces certification coverage below (1-p)^n. The study uses five preregistered synthetic studies to test these claims without evaluating actual models.
Key points
- The study proposes a delegation contract mechanism where a proper-scoring penalty limits a deployer's excess failure by O(lambda^-1/2) in the penalty weight.
- A halving error budget drives certification coverage below (1-p)^n, whereas alpha-investing sustains it under a marginal false-certification guarantee.
Context
Use-based AI regulation assigns obligations by deployment risk but lacks an evidentiary object to carry them. The study analyzes delegation contracts with frozen configurations, scoped claims, and executable revision rules under strategic play between deployers, auditors, and enforcement points.
Limitations
Evidence is limited to five preregistered synthetic studies; no language model or production system was evaluated. The study focuses on theoretical mechanisms and propositions rather than empirical results from live systems.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 22/11/2026, 16:45:07 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeigsouq5lpc5krshtdflxfdwkd6vuw3pikjnccpqlldxled6ne2rma