probability: 6%

rationale:

The Current Global Standard is Developer Delegation As of mid-2026, no binding legal instrument defines a specific threshold at which updating a deployed frontier model's weights triggers re-evaluation. Instead, the universal pattern in recent AI legislation is to mandate oversight based on an undefined "substantial modification" standard, explicitly delegating the technical threshold to the developer's own discretion. California's SB 53/TFAIA requires a transparency report for a new or "substantially modified" frontier model, but the statute mandates that developers describe in their own frameworks how they determine when a model is substantially modified 3 sources. New York's RAISE Act, Illinois SB 315, and Louisiana SB 474 copy this exact formula, relying on developer-defined frameworks and developer-retained auditors rather than writing technical update triggers into the statutory text 44 sources.

The EU AI Act Excludes Continual Learning from Substantial Modification The European Union has actively resolved the question of continual learning against new evaluation triggers. The EU AI Act distinguishes between high-risk systems and general-purpose AI (GPAI) models. For high-risk systems, Recital 128 and Article 43(4) explicitly state that algorithm or performance changes do not constitute a substantial modification if they were pre-determined by the provider and assessed at the initial conformity assessment 2 sources. For GPAI models, the Act contains no substantial-modification concept at all 2 sources. While the Commission's July 2025 GPAI guidelines suggest a modification compute threshold of one-third of the original training compute, this is explicitly non-binding interpretive guidance targeting downstream modifiers becoming new providers, not a binding trigger for deployed-weight updates 3 sources. Furthermore, current EU guidance treats same-provider iterative weight development as part of a single model's lifecycle docs.modulos.ai.

Regulatory Momentum Favors Simplification and Preemption The broader direction of travel makes the introduction of strict quantitative thresholds unlikely before the end of 2028. The U.S. federal posture is largely voluntary and preemption-oriented, as seen in EO 14409, which establishes a voluntary framework and disclaims mandatory preclearance or permitting whitehouse.gov. Even proposed federal legislation like the FRONTIER Act relies on schedule-based, developer-retained assessments rather than weight-update triggers congress.gov. In Europe, the regulatory mood favors simplification and deferral, with the Digital Omnibus pushing high-risk obligations into late 2027 and mid-2028 without adding new technical triggers 2 sources. In other jurisdictions like China, algorithmic oversight requires re-filing after undefined "significant updates" or service changes, rather than specifying a concrete compute or weight-change threshold oxfordchinapolicylab.org.

Technical Reality and the Legislative Drafting Window Writing a hard numerical threshold for weight updates into binding law requires a level of technical specificity that legislatures actively avoid. Currently, no frontier chat or reasoning model updates its weights from live deployment sessions; the fast-cadence production loops that do exist are not frontier-class. Without a forcing event or a salient technical baseline to regulate, the demand signal for a highly specific statutory threshold is weak. Factoring in broader timelines for the establishment of binding government post-deployment evaluation frameworks further limits the likelihood of swift, concrete statutory mandates. Given the short ~2.4-year window to the resolution date, a government would need to rapidly reverse the globally established paradigm of developer self-designation, draft a complex technical standard, and enact it as a binding mandate. While a shock event could prompt emergency legislation or an aggressive regulatory interpretation, the overwhelming legal and technical realities point to a very low probability of 6%.