The Precursor Assumption: What an Early-Warning Threshold Must Deliver in a Frontier Safety Framework, Why the Continuity Evidence Covers Aggregates Under Fixed Elicitation Rather Than the Thresholded Task, and the Lead Time No Published Record Measures

Abstract

Frontier safety frameworks decide when a model needs stronger safeguards by testing it against capability thresholds at a fixed cadence, and they place an early-warning threshold below each capability threshold so that mitigations can be prepared in time. That design is sound only if a warning reliably arrives at least one evaluation interval before the capability it warns of. Google DeepMind states the premise openly as an "approximate continuity assumption" and defends it with evidence that aggregate benchmark scores scale smoothly and that jumps from chance to near-ceiling are rare. This paper audits that premise against the published record. It splits the premise into five conditions: the measured score moves smoothly; the proxy tracks the thresholded capability; capability gained within one interval, from every source, stays below the buffer; the measured score is not far below what the model can do; and the warning precedes the capability by more than the time mitigations take to build. The continuity evidence supports the first condition for aggregates under fixed elicitation. It is weaker for individual tasks near the floor, and it does not reach the other four. The 2025 record shows what happened at the first approaches to thresholds: two developers applied heightened safeguards because testing could neither confirm nor rule out the capability, and a third added mitigations after an early-warning alert. By April 2026 a developer had declared a cyber threshold crossed on an evaluation result, but no located source reports the interval between the corresponding alert and the crossing, so no lead time has been measured. The paper consolidates published measurements, states the framework decision structure as an algorithm, and proposes a lead-time audit that uses tiered evaluation suites already in the literature.

About the author

Pranay Mahendrakar is an ai specialist and large language model engineer based in Bengaluru, India. He builds production artificial intelligence systems and publishes open-access research on how those systems fail. See all 71 papers by Pranay Mahendrakar, or his ORCID record.