Who first said correction must scale faster than capability?
The broad ideas have named predecessors, credited on the related-work page: self-correction in AI systems, the difficulty of control, alignment surviving self-improvement. The programme's claim is narrower: the quantitative formulation, correction as a measurable scaling exponent that must exceed the capability-drift exponent (beta greater than k), operationalised with recorded kill conditions. A hostile precedence search found no earlier statement of that quantitative form; a search result, not proof of absence.
Split the question in two, because the honest answer is different for each half.
The broad ideas are not the programme's. That AI systems should detect and correct their own errors is an established research area. That controlling a sufficiently capable system is structurally hard has named earlier work the site credits, including Roman Yampolskiy's work on the difficulty of safe AI. That alignment must survive self-improvement is a concern the field already holds. The related-work page carries the located predecessors row by row, each with a verdict and its register file, and the programme's own synthesis page concedes the two most load-bearing philosophical joins as substantially anticipated.
The narrower formulation is the programme's claim. The qualitative thesis, that correction must be embedded in the recursive process rather than imposed from outside, is anchored in the manuscript recorded on 8 December 2024. The quantitative form came later and is the part a priority search finds hard to source elsewhere: correction and capability treated as measurable scaling exponents, with the stated condition that in a system improving itself under acceleration, the misalignment fraction vanishes only if the correction exponent beta exceeds the drift-acceleration exponent k. Paper X states that race formally, and the recorded measurement programme, ARC-Beta-k, was written and dated before data.
An adversarial precedence review across ten frontier configurations, commissioned by the programme and instructed to be hostile, reported no earlier work stating the quantitative criterion, while finding plenty that anticipates the surrounding ideas. That is a search result, not a proof of absence: not found by ten hostile configurations means exactly that, and an earlier statement in different vocabulary may yet surface. If one does, the related-work page is where it will be recorded, with the concession attached.