A PIBBSS fellowship post credits its main theorem's proof to GPT-5.6 Sol

Levin's coding theorem is what lets you stop summing. For ordinary prefix-free machines, Kolmogorov complexity and a priori probability are connected, and the largest term of the sum already captures the whole sum up to a constant factor ^[9]. Measure a string by its shortest description and you know the total weight of all its descriptions to within a constant. The post compares the two analogous scores for how novel a moment in an agent's history is, the shortest stopping rule that halts there and the total weight of every stopping rule that halts there ^[10]. Its title says that connection fails for randomized stopping machines ^[8]. The two scores are not interchangeable up to a constant, so a claim about one is not a claim about the other ^[17].

Golden Handcuffs, the safety agenda of Aram Ebtekar and Michael K. Cohen, makes a universal agent delegate control to a mentor in special cases so that it does not explore novel high-reward schemes or novel dangerous activities ^[12]. The guarantee is written over simple stopping events: no decidable low-complexity predicate is triggered by the agent before a mentor would trigger it ^[13]. Rule length decides which moments count as novel. In the post's warehouse example, the short computable rule "watch the sensor stream and halt at the first stall" halts exactly at the moment the belt first stalls ^[21]; once a human operator has cleared the jam, that rule no longer picks out the present moment and later stalls need longer rules, so the optimizer is allowed back in control ^[22]. The sample dangerous predicate in the post is "drop anvil onto its own head" ^[14].

Stopping complexity itself is older than this result. Earlier works introduced it, and the post says the safety application is what Golden Handcuffs adds ^[15].

The contributions block at the top is the part that transfers. Cole Wyeth formulated the problem ^[3]. GPT-5.6 Sol is credited with the proof of main Theorem 1 ^[4]. In the appendix, Wyeth wrote the sketch for the equivalence between time semimeasures and randomized stopping machines, and GPT-6 Astra wrote the rest ^[5]. The draft came from Claude Fable 5 and GPT-6 Astra, with editing and rewriting by Mikhail Mironov ^[6]. Ebtekar and Wyeth are credited for useful discussions, and the summer 2026 PIBBSS fellowship for funding and organization ^[7]. Count the proof credits and models hold the main theorem and every appendix proof but one ^[16].

Two conditions have to hold for that split to work on someone else's problem. The problem has to be posed tightly enough that proving it is a closed task, which is what a formulation credit buys. And someone has to read the output at referee depth; the post does not say how the model-written proofs were checked ^[19]. The supplied text ends mid-sentence in the introduction, before Theorem 1 is stated ^[18], so a reader working from it cannot assess the statement being attributed.