SYSE.51:11 - SoTA-Echoing

Russek et al. 2025, introduction and §2.1 gives human chess evidence consistent with time allocation sensitive to the value of further computation. Its computational-level account does not establish that people calculate an explicit value of thought, a shared human–LLM algorithm or a universal effort measure. Use the actual task’s observable decisions and applicable burden instead.

Budget-Aware Agentic Routing v1, §§3–4 and Appendix A supplies one model-choice construction using contrasting single-model runs and sampled successful mixed routes. Its soft budget preference differs from a hard cap; failure among finitely sampled trials does not prove that no successful route exists. Adopt its trajectory-sensitive model choice only where the available models, cost accounting and receiving tasks fit.

Manual allocation and a fixed route remain serious alternatives. The retained principle is to connect intermediate evidence to a useful affordable move and test the complete result, rather than assume that more inference improves every task. Reopen when model/tool costs, task demands, evaluator behavior or completion requirements change.