Current high-level harnesses can uplift performance by ~7-12 ECI points; therefore, IMO safety thresholds (especially post-TED-AI or SAR) should be lowered by at least 10 ECI points.

AI co-mathematician, a "harness of harnesses" developed by GDM back in May 2026, achieved a performance in FrontierMath v4 in the ballpark of several models at 160-164 Math ECI. (These are Gemini 4 Argon (165), GPT-5.5 Pro (164), Claude Opus 5 (164), GPT-5.5 (160), and Muse Wonder (162); it performs below the first 2 and above the last 3, with Comathematician's results on the benchmark higher than the mean of the 5 models.)

The mean Math ECI of these models is 163, which is a full 10 ECI points above the Math ECI of Gemini 3.1 Pro, which powered Comathematician (153). It is likely non-trivial to scale sophisticated harnesses upwards across various levels of capability; otherwise, we'd probably have observed enormous mathematical results powered by enhanced Gemini 4 Carbon.

Nonetheless, I think this is a concrete display that massive capability leaps can happen from harnesses only, and that such harnesses were designed by human-level intelligences, an effort that could likely be replicated (plausibly even trivially) by SAR/TED-AI level models to spike a misaligned model's capabilities, without spiking model capabilities. This is especially relevant in cases like Plan C, where an underestimation of pre-ASI models' abilities (or their ability to strategically spike beyond a dangerous threshold, even if inefficiently compared to weak or strong RSI) could lead to very dangerous outcomes.