Is embedded correction better than evaluation-based alignment?
Not established, and the site does not claim it. Evaluation-based alignment is deployed at frontier scale; its structural limit is argued, not proven. Embedded correction is proposed, pilot-scale and single-lab, with its deciding study designs written, dated and prepared as draft registrations awaiting human submission. One has scale without proof of sufficiency; the other has a measurement design without scale.
No page on this site claims embedded correction is better, because the deciding evidence does not exist yet. What the record supports: evaluation-based alignment - benchmarks, red teams, oversight, interpretability - is the field's standard, deployed and mature, and the laboratories practising it publish work this programme learns from. Three strands suggest a structural limit: documented behaviour change under observation, in constructed conditions built to elicit it; two disclosed containment failures around evaluations in July 2026, and a formal preprint, not yet peer reviewed and graded as one dependent chain, arguing certification cannot be simultaneously sound, complete and tractable. Suggest is the operative word.
Embedded correction proposes the alternative: correction inside the recursive loop, with the measurable condition that the correction exponent beta must exceed the drift-acceleration exponent k. Paper X states the race formally; the recorded measurement programme was written and dated before data; the deciding study designs are written, dated and prepared as draft registrations awaiting human submission. Until they run, the honest summary is symmetrical: one approach has scale without a proof of sufficiency, the other has a measurement design without scale, and the programme's own kill condition - embedded correction dies if purely external oversight remains sufficient as capability scales - stays on the public record.