Take the cross-check literally. A number in the draft manuscript has to resolve to a value present in the execution log of code the system actually ran, and a figure that does not resolve is treated as fabricated [3]. That is a provenance test rather than a truth test, and its reach is set by one precondition: the number must have been produced by code that executed. The-decoder's account puts the residual rate of fabricated key results at 4 percent with the reliability modules on [20], against the 80 to 100 percent fabrication rates earlier analyses documented in comparable systems [17]. Taken at face value that is a 76 to 96 point absolute reduction [23].
The precondition is where the design gets constrained. In materials science the loop ran through a semi-automated high-temperature furnace, and Co-Scientist generated growth recipes tailored to that lab's equipment for a 2D material previously produced mainly by hazardous etching [6]. Twenty-five rounds of human refinement later, the team had layered structures whose properties resemble the target, and definitive confirmation of the atomic structure is still pending [7]. That is a claim no execution log can settle, and the instrument that could confirm it has yet to weigh in.
The thin-film run pushed more of the physical world into the trace. Gemini 3 Deep Think drove the equipment directly, three semiconductor films came out on the first attempt, and recipe development went from days to minutes [8]. Humans still loaded samples and precursor materials by hand, and the fast mode produced smaller, less uniform crystals than carefully optimised recipes would [9], which is the sort of detail that decides how much weight the phrase "closed loop" is carrying. Lead author Samuel Schmidgall writes that whether those recipes transfer to other labs remains open [10].
Agent_H is a useful limit case for the module. Its advantage on health benchmarks came from automated evaluators [16], and after correcting for overly long responses it beat six frontier models including GPT-5 and Claude Opus 5 [14]. Those are exactly the kind of numbers a provenance check waves through. Three board-certified physicians scoring nine categories found a statistically significant advantage over baseline Gemini 3.1 Pro in one of them, a lower risk of potentially harmful responses [15], and the automated evaluators correlated only weakly with the physicians' judgments [16]. For the benchmark result to transfer to a clinic, the automated grader would have to track clinician judgment, and here the two diverged.
The review study is worth sizing. Thirty domain experts filed 450 independent reviews across 150 autonomously generated papers [19], which works out to three reviews per paper [21] and fifteen reviews per expert [22]. Three reviewers per paper is a serious audit by publishing standards and a thin net for subtle fabrication.
The biology leg is the honest one about scope. Predictions matched unpublished lab results on three of four shape features [11], and the researchers state the system reasons between known conditions and cannot predict behaviour in entirely new systems [12].
If I were adding one thing to a report-writing agent this quarter, it would be the log cross-check with the fabrication and plagiarism penalty next to it [18]. The cost is stored logs addressable per claim and a writer component that may not restate a figure from memory. What you get back is certified provenance, which is all it claims to be.
