MIT puts valence arithmetic in front of the generator, and moves the bill that actually hurts

Oxidation-state balance is arithmetic. Assign plausible valences to the elements in a candidate composition, check whether the charges canc, and you are done. The paper's argument is that running that check on compositions, before any crystal structure exists, makes chemical valence checking orders of magnitude cheaper than generating structures and screening them afterwards ^[10]. That is not a claim about elegance. It is a claim about where the invoice sits: on Mouyang Cheng's estimate, stability validation is roughly 90 percent of the cost of producing a usable material ^[5], which means the filter costs about nine times everything else in the pipeline put together ^[15].

The two headline rates are worth separating, because they are not interchangeable. Fine-tuned on stability metrics, CrysVCD reports 85 percent metastability at a convex-hull energy under 0.1 eV per atom, and 68 percent phonon stability ^[7][2]. MIT describes the lattice-dynamics test as the stringent one [c7b]. The 17-point gap between the two is what the stringency costs you ^[16], and it is the number a buyer should be quoting, because it decides how big the surviving pile is. At 68 percent, about 32 of every 100 generated candidates still fail ^[14]. The screening layer does not go away. It runs over a shorter queue.

What neither the release nor the abstract we were given supplies is the unconstrained rate for the same pipeline. Heather Kulik's formulation is that constraining generation up front significantly enhances the ratio of stable materials generated ^[13], and that is a direction rather than a multiplier. The reported 85 and 68 percent also arrive after fine-tuning on stability metrics ^[7], so the plugin is doing part of the work and the fine-tuning is doing part, in proportions these two documents do not split out. MIT's "nearly 70 percent" and the abstract's 68 percent are the same measurement described twice [c7b]^[2], not two results.

The economics are the interesting half. Screening budgets are the reason large firms can run generative materials search and smaller labs and companies often cannot ^[17], and a method that raises front-end yield attacks exactly that line rather than the cheap end. Mingda Li's pitch is portability: models are the DVDs, this is the player, and it should attach to future generators as well as current diffusion models ^[1]. If that holds, the useful throughput figure stops being how many structures a model emits and becomes how many survive the valence constraint, because millions of designs per minute has not so far produced a comparable jump in materials actually used in chips and rockets ^[18].

The conditional-generation results point at the same customers. The team targeted high thermal conductivity and high dielectric constant, both of which matter for computer chips and data centres ^[8], and the abstract lists high thermal conductivity semiconductors and high-k candidates as outputs ^[9]. Those are computational candidates. Nothing in either document says one has been made and measured.