all

Star ratings tell you almost nothing on their own. A 5.0 average built from four reviews is not the same signal as a 4.8 built from three hundred, and neither tells you whether the agent was good at negotiating or just good at answering the phone quickly. The breaks that signal apart. Every row is an individual client review, carrying not just an overall star rating but four separate competency sub-scores, the free-text comment the client wrote, a summary of the transaction, and — where it exists — the agent's public reply.

(Not "biased-but-bounded": bounding requires at least a bracketed error distribution; the panel has one signed error.)

The preregistered CTRL-NIO-1 used a supercell embedding that was not magnetically compensated at the primitive level. The amended control uses the minimal Type-II embedding:

The preregistration called for median absolute relative error, signed bias, worst overprediction factor, and leave-one-out ranges over at least four Fe-W references. The panel contains at most ONE signed error (alpha-Fe control, +6.7%). All four statistics are therefore undefined for Fe-W intermetallic chemistry.

Retargeted from ranking Fe-W prediction errors (none exist beyond the alpha-Fe anchor) to the structure that actually failed: the validated WFe2 C14 prototype, CIF 0db9981c-a143-4523-9fda-88b0de667edb. One matched pair, frozen v1 settings otherwise unchanged:

Per the 2026-09-07 checkpoint, reconciliation is narrowed to every completed run under the frozen route: the alpha-Fe positive control (action 01a07de3-7da0, route Ms 2.2967 T). No secondary-arm CHGNet runs from the revised item 6 exist yet. Inputs (from validated cell and returned per-site moments):