Total: 1
Two recent independent analyses of Belle II $B^+ \! \to \! K^+ν\barν$ data yield limits on ${\mathcal B}(B^+ \! \to \! K^+ a)$ -- the two-body mode to a light invisible particle such as the QCD axion -- differing by a factor of roughly four; we trace this to the choice of kinematic variable space. The central figure of merit is the resolution in the reconstructed di-neutrino invariant mass $q^2_{\rm rec}$: fine-grained binning resolves the narrow axion signal, while coarse binning dilutes it into a background-dominated range. A BDT axis trained on $B^+ \! \to \! K^+ν\barν$ adds little discriminating power for $B^+ \! \to \! K^+ a$, as this axis is largely uncorrelated with $q^2$. These expectations are confirmed by a set of numerical tests. The subleading shape systematics omitted from our $q^2_{\rm rec}$-based approach {\em lower}, not raise, the $B^+ \! \to \! K^+ a$ limit: by better accommodating the $B^+ \! \to \! K^+ν\barν$ shape, they leave less room for the axion signal, making our $q^2_{\rm rec}$-based bound conservative, if anything. A dedicated reanalysis confirms that the kinematic-axes choice alone accounts for the factor-of-four sensitivity difference, and that the $B^+ \! \to \! K^+ a$ bound varies sizeably within the $q^2_{\rm rec}\timesη({\rm BDT}_2)$ space, depending on the SM-likeness of $B^+ \! \to \! K^+ν\barν$, thus losing the dual-probe feature of our $q^2_{\rm rec}$-based approach. These results point to a broader consideration: likelihoods dominated by BDT variables are of limited use for reinterpretations when the signal shape differs appreciably from the BDT's training signal. We therefore advocate that experimental collaborations publish likelihood projections in physical variable spaces alongside BDT-based likelihoods, to maximise the reinterpretability of their measurements.