Analysis results

Written as correlated 1σ errors, Eq. (1) of the paper reads

δm² / 10⁻⁵ eV² = 7.48 ± 0.10, sin²θ₁₂ / 10⁻¹ = 3.085 ± 0.073, ρ = −0.20,

for all oscillation constraints available as of 2025. The errors are, to a very good approximation, symmetric and linear in the number of standard deviations — which is why the 2σ and 3σ ranges above are so close to simple multiples of the 1σ one, and a departure from that pattern in a future release would itself be news.

The update is dominated by JUNO. The paper's own intermediate row, with SNO+ added but not JUNO, still gives δm² = 7.44 at 2.1 %: it is JUNO that takes the accuracy on δm² from 2.3 % to 1.3 %, while the remaining data shift the best fit slightly and reduce JUNO's own errors by a factor of about 0.84.

How far the fourth digit means anything. The paper is explicit that subleading effects it does not zero out can move the best fits at the level of a few parts per thousand — the fourth significant figure of sin²θ₁₂, sometimes the third once the fourth is rounded. Four digits are quoted so that the 1σ error can carry two, not because the last one is determined.