A developing Drosophila embryo has a problem. Each nucleus needs to know its position along the body axis. It reads this from morphogen gradients — concentration fields that encode location. High bicoid means head; low means tail. The nucleus counts binding events, builds a concentration estimate, and decides its fate.
The physical limit on this measurement is the Berg-Purcell bound: sensing accuracy is proportional to the square root of detected molecules. For the concentrations involved, this means waiting minutes to hours for enough binding events to distinguish neighboring positions. But Drosophila nuclei commit to their fates roughly forty times faster than this limit predicts.
The standard interpretation is that some unknown amplification mechanism exists — a biological trick that somehow circumvents thermodynamic constraints. The actual answer (arXiv:2602.00261) is simpler and more interesting: the cells aren't measuring a concentration. They're predicting a trajectory.
Morphogen gradients build up over time. The concentration isn't static — it follows a structured profile from zero to steady state. Crucially, the rate of approach depends on the eventual steady-state value: positions with higher final concentrations see faster initial buildup. The parameter a ≈ 1.8 for bicoid means that a position heading toward twice the final concentration also reaches it nearly four times as fast in the early phase.
This correlation is information. The first few binding events in the rising phase encode not just the current concentration but the slope — and the slope predicts the endpoint. A maximum a posteriori estimator using this trajectory structure achieves sensing precision of δc/c = 1/(a²√N), which is a factor of a² better than the classical Berg-Purcell limit for the same number of binding events.
The mathematics is Bayesian, but the implementation is molecular. Constitutive production gives the cell a clock. Dimerization performs multiplication. Enzyme kinetics take square roots. The entire prediction circuit is physically embodied — no central processor, no digital computation, just the natural dynamics of molecular interaction computing a Bayesian estimate.
What the cells discovered is that the distinction between measuring and predicting is illusory when the signal has structure. Measuring a random variable requires enough samples to average out noise. Predicting a structured trajectory requires far fewer samples because each data point constrains the model. The early observations don't just reduce uncertainty about the current value — they collapse uncertainty about the entire future.
The Berg-Purcell limit assumed the worst case: an unknown, static concentration. It calculated the cost of ignorance. The cells pay less because they know more — not about the specific concentration, but about the shape of all possible concentrations. The prior is the shortcut.