The Regularizer That Switched Off the Experiment: Why Our Protein-Interaction Null Was About the Encoder, Not the Prior

Does a typed protein-interaction prior improve prediction of how primary human T cells respond to genetic perturbation? We report two failures, and the more useful one is not the one we went looking for. The first is an instrument failure that nearly published itself: a textbook edge-sparsity regularizer, written as an unnormalized sum over the edges of each sampled neighborhood, is roughly $103\times$ the response term at its default weight and drives every learnable edge gate to $\sim$$10^{-7}$ inside the first epoch, in all five seeds. Training completes and reports a plausible score, so a graph-versus-no-graph comparison silently becomes no-graph versus no-graph. The second is the null itself. With the gates demonstrably live, the evidence-gated graph is indistinguishable from an expression-only baseline ($\Delta$ systema $=-0.0009$, 95\% CI $[-0.0072,+0.0054]$, $n{=}5$, family-wise corrected), and the typed null holds across seven further Perturb-seq screens. The one contrast we were entitled to call significant then failed its own replication on a harder split. Of five candidate causes we adjudicate, the two that carry the argument are not about biology. Our baseline was never graph-free: zeroing its graph-derived features costs $-0.0036$ systema under correction, more than the message passing over eight million typed edges we set out to measure. And an arm we had never run overturned the reading: strip the typing and gating, keep the raw topology, and that plain arm beats the baseline on our own screen ($+0.0043$, $n{=}7$, surviving Bonferroni and Holm) and on Replogle RPE1 ($+0.0675$), and loses to it on Norman ($-0.0790$); three of eight datasets clear correction, in disagreeing directions. So the prior is not what fails but the encoder we built to exploit it, and a "graph versus no graph'' null can be an artifact of the encoder rather than evidence about the graph. The null is also instrument-bound: the gated arm it rests on clears at no injected signal up to $0.40$ response SDs, where our best graph arm clears at $0.02$. We close with the checklist we wish we had used.