Exact Inference in Fixed-Effect Regressions with Concentrated Identifying Variation

In saturated fixed-effects regressions, Gaussian inference depends not on total identifying variation but on its concentration, measured by the self-normalized leverage $\lambda_n$ of the residualized treatment. When finitely many score weights remain persistent, the $t$-statistic converges to a convolution of raw errors and a Gaussian component. At full concentration, its null distribution varies across symmetric error laws with equal variance, so no fixed critical value is uniformly valid. We instead construct nuisance-annihilating contrasts from the design alone. These eliminate the fixed effects identically and yield finite-sample exact sign-flip inference under symmetric, arbitrarily heteroskedastic errors, with no homogeneity assumptions or restrictions on the fixed-effect dimension. In two-way designs, admissible contrasts form the cycle space of the observation multigraph. Their efficiency is summarized by an observable capture ratio $\kappa$, which equals Pitman efficiency. The resulting design problem involves a capture--granularity trade-off: coarse supports maximize capture but reduce the number of randomization signs. Cycle packing provides sufficiently granular supports. On matched employer--employee data, a structure-exploiting algorithm achieves $\kappa \approx 0.51$, compared with $0.26$ for naive packing. In the Grunfeld investment regression, realized score concentration is $0.739$, corresponding to $N_{\mathrm{eff}}^{\mathrm{score}}=1.80$, while $32$ valid supports attain $\kappa=0.627$. The resulting exact $95$ confidence interval is $[0.150,0.450]$. A worker--firm application demonstrates scalability to large networks.