The bispectrum, the three-point correlation of primordial fluctuations, is where early-universe models are supposed to part ways. Planck's final analyses constrain the amplitude in the standard templates to f_NL local = -0.9 +/- 5.1, equilateral = -26 +/- 47, and orthogonal = -38 +/- 24, with the PR4 reprocessing giving f_NL local = -0.1 +/- 5.0: consistent with zero in every shape. Scale-dependent extensions, in which f_NL runs with wavenumber, are far more weakly constrained; the data allow substantial running, some modal bins show mild excesses of no individual significance, and the constraining power degrades steeply away from the templates the estimators were built for.
The discomfort for ΛCDM is structural rather than statistical. Single-field slow-roll inflation predicts f_NL of order the slow-roll parameters, around 0.01, through the Maldacena consistency relation: a sharp, falsifiable target, but four orders of magnitude below Planck sensitivity. Multi-field, curvaton, and feature models can produce nearly any amplitude, sign, and scale dependence. The result is a sector where the standard model makes either a prediction it cannot test or no prediction at all: a null result excludes nothing that matters, and any future detection of scale-dependent non-Gaussianity could be retrofitted by some inflationary variant. The bispectrum, advertised as inflation's discriminating test, currently discriminates almost nothing.
The standing is a stalemate awaiting sensitivity: CMB-S4 and 21-cm surveys target sigma(f_NL) near 1, and galaxy surveys (DESI, Euclid, SPHEREx) probe scale-dependent bias signatures of local non-Gaussianity down to similar levels. Whether the three-point sector finally selects among origins models this decade depends on whether any framework on the table makes a specific, scale-resolved prediction narrow enough to kill.