Gamma-ray bursts split into two categories: those from collapsing massive stars (collapsars) and those from merging neutron stars. The traditional separator is duration — long bursts are collapsars, short bursts are mergers. This heuristic works most of the time but fails on ambiguous cases, and the boundary is not clean.
This paper builds a classifier on 24 GRBs with spectroscopically confirmed progenitors — the gold standard, where the mechanism is known from host galaxy and afterglow analysis, not inferred from the burst itself. The classifier uses peak energy, isotropic energy, and duration. The finding: peak energy carries approximately 5× the discriminative power of duration.
Spectral hardness — how energetic the typical photon is — is the primary physical signature. Mergers produce harder spectra relative to their total energy output. Collapsars produce softer spectra. The same total energy, distributed differently across photon frequencies, reveals the engine.
The result reframes the classification problem. Duration was never the physical variable — it was a proxy that happened to correlate with the real discriminant. The short/long classification worked because mergers tend to be both shorter and harder, and collapsars tend to be both longer and softer. But when the two properties disagree, hardness wins. The spectrum is the fingerprint; duration is the shadow.