Down-Type Jet Identification in Fully Hadronic $t\bar t$ Events

Abstract

Fully hadronic $t\bar t$ events carry the largest branching fraction and provide the down-type quark as a near-maximal spin analyzer ($\beta_d\simeq1$), yet they are rarely used for spin-correlation and entanglement studies because the down-type jets must be extracted from a difficult jet-to-parton reconstruction in the presence of a large QCD multijet background. We develop a two-stage machine-learning reconstruction for this channel: a GNN+Transformer network assigns all reconstructed jets into a legal six-jet top-pair candidate, and a second hybrid classifier resolves the four down/up hypotheses inside the two hadronic $W$ decays. The second stage combines an assignment scorer with an auxiliary conditional diffusion head; we find that the diffusion assignment score can rank hypotheses only when trained with a margin ranking objective, while standard denoising objectives leave it at the random baseline, and the two objectives are largely decoupled. We evaluate the method against a calibrated QCD background and an unmatched $t\bar t$ component using both reconstruction and spin observables. Compared with a ST1 only benchmark with random down/up labels, the learned ST2 classifier improves both the effective spin-analysis factor and the purity-weighted six-parton exact fraction at fixed event selection. A direct check of the standard spin-correlation coefficient $D$ further gives an ML-reconstructed value compatible with the truth-level value within the resampled spread. These results show that down/up identification can retain useful spin-correlation information in the all-hadronic channel.