The reconstruction menu, and the gates that prune it

Reconstruction is detector-wide. Every event that enters LHCb must be clustered and written under a latency and throughput budget that tightens at Upgrade II. The architecture that meets those constraints is a searchable object, not a per-analysis add-on.

This is the inner loop. It does not live on the hypothesis-space diagram. That page ranks what is still worth measuring. This page ranks how to extract the object, under gates.

How to read the family tree

The menu is short and named. We do not enumerate every network.

Left: classical replacement. The cellular automaton is on the tree because Graph Clustering replaced it. The dashed edge is that replacement. Graph Clustering is the published Run 3 gold standard (Valls Canudas et al., Eur. Phys. J. C 83, 169 (2023)). At high occupancy it needs offline calibration and does not learn overlapping showers. Upgrade II can condition it. Conditioned is not dead.

Right: learned families. Graph nets (GarNet), transformer-style calorimeter reconstruction (UzzieNet), deep sets, and state-space models. UzzieNet is a refinement of GarNet: node-centric linear-cost attention. Distillation varies only inside a family, under an explicit HLT1 latency cut. A student can pass a throughput gate the teacher fails. Deep sets and state-space models are siblings on the menu, not children of GarNet. An unscored family stays amber. The gate does not invent a prune.

Edges are replacement, refinement, or inside-family compression. They are never “also a calorimeter network.”

The two gates

Throughput. Offline is loose. Run 3 HLT1 is the Allen path we actually have to live with. Upgrade II is tighter still: occupancy, PicoCal geometry, timing layers. A family that cannot meet the budget is conditioned or gone, even if its resolution is beautiful overnight on a GPU.

Resolution. Inclusive energy is one demand. Overlapping showers are another. A vertex-less radiative photon, the \(\Lambda_b\to\Lambda\gamma\) object, is harder than an inclusive cluster. Flip that topology and the same menu is re-coloured. The architecture that wins on a given topology is a diagnostic of problem structure, not a trophy.

Gate colours on this page are schematic. CHEP 2026 numbers that are real stay in the inspector: UzzieNet up to 8× faster than message-passing; Graph(GarNet)→MLP about 95% smaller and beating the teacher on energy resolution; GarNet ONNX up to 5× CPU and about 2× GPU at FP32 parity near \(10^{-7}\). Those are measurements against teachers and runtimes, not a published claim that the search already beats Graph Clustering at Upgrade II.

Two oracles, one human gate

The cheap oracle is a surrogate under FLOP and wall-time constraints. It is a filter, not calorimeter truth. The expensive oracle is Allen-scale PicoCal reconstruction. If they disagree, the lie is published; the ranking that counts is Allen’s.

The human gate keeps what counts as an acceptable model of systematic uncertainty. The \(\Lambda_b\to\Lambda\gamma\) revival found an efficiency-ratio discrepancy from two compensating errors whose branching ratio still looked fine. A beautiful excess that is two mistakes does not pass.

This page is an AI-assisted first version of the post.