The reconstruction menu, and the gates that prune it

Mon, 31 Aug 2026 00:00:00 +0000

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.

The space of the Standard Model, and the space still open

Mon, 31 Aug 2026 00:00:00 +0000

A particle theory is not a list of Lagrangians. It is a tree of decisions: what is allowed now, what the data kill, and when a failure is not a death but a reason to grow a new degree of freedom.

The Standard Model is what that tree looks like after the decisions have already been made. Beyond-Standard-Model (BSM) physics is the same kind of tree, still growing. Oracle-Guided Triage and architecture Search (OGTS) is a proposal to write that tree as a searchable object, then spend computing time along it instead of scanning every leaf.

How to read the two chambers

Left: Standard Model. Constraint availability by era, not a history lecture. In 1933 Fermi’s four-fermion contact is admitted because it fits beta decay. The 1950s rank five Dirac bilinears; angular data mildly prefer scalar and tensor couplings. Wu’s parity-violation experiment flips that ranking: S, T, and P are hard-pruned, V−A is admitted. When the contact term violates unitarity at high energy, the branch does not die. It expands: a massive vector W. The 1970s admit SU(2)L × U(1)Y with spontaneous symmetry breaking, prune a pure SU(2) without hypercharge, and keep colour as a separate sector. Minimal SU(5) still looks healthy at triage and dies at verify (proton lifetime).

Right: BSM futures. Same schema, experimental grain. The root is the radiative family b → sγ. An EFT child is the Wilson coefficient C7 (operator O7). Its exclusive children, Bs → φγ (active) and Λb → Λγ (conditioned), are siblings: same grain, not parent and child of each other. A Type-II two-Higgs-doublet model with a light charged scalar is a UV child of the family, not of either exclusive mode. Inclusive B → Xsγ verify-kills it. A photon in some other radiative decay does not make that decay an O7 child.

Edges are refinement or UV completion. They are never “similar final state.”

What “computational allotment” means here

OGTS is two loops with a human gate.

The outer loop ranks hypothesis nodes. Cheap triage asks whether a term is even admissible (Lorentz, anomalies, parent still alive). Expensive verify asks whether it matches this era’s data. Expand happens on a structural FAIL such as unitarity: spawn a child with a new field, do not empty the branch. Status is part of spend: active nodes can take compute; conditioned nodes wait on a reconstruction or systematic gate; pruned nodes get zero. If O7 is excluded, exclusive children leave the ranked list. That is the graph doing work a flat scan cannot.

The inner loop is a different object: a named reconstruction menu pruned by throughput and resolution. That family tree does not live on this diagram. This page is only the outer loop, hypothesis space and allotment. A vertex-less radiative photon is a harder calorimeter object than an inclusive cluster; that fact conditions Λb → Λγ here.

Track thickness on the BSM side is a schematic share for this demonstrator. It is not a published OGTS ranking, and it is not a claim that the search already beats legacy allocation. Stage 0 on the left is the calibration problem whose answer is already known. The open question is whether the same grammar transfers to LHCb channels and to SMEFT.

The human gate keeps two decisions: what counts as a legitimate node, and what counts as an acceptable model of systematic uncertainty. The Λb → Λγ 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.