# Charged pT from a tracker fit with a learned correction Enable these set-decoder settings on a hit model. The example uses the CLD 2 T field; for CLIC set `tracker_group_magnetic_field: 4.0` explicitly: ```yaml query_init: input-conditioned hit_grouping: true tracker_group_readout: true tracker_fit_correction: true tracker_fit_max_log_correction: 0.4054651081081644 # log(1.5) ``` This is an opt-in branch inside the existing particle-set model, not an external tracking stage. Defaults and the legacy tracker-anchor mode are unchanged. The new branch adds a zero-initialized linear correction head and requires training (or fine-tuning with explicit handling of its new weights). Existing checkpoints remain compatible with their original configuration. ## Prediction The existing group readout fits tracker coordinates with a centered, scaled transverse-circle estimator in FP64 and three Huber reweighting iterations. It converts the radius to pT using the configured magnetic field (2 T by default). The fit requires at least four finite hits, three detector surfaces, a radial span of 0.1 m, sufficient conditioning, relative residual below 0.02, and pT between 0.1 and 10,000 GeV. These gates reject underconstrained or inconsistent groups; they do not guarantee that a mixed group belongs to one particle. Nonfinite coordinates are excluded. For an eligible tracker slot predicted as charged hadron, electron or muon: ``` log(pT) = log(pT_fit) + max_log_correction * tanh(correction_head(slot)) ``` The default bounds pT / pT_fit between 2/3 and 1.5. At initialization, the correction is zero and the output equals the fit. The correction head and final addition use FP32 even under BF16 autocast. The head receives the final normalized slot, which includes detector embeddings and fit summaries. Physical PID selection matches inference (argmax over classes 1–5, excluding the empty class). Neutral predictions, invalid fits and slots with existing calorimeter anchors retain the ordinary scale readout. Energy and direction are always learned by the ordinary heads in this mode; this first version does not fit a longitudinal slope, derive energy from a mass hypothesis, or model material effects. A bounded correction cannot rescue a grossly wrong fit that passes the gates. Charge/PID mistakes can route a slot to the wrong branch; monitor species efficiencies as well as momentum residuals. The same head is used for intermediate auxiliary predictions. The geometry fit and hard membership remain detached; the correction and ordinary heads train jointly with the existing losses, and hit ownership learns through its supervised grouping objective. This is not a differentiable track fitter. `tracker_fit_fraction` and `tracker_fits_per_event` count eligible geometrical fits after excluding slots with calorimeter anchors and before the charged-PID gate, not the number of final charged predictions that use them. ## Comparison `configs/training/scenarios/cld_pf_heptv2_fit_correction_50k.yaml` defines three 50k-step runs: ordinary grouped regression, legacy tracker anchors, and the new bounded charged-pT correction. Use the normal scenario runner. The correction limit is an initial hypothesis, not a tuned resolution guarantee. Evaluate fit coverage, charged-particle efficiency, uncensored momentum tails, and central residual widths versus truth pT and eta. Compare with truth-owned hit fits before interpreting limitations as detector resolution. No training result is implied by the synthetic fit and forward/backward tests. The completed 50k comparison found small width gains accompanied by lower charged efficiency and higher tails. This branch remains experimental; see [the evidence summary](cld_hit_results.md) for matched results.