xcross-attempt-v1

xCrossAttempt v1 (sc_extended) — Cross-Attempt Propensity from Tracking State

Read this first. This repo serves the sc_extended (owner-tier) variant. It is NOT bundled with the silly-kicks wheel because it is trained on restricted owner-tier data that cannot be redistributed inside a PyPI package — a licensing constraint. Only the learned parameters are published here. If you do not have owner-tier access, use the bundled default variant instead — see Which variant should I use?.

Model Description

XCrossAttemptModel is a deterministic-XGBoost classifier estimating P(the in-possession team attempts a cross within ~1 s of a tracking frame) — a STATE-anchored framing, reframed from the sender-level event treatment in Cao et al. (arXiv:2505.11841).

It carries 7 of the paper's 8 confounders (crosser-position, #7, is omitted — no faithful tracking-only proxy) plus a novel, isolatable GK-position confounder block, which is the paper's headline gap and the reason this model exists inside the GKDV research arc (TF-17 → TF-19).

• 16 features, faithful (velocity-bearing), goal-relative coordinates via the shared _geometry helper
• Domain filter: alive-ball, wide-area
• The sc_extended model is fit on the owner-tier corpus (IDSSE + SkillCorner incl. owner-tier); base rate 5.0% positive.

Which variant should I use?

Why this variant is HF-only

sc_extended is HF-only for licensing, not quality: it is trained on restricted owner-tier SkillCorner data that cannot be redistributed inside the PyPI wheel (ADR-038). Only learned parameters are published here — no raw provider tracking data (only split thresholds, feature indices and leaf values are stored — no per-sample training data).

The Hub sc_extended repo is the owner-tier archive: it holds the owner-tier model independent of the wheel-bundle selection gate, which decides only what ships in the wheel (ADR-071). This artifact was produced with that operator override (--ship-variant sc_extended); the gate's verdict and per-fold deltas are recorded in metrics.json (candidates.paired). training_commit: 4095251.

Held-out CV (5 folds, out-of-fold)

All four acceptance gates pass (enough_usable_folds, pr_auc_gt_base_rate, brier_lt_base_rate_brier, log_loss_lt_uniform). Estimates are CV, not the shipped fit.

TF-19 GK-substitution probe

The frozen GK-substitution probe (gk_substitution_probe in metrics.json; 200 frames; ADR-037's two-prong gate — ratio ≥ 2.0 × the nearest-defender control and an absolute floor ≥ 0.01):

The GK-block ablation shows the GK confounder block does carry signal (removing it drops held-out PR-AUC by 0.0089), but the substitution probe does not clear the frozen gate. Per ADR-037 this is a gated_clean_fail — TF-19 routes to GK feature engineering, explicitly not "no signal." Do not build a TF-19 consumer on this surface. (The position-only sibling — xcross-attempt-position-only-v1 — does clear the gate; see its card.)

Usage

from silly_kicks.tracking import XCrossAttemptModel

model = XCrossAttemptModel.from_variant("default")       # recommended, bundled, offline
model = XCrossAttemptModel.from_variant("sc_extended")   # this repo, downloads from the Hub

Requires pip install silly-kicks[xcross] and silly-kicks >= 4.74.0 (the sc_extended_position_only sibling repo requires >= 4.94.0, which introduced its variant key — ADR-070).

The >= 4.74.0 floor is a hard requirement. These weights are on the corrected goal-relative transform (geometry_version: goal-relative-2); ADR-051 found the previous transform was chiral (an x-only mirror at one goal end, identity at the other), so one physical scene scored differently depending which end the attacking team attacked. load()'s feature-contract prong is fail-closed, so an older silly-kicks refuses these weights with IntegrityError. from_hub() takes no revision argument yet, so treat the library version as the pin; prior revisions are addressable by commit SHA.

Integrity and load-time guards

load() is fail-closed on two independent checks: (1) SHA256SUMS verified before anything is parsed; (2) chirality fingerprint (ADR-040) — the model re-runs its own outputs on a fixed y-asymmetric probe frame and compares to the recorded fingerprint, raising on a mismatch and on a missing one. A base_score guard handles the xgboost 3.x bracketed-string serialization that 2.x silently drops to 0.5.

Limitations

• Not the bundled model (restricted corpus). This is a redistribution limit, not a performance one.
• tf19_ready = false (see the TF-19 section) — do not build a TF-19 consumer on this surface.
• Trained on an owner-tier corpus that is heavily one-club (the 98 owner-tier additions are a single club), so club/style confounding is real and unquantified here.
• SkillCorner keepers are detected in only ~19.6% of frames (~80% interpolated), which is why GKDV measurement is registered to Gradient Sports frames only (ADR-038 §5).
• Estimates are cross-validated, not a held-out test of the shipped fit.

References

See the NOTICE file in the silly-kicks repository for full bibliographic citations.

• Cao et al. "Framing Causal Questions in Sports Analytics: A Case Study of Crossing in Soccer." arXiv:2505.11841 (2025).
• Decisions: ADR-011 (trained-model lifecycle), ADR-015 (causal-validation port), ADR-037 (TF-19 re-gate), ADR-038 (corpus + visibility), ADR-040 (chirality enforcement), ADR-070 (position-only Hub variant), ADR-071 (owner-tier archive).

Model Files

More Information

https://github.com/karsten-s-nielsen/silly-kicks

Wheel sibling (4.134.0)

In silly-kicks 4.134.0 the wheel's bundled default was re-fit on the corrected floor velocities (TF-65 / ADR-120, training_commit=a34e992). This Hub owner-tier sc_extended artifact is re-fit on the same corrected floor velocities this cycle (silly-kicks 4.138.0): predictions change within noise from the prior velocity-stale artifact (PR-AUC 0.189). No label defect — xCross was always type_id-filtered. The re-fit artifact is pushed by the ADR-088 model-publish step.