silly-kicks/xcross-attempt-v1

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

Read this first. This repo serves the sc_extended variant, which is NOT bundled with the silly-kicks wheel — it is trained on restricted owner-tier data and cannot be redistributed there. That is a licensing constraint, not a quality one: sc_extended cleared its pre-registered paired test. Pick the variant that matches your situation — 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, goal-relative coordinates via the shared _geometry helper
• Domain filter: alive-ball, wide-area
• Trained on 1,209,333 rows / 39,766 positives (5.0% positive rate)

Which variant should I use?

Variant	Corpus	Where it lives	Use it?
default (public)	17 matches — SkillCorner + IDSSE, redistributable	bundled in the wheel	Default choice — offline, fully reproducible, no restricted data
sc_extended	+ 98 owner-tier SkillCorner matches (179 total)	this repo (HF-only)	Yes, if you can accept a Hub download and the corpus caveats below — it beat public on held-out folds

sc_extended is HF-only because it is trained on restricted data and cannot be redistributed inside the PyPI wheel. Only learned parameters are published here — no raw provider tracking data, and no artifact from which raw rows can be reconstructed (every booster leaf aggregates ≥ 14 samples by the binding min_child_weight floor).

Why this variant is HF-only (it did NOT fail its paired test)

sc_extended CLEARED the pre-registered paired test. It is HF-only because it is trained on restricted owner-tier data and cannot be redistributed inside the PyPI wheel (ADR-038) — a licensing constraint, not a quality verdict.

The Stage-B run's own record:

"shipped": "sc_extended",
"why": "sc_extended clears; full does not dominate it -- ties go to less data"

Stage-B deltas vs public, held out on public folds: [-0.008, +0.009, +0.005, +0.029, +0.023] — positive in 4 of 5 folds, mean +0.0117. The registered rule (scripts/_paired.py::clears_rule: positive in ≥ K−1 of K folds and a positive mean) returns True.

Do not misread the BUNDLED artifact's paired block

The bundled default artifact carries a different run's record — Stage A, whose corpus contained no owner-tier SkillCorner rows at all. There, cand_masks["sc_extended"] = is_public | is_sc_private with is_sc_private empty, so the sc_extended candidate was the same mask as public: the same model scored against itself. Its deltas are [0.0, 0.0, 0.0, 0.0, 0.0] by construction — a degenerate self-comparison, not a measured null — and since the rule requires strictly positive values, zero fails and the fixed sequence stops. That is why public is the bundled default.

Reading those zeros as "sc_extended failed" is a trap; an earlier revision of this card fell into it. docs/research/tf19_pr2/decision_table.md has the Stage A / Stage B split correct.

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

Metric	Value	Baseline
PR-AUC	0.1882 (± 0.0102)	base rate 0.0500
Brier	0.0437	base-rate Brier 0.0475
Log loss	0.1659	—

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.

The TF-19 result this variant exists to support

The frozen GK-substitution probe (TF19_PROBE_RATIO=2.0, TF19_PROBE_ABS_FLOOR=0.01, both frozen before the run; tf19_ready requires both prongs):

Metric	bundled public	this sc_extended
gk_median_abs_delta	0.002417	0.009697
nearest_def_median_abs_delta	0.001718	0.004380
ratio (gk / nearest_def)	1.41× — misses	2.21× — clears
abs floor ≥ 0.01	4.1× short	misses by ~10% relative
tf19_ready	false	false

Both models fail the gate. They fail differently: the bundled model misses both prongs by a wide margin; this one misses a single prong by ~10%. Per ADR-037 §4 the verdict is gated_clean_fail, which routes TF-19 to GK feature engineering — explicitly not "no signal."

Probe run out-of-fold on held-out Gradient Sports matches 10502 / 10503 (probe_sample_in_training_folds records false for both).

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 ≥ 4.51.0 (earlier versions have no sc_extended routing).

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. It raises on a mismatch and on a missing one, because every pre-PR-2 artifact belongs to the y-mirrored mis-served class.

A base_score guard also handles the xgboost 3.x bracketed-string serialization that 2.x silently drops to 0.5.

Limitations

• Not the bundled model (restricted corpus, so it cannot ship in the wheel). This is a redistribution limit, not a performance one.
• tf19_ready = false. Do not build a TF-19 consumer on this surface.
• Trained on a corpus that is 179 matches, heavily Real Madrid — the 98 owner-tier additions are one 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).

Model Files

File	Purpose
model.json	XGBoost booster (pickle-free)
metadata.json	features, hyperparameters, chirality fingerprint, provenance
metrics.json	CV metrics, GK-substitution probe, ablation, permutation importance
SHA256SUMS	integrity manifest, verified by load()

More Information

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