ExperimentMixed

P1: holding the last team through ball gaps barely moves the score

Keeping the last team for up to 1 s when the ball is missing lifts macroF1 a little and leaves spellF1 untouched. The reason is in the gaps: 88% of them last over five seconds, so a one-second hold covers only 5% of the missing frames.

baseline-holdv1spellF1 · public-5min0.0441= 0.0000 vs 0.0441baseline-holdv1macroF1 · public-5min0.4604▲ +0.0031 vs 0.4573

Context

The study's first idea, P1: no ball is not a dead ball. When the tracking loses the ball, keep the last team instead of inventing a stoppage.

Hypothesis, written before any real run: a short hold avoids artificial breaks, but may delay real turnovers. Limit fixed at 25 frames (1 s), not tuned on the public window.

What we did

A causal nearest-player reader with a hold: after the last valid team, repeat it for at most 25 frames, then answer X until new evidence. The state resets at each half, and a half that starts with no evidence starts with X.
No smoothing, so the only difference between the two runs is the hold.
A controlled comparison: hold 25 against hold 0, same commit, inputs and Docker environment, each run twice with identical output.
An aggregate diagnosis of the ball gaps over both halves.

python docker/pipeline.py modelo --config configs/baseline_hold.json --data-dir <local data> --repetir
python docker/pipeline.py modelo --config configs/baseline_hold_disabled.json --data-dir <local data> --repetir

Result

Public window (17 true possessions):

variant	spellF1	macroF1	predicted possessions	`X` frames
`baseline_smooth50` (reference)	0.3256	0.6207	26	0%
hold 0	0.0441	0.4573	128	11.8%
hold 25	0.0441	0.4604	128	10.4%

The ball gaps explain why:

40.7%of the frames in both halves have no ball

150gaps in the ball tracking

88%of them last more than five seconds

5%of the missing frames covered by a 1 s hold

A one-second hold fixes the short gaps, and there are very few of them. Almost all the missing ball comes in long stretches, which belong to P6.
The scorer already holds for us. It fills every X with the previous letter before counting possessions, so a hold and an abstention end up almost the same for spellF1.
Flicker is the real problem. 128 predicted possessions against 17 true ones caps spellF1 at 2 × 17 ÷ 145 ≈ 0.23 before a single possession is matched — the count ceiling from the study.

Not every frame without a ball needs fixing. The tracking flickers far more than it disappears.

Decision

Keep the hold as a reproducible building block. Do not replace baseline_smooth50 with it, and do not tune the limit on the public window. The comparison with smooth50 is not controlled: smoothing and masks change too.

Stability belongs to P4 (temporal decoding) and control events to P3; real stoppages to P2. Long gaps stay for P6.