TELLAUDIT REPORT № TA-2026-0042
RHP №21 · bullpen audit
- PITCHES
- 42 (24 FB / 18 CH)
- CAMERA
- SIDE VIEW · 30FPS
- SESSION
- APR 18, 2026
- STATUS
- FINAL
An opponent watching nothing but your glove can call fastball-or-changeup 76% of the time, on pitches the model never trained on. One cue carries all of it, it's visible from the on-deck circle, and it has a two-week fix.
What an opponent's model would score
| CUE | PAIR | TRAIN ACC | HOLDOUT ACC | AUC | P | VERDICT |
|---|---|---|---|---|---|---|
| Glove drive (hand break) | FB / CH | 78.6% | 76.2% | 0.91 | <0.01 | CONFIRMED |
| Leg-lift tempo | FB / CH | 64.3% | 57.1% | 0.61 | 0.11 | WEAK |
| Pre-set wiggle | FB / CH | 57.1% | 52.9% | 0.58 | 0.31 | CLEAN |
| Hand-break timing | FB / CH | 55.0% | 51.4% | 0.55 | 0.44 | CLEAN |
| Head position | FB / CH | 52.4% | 50.0% | 0.52 | 0.62 | CLEAN |
| Glove set height | FB / CH | 51.2% | 48.6% | 0.50 | 0.71 | CLEAN |
BASELINE = 50% COIN FLIP · HOLDOUT = 14 PITCHES THE MODEL NEVER SAW · PERMUTATION-TESTED
Tell № 01 · glove drive at hand break
Your glove drives 2.2 inches past its fastball set at hand break, on 16 of 18 changeups in this session. Measured reach on the heater, overdrive on the change.
CONFIRMEDOPPONENT'S VIEW
"Glove punches long, sit changeup. Reads from the on-deck circle, no binoculars needed. Relay is one word."
THE FIX · ONE-PATH DRIVE
One glove path to one finish spot, the same reach on every pitch. Mirror work ×20 daily for two weeks, then a 30-pitch re-audit. Target: holdout accuracy back under 60%.
AUC 0.91 · p<0.01 (PERMUTATION, N=2000) · TRAIN 78.6% (n=28) · HOLDOUT 76.2% (n=14)
Watch item · leg-lift tempo
Lift-to-break ran ~80ms slower before changeups in training data, but the edge didn't survive the holdout (57.1%, p=0.11). It's suggestive and still unproven. We'd need more pitches before we'd call it actionable, and we'll say so plainly. We track it across your next session; if it's real, more pitches will say so.
WEAKAUC 0.61 · p=0.11 · BELOW THE CONFIRMATION BAR BY DESIGN
Tested clean
Glove set height
✓ NO SIGNAL
Pre-set wiggle
✓ NO SIGNAL
Hand-break timing
✓ NO SIGNAL
Head position
✓ NO SIGNAL
Four of six cue families came back indistinguishable from a coin flip. Those are real results too. They tell you where not to spend practice time.
Methodology
SPLIT: 28 train / 14 holdout, stratified by pitch type. Reported accuracy is holdout-only.
SIGNIFICANCE: permutation test, 2,000 shuffles, against the 50% baseline. Below the bar → WEAK or unreported.
POSE PIPELINE: markerless keypoints from your video; cues are interpretable geometry (heights, timings, widths), never a black box.
WRITE-UP: the narrative is generated strictly from validated findings. The language model cannot add a tell, and that's enforced in code.
THIS REPORT, ON YOUR PITCHER
First audit is free for pilot facilities
Ten slots. Film a bullpen this week, have this report by the weekend.
Claim a pilot slot →GENERATED BY `tellaudit demo` · SYNTHETIC PITCHER · NO PLAYER PERMISSIONS NEEDED