Toledo Football IQ

Coaches and analysts correct play boundaries, formation labels, route labels, coverage labels, player identity, and metrics. Every correction becomes structured Toledo-specific training data. This is the competitive moat: private, corrected, domain-specific data beats any off-the-shelf model at Toledo football terminology.

The platform separates generic football labels from Toledo-specific call language. A concept can have a generic label like "mesh" while also mapping to Toledo's internal terminology. This keeps ML portable while making the product feel native to staff.

Each clip receives an analytics-safe score based on field visibility, homography stability, player scale, motion blur, drone drift, and occlusion density. Weak calibration suppresses precise speed, separation, and distance metrics before they can mislead coaching decisions.

Identifies what Toledo is unintentionally revealing: formation-to-play tendencies, motion-to-play tendencies, field-zone tendencies, personnel tendencies, and pre-snap tells. Shows what a prepared opponent would see.

Coarse, coach-useful measures from overhead footage: pad level, hip flexion, torso angle, pass-set depth, strike timing, shoulder-over-knee alignment, deceleration step count, hip sink depth, stride symmetry, QB shoulder-hip separation. These remain experimental until reviewed by the relevant position coach.

Longitudinal development profiles: roster bio, role summary, development goals, coach notes, best teaching clips, best recruiting clips, corrected metrics, position benchmarks, player-facing summaries, workload visibility controls, and weekly snapshots. This turns the platform into a player-development system, not just a film tagger.

Coaches select a clip and find similar reps based on movement, spacing, formation, route concept, or play embeddings (pgvector). Valuable when labels are incomplete or when scouting opponent concepts.

A later phase targets period-break feedback: clip-ready metrics to position coaches within 5–10 minutes between periods. Full sessions processed within 60 minutes of practice ending. Lightweight models run during practice; full-quality runs run overnight.

A coach who does not trust the system will not use it. Building trust — through evidence links, correction paths, confidence scores, and conservative suppression — is more important than adding advanced features early.

The coach-correction flywheel must exist before pose-lite biomechanics, similar-rep search, or zero-shot discovery. Without corrections, there is no Toledo-specific training data. Without Toledo data, the models are generic.

A model that reports confident metrics on a poorly calibrated clip is actively harmful. Suppressing metrics when calibration is weak is a feature, not a limitation.

Dynamic homography, camera drift, high-occlusion football scenarios, and helmeted players with similar uniforms make drone footage harder than broadcast footage. These constraints need to be part of the design, not discovered during development.

A system that takes 24 hours to process a practice session cannot change practice behavior. Same-session feedback — even lightweight — is worth more than overnight perfect analysis.

Frontier experiments — zero-shot discovery, quantum AI sandbox, advanced health fusion — need isolation from the production coaching workflow. Experimental outputs must be clearly labeled and role-gated.

Effort, load, and biomechanics data touch athlete welfare. Role-based access, conservative labeling, and a strict no-automated-medical-inference rule are non-negotiable from the start, not retrofitted later.