SnatchPhysics Academic Evaluation White Paper

SnatchPhysics is a desktop academic evaluation prototype for reviewing Olympic weightlifting movement from uploaded video. The first focus is the snatch, but the long-term framework is designed for movement-based sports such as sprinting, baseball, softball, throwing, jumping, and other sport-specific movement templates.

The purpose of SnatchPhysics is not to replace a coach, force plate, 3D motion-capture laboratory, or formal biomechanics validation study. The purpose is to create a transparent review environment where video evidence can be measured, annotated, interpreted, compared, exported, and preserved over time.

The current Academic Evaluation version separates the workflow into five layers:

1. Video Evidence
2. ATB Measurement
3. Coach Review
4. Experimental AI Interpretation
5. Dual Review and Recursive Learning Records

Core principle: ATB is evidence. Coach is judgment. AI is experimental interpretation.

The original product design contained many SaaS-style features and experimental interfaces. For academic review, the system needed a stricter workflow to answer a more serious question:

Can video-derived movement signals be organized into a useful, honest, reviewable teaching and evaluation environment for coaches and sport-science reviewers?

To answer that, the portal needed:

• desktop-only controlled upload flow
• clear measurement boundaries
• camera-confidence disclosure
• validity gates
• no hidden dummy metrics
• no silent fallback values
• coach review separated from AI interpretation
• exportable evidence
• engineering ledger transparency
• a foundation for longitudinal learning

SnatchPhysics uses an academic layering model to avoid mixing measurement, human interpretation, and AI interpretation.

Video Evidence

The video is the source evidence — the raw movement record uploaded by the reviewer. It should answer what clip was reviewed, not whether the athlete was good or bad.

ATB Measurement

ATB Measurement is the measurement layer. It uses video-derived pose tracking and movement signals to produce structured outputs such as pose/measurement-confidence proxy, phase segmentation, 2D projected joint angles, wrist/grip bar-path proxy, timing metrics, diagnostic velocity proxies, unavailable/N/A metric flags, validity state, and measurement warnings. It should answer what the system measured or estimated from this specific video, not the final coaching verdict.

Coach Review

Coach Review is the human judgment layer. The coach records observations, agreement or disagreement with ATB outputs, video-quality limitations, technical priorities, coaching cues, review confidence, and session notes. The human coach remains the expert.

Experimental AI Interpretation

Experimental AI is an optional interpretation layer. It does not create measurement data. It reads completed ATB Measurement outputs and translates them into possible coaching language, disclosing validity state, measurement confidence, N/A metrics, proxy metrics, suppressed metrics, and whether the result is VALID, LIMITED_DIAGNOSTIC, or REJECTED_ACADEMIC. It is not a coach replacement.

Dual Review

Dual Review places ATB evidence, coach review, and AI interpretation side by side to identify where coach and AI agree, where they disagree, what needs human review, which measurements were limited, and what should be tracked over time.

After building the first Academic Evaluation version, SnatchPhysics was audited as a scientific instrument. The audit asked:

• Are any metrics dummy, simulated, hardcoded, or placeholder?
• Are any missing measurements silently replaced with elite-reference values?
• Are proxy metrics labeled as proxies?
• Does AI know the measurement validity state?
• Do exports preserve provenance?
• Can this become a recursive learning center and dataset bank?

The initial audit score was 53 / 100. The architecture was promising, but the measurement layer still contained prototype shortcuts that could not remain in an academic review path. The score was not used for marketing — it was used as an engineering discipline tool.

Guiding rule: No academic metric, graph, score, AI critique, or export may use dummy, placeholder, hardcoded, simulated, silent-fallback, or elite-reference-substitution values unless clearly labeled demo/research-only and excluded from evaluation.

Key remediation actions included:

• removing the synthetic MoveNet server stub
• returning NO_POSE_DATA when client pose data is missing
• removing hardcoded lockout symmetry
• removing silent elite-reference fallback substitution
• labeling velocity and acceleration as timing-derived proxies
• disabling hip-only COM output for academic review
• marking Power and Force as not computed
• sharing validity rules between ATB and AI systems
• improving PDF/CSV/JSON export provenance
• including Coach Review in exports
• locking down lifter history RLS
• adding analysis_versions for append-only measurement snapshots
• adding sessions, session_lifts, and movement_templates foundations
• adding a Scientific Status page
• adding an Engineering Ledger

Post-sprint estimated score: approximately 82 / 100, pending final re-audit. This does not mean SnatchPhysics is a validated laboratory instrument. It means the system has moved closer to being an honest academic teaching and review prototype.

Every metric should be one of: measured, estimated, proxy, experimental, unavailable, or suppressed.

• Measurement Confidence™: Pose / Measurement Confidence Proxy
• Bar Path: Wrist / Grip Proxy
• Angle HUD: 2D Projected Angles
• Velocity / Acceleration: Timing-derived proxy until frame-differential calculation is implemented
• COM: Unavailable until full weighted segment-model COM is implemented
• Power / Force: Not computed
• Lockout Symmetry: N/A until pose-derived computation is implemented

This language is intentionally conservative.

Every analysis receives a validity state.

VALID — the system produced real measurement outputs under acceptable video and analysis conditions.

LIMITED_DIAGNOSTIC — the system produced real outputs, but camera angle, visibility, calibration, proxy quality, or missing metrics reduce confidence. AI interpretation may run with caveats.

REJECTED_ACADEMIC — the system blocks scoring, AI interpretation, or comparison when critical measurement requirements fail (missing pose data, placeholder engine, failed analysis, synthetic data, no valid keypoint frames).

These validity states protect reviewers from treating weak data as strong measurement.

Best measurement comes from best video setup.

The current camera-confidence value is a pose/measurement-confidence proxy based on how well the system can see and track usable movement signals. It is not a physical camera sensor measurement.

Measurement Confidence™ can be affected by camera angle, framing, lighting, occlusion, full-body visibility, bar/hand visibility, motion blur, view type, and scale/calibration quality.

A camera-limited result should not be treated as a judgment of the athlete's true ability. A world-class athlete can produce a limited diagnostic result if the video is not suitable for measurement.

SnatchPhysics treats every uploaded lift as a complete scientific observation, not merely a media file.

A scientific observation has four components:

1. Athlete — who was filmed.
2. Movement — what was performed.
3. Recording Environment — how the evidence was collected.
4. Measurement Results — what the ATB engine produced from that evidence.

Capture Metadata™ records the recording environment: camera model, device type, resolution, frame rate, duration, codec, container, orientation, camera distance, camera height, camera angle, tripod usage, drone usage, indoor or outdoor environment, lighting conditions, background complexity, and capture notes. Technical properties are read automatically from the video file wherever possible; the remaining fields are entered by the user and stored as part of the permanent research dataset.

Why this matters:

• Two identical lifts recorded under different conditions can produce different measurement confidence.
• Future researchers can correlate recording conditions with measurement repeatability, pose stability, and diagnostic signal quality.
• Capture Metadata is preserved in the lift record and in every immutable analysis_versions snapshot.

Reusable Capture Profiles™ reduce repetitive data entry by storing common recording setups (for example, "Garage Gym – Canon 1080p60"). Auto-detected video metadata always overrides stored values when appropriate.

A video records a performance. Capture Metadata records the conditions under which that performance became scientific evidence.

This design decision is recorded in Engineering Ledger Entry #010.

For the Academic Evaluation version:

• MoveNet is the academic baseline engine.
• YOLOv8 Pose is Research Lane only until tested and validated against MoveNet and coach-reviewed reference clips.

This avoids mixing unvalidated engine outputs with the official academic measurement layer. Research Lane outputs must be labeled research-only, not academic baseline, not official measurement, and not used in default AI interpretation unless explicitly selected.

Experimental AI is separated from ATB Measurement. AI can only interpret completed ATB Measurement outputs.

AI must not invent missing measurements, hide camera limitations, treat proxy values as lab measurements, replace coach judgment, produce definitive athlete claims from limited video, or use research-lane outputs unless explicitly enabled.

AI can summarize measured ATB outputs, explain caveats, generate possible coaching language, identify review questions, and compare its interpretation with coach notes.

The long-term objective is for SnatchPhysics to become a recursive learning center — each reviewed movement becomes part of a growing evidence base.

Each lift can preserve raw video metadata, ATB measurement snapshot, validity state, measurement confidence, proxy and N/A metric flags, coach review, experimental AI interpretation, dual review summary, comparison records, reference overlays, tuning profile, export package, and engine and prompt versions.

The system should learn not by pretending AI knows everything, but by accumulating structured review evidence over time.

The remediation sprint added early foundations for a dataset bank: analysis_versions, sessions, session_lifts, and movement_templates.

The first movement template is the snatch. Future templates may include sprint start, sprint acceleration, baseball swing, softball swing, pitching, throwing, jumping, track and field throws, and rehabilitation movement screens.

The dataset bank must preserve measurement context and validity state so future models do not learn from mislabeled or low-quality data.

SnatchPhysics may be useful as an academic teaching tool because it can demonstrate why camera setup matters, why 2D projection is limited, how pose tracking works, why measurement confidence matters, why proxy metrics must be labeled, how phase segmentation can be reviewed, how coach judgment differs from AI interpretation, how validity gates protect scientific integrity, and how a dataset can grow from repeated review.

The teaching value depends on honesty. The system should teach uncertainty, not hide it.

• a validated lab instrument
• a force-plate replacement
• a 3D motion-capture replacement
• an injury prediction system
• a definitive coaching authority
• a fully validated biomechanics system
• a system that can judge an athlete independent of video quality

• a desktop video review prototype
• a movement measurement transparency system
• a scientific observation preservation system with Capture Metadata™
• a teaching and evaluation environment
• a coach-review workflow
• an experimental AI interpretation environment
• a recursive learning foundation
• a dataset-bank prototype

• 2D projected angles only
• wrist/grip proxy for bar path
• single-camera geometry
• measurement confidence is a pose/measurement-confidence proxy
• velocity and acceleration are currently timing-derived proxies
• COM is unavailable until full weighted segment-model COM is implemented
• Power and Force are not computed
• YOLOv8 Pose is research-only
• AI interpretation is experimental
• formal validation against lab-grade systems has not been completed

The portal includes an Engineering Ledger documenting audit findings, score changes, scientific remediation, known limitations, affected files, export changes, and future actions. The first entry documents why the portal was scored 53/100, what was fixed, and why the post-sprint estimate improved to approximately 82/100 pending re-audit. Entry #010 documents the introduction of Capture Metadata™ and the shift from storing uploaded videos to preserving complete scientific observations. The ledger exists to show scientific discipline, not marketing.

Measure honestly. Interpret carefully. Learn recursively.

The current version is not presented as a validated laboratory instrument. It is presented as a serious academic teaching and review prototype designed to make movement video more measurable, reviewable, explainable, and learnable over time.

Final statement: the system measures movement signals, the coach interprets them, AI provides optional experimental interpretation, and the learning center preserves the evidence.

A video records a performance. Capture Metadata records the conditions under which that performance became scientific evidence.