What can on-device swing analysis measure?

From one clear side-view swing, Rundown Swing Lab can estimate where visible body landmarks appear over time, infer approximate swing phases, and grade only the checkpoints that clear its visibility, confidence, phase, and runtime gates. It can also attempt an image-based bat path, but the path appears only when the current tracker verifies both ends of the bat across enough frames.

It cannot measure ground force, pressure, weight transfer, joint torque, muscle activity, true three-dimensional depth, true bat speed, exit velocity, or injury risk. Those require sensors, calibrated cameras, force plates, ball tracking, medical expertise, or validation that one phone clip does not provide.

The clearest way to use it is as a coach’s second look at a video, not as a laboratory verdict.

Capability and limitation table

Question	What the current Swing Lab can do	What it cannot honestly claim
Can it find the hitter?	The current MediaPipe reader attempts to locate 33 body landmarks on each analyzed frame and carries a visibility value for each landmark.	A landmark is a model estimate, not a marker attached to an anatomical joint center. Occlusion, blur, framing, lighting, orientation, and fast motion can move or hide the estimate.
Can it find swing phases?	It infers stance, load, stride, launch, contact, extension, and finish from body-motion proxies.	It does not directly observe force transfer or guarantee the true instant of bat-ball contact. The body reader’s “contact” remains a wrist-speed proxy; a separately verified bat read supports only its own bat-path claim.
Can it check body positions?	It attempts 11 named checkpoints: Stance Width, Load Tempo, Stride Reach, Hip-Shoulder Separation, Back-Elbow Slot, Hips Fire First, Back-Heel Release, Quiet Head, Front-Leg Block, Posture Hold, and Finish Balance.	It may not grade every checkpoint. A label or band is not a clinical measurement, a universal youth standard, or proof that one mechanic caused an outcome.
Can it compare movement through the swing?	It can estimate screen-space displacement, body-normalized ratios, selected joint or segment angles, and relative timing between inferred phases.	It cannot recover metric-accurate depth or a full three-dimensional reconstruction from the recorded scene. Some “3D” inputs are model-inferred coordinates from one 2D view, not multi-camera motion capture.
Can it call hips before shoulders?	It calls the relative hip-versus-shoulder order only when its image-plane read and model-derived world-coordinate read agree and the frame spacing is sufficient.	It does not report a full laboratory kinematic sequence or validated absolute pelvis, torso, arm, hand, and bat angular velocities.
Can it show a bat path?	It draws the automatic tracer only when image evidence passes the motion, endpoint, sweep, continuity, handedness, and renderer gates.	It cannot promise a tracer on an ordinary clip. Current measured reach is 2 of 10 original iPhone clips. A refusal draws no substitute hand-path line and no guessed bat position.
Can it report bat speed or attack angle?	The shipping report currently does not show a bat-speed number, mph, exit velocity, or attack-angle number.	One uncalibrated side-view clip has no reliable metric scale or complete depth path for true bat speed. The current tracer reach also does not support marketing an automatic bat metric.
Can it give coaching feedback?	When enough checkpoints are readable, it can show On Track, Work This, Couldn’t Call It Cleanly, or Couldn’t Read states and select a bounded first focus from supported evidence.	It cannot prove that a drill will improve performance, diagnose the hitter, replace the coach, or rank an athlete for selection or recruiting.
Can it compare sessions?	It can keep a short, device-local history of the date, sport, age band, readable/on-track tally, and first suggested focus.	It does not create a cloud athlete profile, sync a video library, or establish that two differently filmed clips are measurement-equivalent.

What the 11 checkpoints actually are

The checkpoint names are coaching shorthand. Underneath them, the current build reads these narrower signals:

Checkpoint	Current camera-derived signal	Important boundary
Stance Width	Ankle span divided by shoulder span at the inferred stance frame	An internal coaching range, not a published universal stance-width standard
Load Tempo	Inferred load-to-plant time relative to inferred plant-to-contact time	Phase timing is frame-rate limited, and a missing true plant can withhold downstream claims
Stride Reach	Front-ankle horizontal travel from load to plant, normalized to estimated body height	It is travel in the image, not landed foot-to-foot stride distance or force production
Hip-Shoulder Separation	Peak separation from model-derived hip and shoulder coordinates, with a 2D cross-check	A camera estimate, not calibrated torso-pelvis motion capture
Back-Elbow Slot	Rear-elbow angle and limited elbow-versus-wrist position signals	Occlusion and camera direction make front-versus-back calls fragile; unsupported faults are withheld
Hips Fire First	Relative peak order of hip and shoulder turning, cross-checked two ways	No absolute lab-grade angular velocity or full body-to-bat sequence
Back-Heel Release	Model-derived rear-foot pitch change from stance toward contact	The foot points are noisy and require a higher visibility floor
Quiet Head	Nose displacement from launch to contact, normalized by an estimated head width	Image displacement only; it does not measure vision, balance systems, or cause of movement
Front-Leg Block	Lead-knee angle and change between an inferred plant and contact	Without a detected plant, the read cannot support a force or bracing claim
Posture Hold	Change in the image-plane trunk angle from stance toward contact	A 2D posture estimate, not spinal loading or a medical assessment
Finish Balance	Mid-hip offset from the ankle base plus a short stillness check	A pose proxy, not center of mass, center of pressure, or ground-reaction force

For ages 10-13, the current interface emphasizes sequence and stability rather than professional degree targets. Ages 14-18 can receive tighter bands only where the build has a supported rule. The page itself warns that most published ranges are baseball-derived and that fastpitch bands remain more forgiving until more sport-specific evidence is loaded.

Why this is not a biomechanics lab

Google’s Pose Landmarker returns normalized image coordinates and model-estimated world coordinates for 33 body landmarks. Its own model card says the depth value is inferred by fitting synthetic GHUM data to 2D point projections, is not metric, and is out of scope for applications that require metric-accurate depth. The model is intended for single-person pose and fitness uses, and its quality can degrade with motion, noise, lighting, scale, orientation, and occlusion.

That is useful input for a bounded coaching read. It is not the same instrument as a biomechanics lab.

Published baseball and fastpitch biomechanics studies use equipment that one phone view does not have. A foundational baseball study used three-dimensional kinematic and kinetic data. A collegiate fastpitch study used eight motion-capture cameras at 200 Hz, reflective markers on the bat, and a three-dimensional computer model to calculate bat trajectories, forces, torques, work, and power. Other batting studies add force plates to measure ground-reaction forces.

The distinction is simple:

• Kinematics describe motion, such as position, angle, and velocity.
• Kinetics describe forces and moments that cause or accompany motion.
• One side-view video can support selected motion estimates when the image is clear.
• It does not directly measure forces, pressure, torque, muscle activity, or energy transfer.

Single-camera research also has to be validated for the exact model, task, movement speed, view, and target metric. Accuracy reported for squats, gait, or jumps cannot be transferred to a youth swing or to Rundown without a swing-specific ground-truth study.

What Swing Lab refuses

Refusal is part of the result. The current build uses several separate gates:

Condition	What the current product does	What it does not do
A landmark or checkpoint is hidden, missing, implausible, engine-limited, or low confidence	Leaves that checkpoint ungraded or says it could not call it cleanly	Does not convert low confidence into On Track or Work This
Fewer than five checkpoints are readable	Shows a reshoot result without an overall tally, fault, or saved history entry	Does not grade from a thin subset
Fewer than 60% of analyzed frames have usable pose coverage	Refuses the mechanics read and explains the coverage problem	Does not stretch or fill the missing majority into a confident report
The inferred phase order or timing is physically implausible	Refuses the mechanics grade and identifies whether the cause is the footage or the reader	Does not blame the hitter for a broken phase solution
The browser tab is hidden or the runtime stalls enough to invalidate sampling	Discards the run and asks for a clean retry with the tab visible	Does not call the clip poor when the runtime was the problem
The browser uses the current unstable backup pose reader	Shows the clip and phases but no checkpoint stamps, fault, overall tally, or history write	Does not reuse a grade that changed across repeat runs
The body read passes but the bat tracker fails	Keeps supported body checkpoints, states that the bat path is unavailable, and draws no bat ink	Does not draw a wrist path and call it a bat path
The body phase read fails but an independent bat read passes	May show only the verified bat-path snapshot while body mechanics remain ungraded	Does not let a bat read backfill unsupported body mechanics

This is stricter than “show a confidence score.” A refused value never becomes a quiet input to a headline or drill.

The current automatic-tracer reliability result

Rundown tested the shipping tracer against ten original founder-recorded iPhone clips, IMG_4452.MOV through IMG_4461.MOV. Each was an original 4K/120 fps file, not a compressed derivative.

Result	Count	What rendered
Accepted automatic bat paths	2 of 10	Both accepted models rendered positive bat-path segments
Refused automatic bat paths	8 of 10	All eight rendered zero bat segments
Overall automatic reach	20%	Below the internal 70% launch gate

The current classical image tracker often has only a one-to-three-pixel bat to work with through blur and motion. A higher-resolution region-of-interest experiment made the result worse, reaching 0 of 10, so it was removed from shipping source. The confidence floors were not loosened to make more clips pass.

This result proves two narrow things: the automatic tracer is not ready to be called reliable, and all eight observed refusals failed closed. It does not establish a 20% success rate for every phone, hitter, environment, or clip. Ten clips from one founder corpus are too small and too concentrated for that claim.

A separate structural review run on August 17 passed 28 of 29 checks but failed the no-hole core-sampling assertion for the planner's 240 fps, eight-second case. That unresolved failure does not change the measured 120 fps corpus counts above, but it blocks any broader claim that the sampling plan preserves every launch-to-extension frame pattern. Until that gate is green, 240 fps continuity is unverified.

What stays on the device?

Data	Current handling	Boundary
Selected swing video	Read from the local file inside the browser. Rundown does not upload, receive, store, or view it.	The browser keeps the selected file available in the open tab for a one-tap retry. Closing the tab removes that session access.
Decoded frames, pose landmarks, and bat evidence	Processed in browser memory and not sent to Rundown, Supabase, or analytics.	They are working data for the current analysis, not a cloud record.
Results history	A maximum of 20 short summaries can be stored in that browser’s local storage.	The saved object contains time, sport, age band, first focus, and readable/on-track tallies. It does not contain video, raw landmarks, file names, or raw checkpoint values. Clearing browser data removes it.
Reader software and fonts	The page downloads its code, pose reader, model files, and fonts. The browser can cache them for later use.	The third-party delivery services receive ordinary request information such as IP address and browser type. No video, landmark, or result data is included in those requests.
Account and analytics	Swing Lab requires no player account and the current Swing Lab path has no product analytics or Supabase result write.	This statement applies to the current Swing Lab analysis path, not every other Rundown account or roster surface.

“On device” describes where the video analysis runs. It does not mean the page makes zero network requests. The software and model must reach the browser before the local analysis can run.

What should a coach use it for?

Use a supported read to start a conversation:

• Compare the video with the checkpoint the product could actually read.
• Pick one supported focus instead of treating every band as a diagnosis.
• Refilm when framing or coverage was the limitation.
• Keep the same side view when comparing sessions, while remembering that a history tally is not a controlled study.
• Ask the athlete what they felt and use the coach’s live observation as context.

Do not use it to clear an injured player, prescribe rehabilitation, claim a bat-speed gain, evaluate college potential, or overrule a coach or clinician who has better evidence.

Quick answers

Is on-device swing analysis accurate?

It can be useful for the specific landmarks and checkpoints that pass the current gates. Rundown has not completed a ground-truth validation study that supports a general accuracy percentage for swing mechanics. The automatic bat tracer currently reaches only 2 of 10 clips in its small original-iPhone corpus.

Does Swing Lab measure a true 3D swing?

No. It uses one side-view video plus model-derived pose coordinates. Those coordinates can support selected estimates, but they are not calibrated multi-camera motion capture and do not provide metric-accurate scene depth.

Does it measure weight transfer?

Not directly. Swing Lab can estimate visible body movement, but it cannot measure how much force or pressure each foot applies. It does not have force or pressure sensor input.

Does it measure bat speed or exit velocity?

No. The current shipping report does not show bat-speed mph or exit velocity. It also does not track the batted ball.

What happens when it cannot read the bat?

The bat path is withheld, the report says it is unavailable, and no bat ink is drawn. Supported body checkpoints can still appear because the pose read and bat read are separate claims.

Is the video uploaded?

No. The current Swing Lab reads the selected file in the browser. Rundown does not receive the video, landmarks, or results. A short result summary can remain in local browser storage until that browser data is cleared.

See what this early build can read

Try Swing Lab in Rundown’s early-access preview. Use one side-view swing and expect the product to leave unsupported checkpoints or the bat path ungraded.

---