# Technique and prototype review

Built 16 October 2026. These are authored demonstrations on the existing Signal mannequin,
with the same skeletal motion used by the interactive scene and guide. This is a review
prototype, not a professionally certified technique library.

The portable mesh omits covered torso/hip skin to prevent clothing overlap artifacts.
The avatar casts shadows onto its surroundings; smooth lighting on the avatar keeps
its grips and joint positions readable without harsh self-shadow patterns.

## Bench press

The existing bench rig is reused. Feet, hips, upper back and head remain supported.
The grip width is solved for nearly vertical forearms at the bottom; the bar follows a
shallow path from lower chest to above the shoulders. The closed grip includes articulated
fingers and a wrapped thumb. A separate setup scene places the supported bar over the face
and demonstrates shoulder-girdle retraction/depression before lifting.

Measured: 67.2 cm between grip centres; 39.55 cm vertical bar travel; bottom elbow-to-torso
angle 50.3°; elbow flexion 4.3–123°. The hands have 18° wrist extension in the existing rig;
this is a modest extension, not a perfectly straight wrist. Grip-centre error stays under
0.002 mm. The setup clip illustrates positioning; it does not teach a complete unrack/rerack.

References: [NASM bench press](https://www.nasm.org/resource-center/exercise-library/barbell-bench-press),
[NSCA coaching foundations](https://www.nsca.com/education/courses/foundations-of-coaching-lifts/foundation-of-coaching-lifts--bench-press-quiz/).

## Flat dumbbell chest press

This is a chest press on a flat bench, not an overhead shoulder press. Body support is
reused from the bench rig. Dumbbells lower together with elbows angled from the trunk,
then move upward and slightly inward without colliding. Wrist orientation is neutral.
Handle centres are derived from the same hand transforms that pose the fingers.

Revision 03 corrects the shallow bottom position against real demonstration footage.
The previous handles stopped about 12 cm above this avatar's central chest. They now
reach the evaluated chest surface at 0.67059 m, with elbows about 15.44 cm below the
shoulder joints. Handle travel is 41.64 cm; the top also changed to match the visible
soft elbow bend. The fitted elbow flexion range is 23.66–131.80°.

The reference shows slight medial forearm inclination and less convergence at the top.
The fitted forearm inclination is 2° at the top, 15° around mid-height and 10° at the
bottom, with a 55° upper-arm direction from the trunk. These are fitted avatar parameters,
not measured human 3D joint angles or a prescription for every person. Hand orientation
follows the forearm, retaining neutral wrists and closed grips throughout the movement.
The measured top/bottom handle separation ratio is about 0.81 in the comparison view,
closer to the reference's approximately 0.88 than the previous 0.56, but not identical.
Getting heavy dumbbells into/out of position is not animated.

References: [ACE dumbbell chest press](https://www.acefitness.org/resources/everyone/exercise-library/19/chest-press/)
and the [U.S. Marine Corps demonstration](https://www.dvidshub.net/video/636920/dumbbell-bench-press).

### Actual video tracking and fitting

RTMPose-m ran through OpenCV on the two cropped video views. An initial comparison
processed 12 timestamps in each view at three rotations (72 inferences). On the six
manually reviewed phase/view combinations, the selected shoulder/elbow/wrist chain
had mean image errors of 2.51 px in the front view and 2.73 px in the oblique view;
16 of 18 points were within the reviewer's stated uncertainty. This is a small fitting
sample, not held-out validation. The tighter crop also changed compared with the earlier
MediaPipe experiment, so the result is not a general model benchmark.

A further 37 source frames (74 inferences) supplied the near-wrist height progression
for one real repetition. Reviewed endpoints, monotonic smoothing and cubic Hermite
interpolation provide the curve actually used by the animation. Its phases are slowed
and given explicit holds in a four-second instructional loop. Reviewed bottom/middle/top
poses constrain the spatial path; fixed bone lengths, supported body contacts and
explicit finger/handle constraints remain part of the rig.

The cameras are uncalibrated, the two selected wrists are different anatomical sides,
and image height is not a metric 3D trajectory. Shoulder depth, forearm rotation and
fingers cannot be recovered reliably from this footage. The result is a video-guided,
manually reviewed fit to the existing avatar, not automatic 3D motion capture. The
comparison and selected tracking overlays are available on the reference-check page;
full raw coordinates and model weights remain local QA materials.

The earlier statement that the 8–98° profile was cross-checked against the reference
was too strong: it verified wrist consistency but failed the visible depth comparison.
That profile is preserved locally as before evidence and is replaced in this revision.

## Leg press

A 45° sled moves with the feet; the mannequin's pelvis and torso stay supported. The
rails and carriage share the same travel line. Hands hold fixed side handles, feet stay
fully on the platform, and knees track in the same plane as the feet. The guide explains
that depth must stop before the pelvis rolls off the pad.

Measured knee flexion: 8–90°. Sled travel: 23.42 cm. The sampled minimum sole-to-platform
gap is 0.25 mm throughout the rep. This machine is schematic: safety releases and load
selection are explained in the narration but are not animated machine controls.

Reference: [NASM leg press](https://www.nasm.org/resource-center/exercise-library/leg-press).

## Standing dumbbell biceps curls

Both arms curl together with palms forward/up, upper arms held slightly forward and still,
neutral wrists, and a quiet torso. The dumbbells clear the clothing at the bottom. This is the fully supinated variation; it does not add a mid-repetition
forearm twist. The 4-second loop includes controlled lowering and a short bottom pause.

Measured elbow flexion: 7.84–114.13°. Grip-centre error remains below 0.001 mm.

References: [Mayo Clinic dumbbell curl](https://www.mayoclinic.org/healthy-lifestyle/fitness/multimedia/biceps-curl/vid-20084675),
[ACE curl mechanics](https://www.acefitness.org/resources/everyone/exercise-library/70/bicep-curl/).

## What was checked

- Per-frame joint lengths and grip centres in the authored geometry.
- Revised dumbbell press: every exported frame checked for hand/forearm alignment,
  the fitted forearm inclination, loop continuity and stable support joints. Evaluated head, back,
  glutes and shoe contacts remain stable. Soft-tissue surfaces enter the bench pad
  by roughly 3–12 mm; this represents foam contact, not a floating torso.
- Revised dumbbell press: 97-pose surface-intersection checks found no dumbbell-head
  intersections with the body, clothing, bench, or opposite dumbbell.
- Exported joint positions against Blender source at start, bottom/midpoint and loop end.
  Maximum position discrepancy across the four exports: below 0.001 mm.
- Identical start/end bone matrices for all four interactive loops.
- Evaluated shoe geometry: 0.25–1.25 mm floor clearance; 0.25 mm platform clearance.
- Curl dumbbell-head surfaces versus body and shorts at 17 poses: no intersections.
- Close-up visual checks of each grip and leg-press feet, plus overview and side views.
- Browser orbit, scrub, half speed, camera presets, and guide/3D switching.
- Phone (390×844) and tablet (820×1180) layout emulation.
- HyperFrames runtime, layout and text-contrast validation for all four guides.

These measurements test consistency and contacts. They are not an exhaustive mesh collision
audit or an independent biomechanics assessment. A qualified trainer's review and physical
iPhone/iPad testing remain necessary before release. Narration is generated with the specified
ElevenLabs v3 voice, aligned from provider timestamps, and supplied with optional captions;
audio listening review remains a human review item.
The shoulder-blade setup shot remains subtle and partly concealed by the clothing and
bench; its explanation currently relies mainly on the narration.

Evidence: `qa/kinematic-summary.json`, `qa/source-geometry.json`, `qa/browser-checks.json`,
`qa/guide-checks.json`, `qa/curl-clearance.json`, the PNG review frames, and the rendered-video verification report.