Measured product lead custody receipt.
github.com/Zer0pa/ZPE-MocapRESEARCH-READY
LIVE
RECALL OPEN
Motion Clips Searchable By Shape
A searchable motion archive — find any movement, not just any file · ZPE-Mocap · PyPI zpe-mocap 0.1.1 · github.com/Zer0pa/ZPE-Mocap
A motion-capture archive stores everything and finds almost nothing. A choreographer, animator, or biomechanist looking for a specific gesture starts at the filename and ends up scrubbing.
ZPE-Mocap changes what the archive can answer. It fingerprints BVH skeletons into a motion index: give it a movement, get back the clips that match in 0.826 ms, with the archive itself 18.77× smaller. Playback reconstruction and semantic naming are not in scope here.
01 · THE GAPSTORED, NOT SEARCHABLE
A motion archive captures everything and finds nothing — every search starts at the filename.
02 · MARKETSADJACENT CONTEXT
Motion capture sits inside these markets; none of them can yet search the archive beneath the file.
03 · VALUE
BVHINDEX
Every BVH archive that cannot yet be searched by the movement inside it.
04 · INSIGHT
Motion capture stores the moment. ZPE-Mocap retrieves the movement.
05.1 · CURRENT TECHA LIBRARY WITH NO INDEX
A BVH archive is a library with no index. Files are named by shoot, take, or date. Finding a specific gesture means scrubbing clips manually or trusting sparse metadata. At scale, movements are effectively lost.
05.2 · OUR TECHGIVE IT A GESTURE
ZPE-Mocap fingerprints BVH skeletal trajectories into a compact motion index. Same-source queries return candidates at p50 0.826 ms with Recall@10 0.583 over 24 held-out windows. The index itself compresses raw BVH float32 by 18.77× on the 10-clip CMU mean — the archive becomes searchable and lighter at once.
05.3 · BENCHMARKSBOUNDED CMU EVIDENCE
Compression18.77×
Query p500.826ms
Recall@100.58324 windows
Checks7/710-clip CMU
Scope: 10 CMU clips, 24 held-out windows. Playback and semantic naming not claimed.
06 · MEASUREMENTFIXTURE-BOUNDED METRICS
Every metric is bounded to its fixture window, no broader claim.
06.1 · COMPARATIVE PERFORMANCE · CMU BVH FIXTURE
Evidence: 2026-04-24 retrieval bundle · 10-clip CMU fixture · 24 held-out windows · BVH float32 baseline · Recall@5 0.417 · Recall@1 0.125 · p99 1.191 ms · Playback not claimed.
07 · KEY METRICSBOUNDED CMU EVIDENCE
07.1 · MEAN COMPRESSION
18.77×
vs raw BVH float32 · 10-clip CMU mean
07.2 · RECALL @ 10
0.583
same-source held-out · 24-window set
07.3 · QUERY p50
0.826ms
same-source retrieval · p99 1.191 ms
07.4 · REPO CHECKS
7 / 7
README verification · fixture / search
07.5 · PLAYBACK CLAIM
none
not playback-grade · not the design target
08 · RETRIEVAL SCOPEWHAT DETERMINISTIC MEANS HERE
Committed fixtures, bounded retrieval, no playback claim.
08.1 · WHAT THE EVIDENCE ANCHORS
The word deterministic is narrow here. The public evidence anchors byte-stable canonical payloads, a stable suffix-index retrieval path, and a fixed 10-clip CMU fixture manifest. Public three-platform parity is not yet anchored.
Retrieval evidence is same-source held-out-window search: Recall@10 = 0.583, p50 0.826 ms. That is shape-fingerprint matching, not semantic labeling. Playback fidelity sits outside the design target and outside the claim.
08.2 · HONEST BLOCKER
Honest Blocker ·
No playback-grade reconstruction. No semantic action retrieval. No broad motion platform. Recall@1 sits at 0.125. The CMU compression scale is 10 clips. Retrieval evidence is 24 held-out windows. The public PyPI release zpe-mocap 0.1.1 is stale pending the 0.1.2 cut.
09
ONE MOTION FINGERPRINT opens five paths.
09.1 · THE AMBITION
Motion Capture Memory means a BVH archive you can search by what the body did, not by when the file was saved. Once a skeletal movement is a compact searchable fingerprint instead of a raw stream, retrieval replaces recollection as how studios, labs, and robotics teams operate their motion archives.
09.2 · WHAT WORKS NOW
Working today: same-source fingerprint search at p50 0.826 ms, 18.77× CMU compression, Recall@10 0.583.
09.3 · WHAT'S STILL OPEN
Still open: semantic retrieval, playback reconstruction, broader corpora, recall lift, PyPI 0.1.2 release.
09.4 · ARCHIVES · NEAR-TERM (12–24 MO)
Mocap archives become searchable libraries
An animation supervisor looking for a specific limp, recoil, or hand gesture types a reference clip instead of scrolling through filenames. The decades of capture sitting on studio drives stop being write-only storage and start answering questions.
09.5 · STORAGE · NEAR-TERM (12–24 MO)
Studios stop throwing away takes
When a session shrinks to roughly five percent of its raw size and stays queryable, a games or VFX studio can keep every alternate take rather than picking three to archive. The “we deleted it” conversation with directors goes away.
09.6 · TRAINING DATA · MID-TERM (24–48 MO)
Robotics training sets get curated
A humanoid-robotics team preparing imitation-learning data can deduplicate demonstrations at the movement level instead of by file hash. Near-identical takes get collapsed, rare gestures get up-weighted, and policy training starts from a balanced motion library rather than a filename pile.
09.7 · BIOMECHANICS · MID-TERM (24–48 MO)
Sports labs query by movement pattern
A sports-biomechanics analyst comparing a pitcher's delivery across two seasons stops watching tape and starts running queries: every jump with this hip-knee profile, every gait phase with this stride asymmetry. Longitudinal motion research becomes possible against a full-session archive.
09.8 · INDUSTRY STANDARD · PARADIGM (48 MO+)
Movement gets a shared vocabulary
Animation studios, biomechanics labs, robotics teams, and XR engineers cite the same gesture across capture rigs and file formats. A movement becomes something that can be referenced, compared, and reused across organizations — a shared language for what bodies do, not just what cameras recorded.