Deterministic hand transport for XR two-hand streams: 23.90x compression vs raw (2.82x past Ultraleap's 8.47x), 6.63x smaller frames, and 9.5x better fidelity than the Ultraleap VectorHand proxy — on the ContactPose aggregate lane.

The codec package is real. The ContactPose benchmark lane is real. Public release is BLOCKED: the modern comparator gate is 0/5 and runtime closure for Unity/Meta remains external. Transport numbers are same-machine local proxies, not vendor runtimes.

ZPE-XR

What This Is

ZPE-XR is a deterministic transport codec for two-hand joint streams in the Zer0pa 17-lane codec portfolio. It targets XR platform teams that care about packet size, transport determinism, and replay behavior more than a generic compression headline.

Transport behavior on the ContactPose aggregate lane (five sequences): 23.90x compression vs raw, 0.057 ms mean encode+decode latency, 0.479 mm mean position error, and 6.63x smaller frames than the Ultraleap VectorHand local proxy. The package is real and published on PyPI. Public release is BLOCKED: the modern comparator gate is 0/5 (float16+zlib wins on fidelity), Photon semantics remain narrower, and Unity/Meta runtime closure is external.

Key Metrics

Competitive Benchmarks

The benchmark story is mixed on purpose. ZPE-XR carries a strong transport surface on the ContactPose lane — 6.63x smaller frames and 9.5x better fidelity than the Ultraleap VectorHand proxy — but the closest modern local baseline (float16+zlib) still wins on fidelity (0.277 mm vs 0.479 mm). That is the honest reason the release posture stays PRIVATE_ONLY. Release posture per phase5 decision: PRIVATE_ONLY — public package shipped on PyPI v0.3.1; vendor-runtime closure (Unity/Meta) is external; comparator gate is 0/5 vs float16+zlib. All comparator rows are same-machine local proxies, not vendor runtimes.

Key ratios vs closest open-transport comparator (Ultraleap, same-machine proxy):

• Bytes: ZPE 6.63x smaller on ContactPose real data (25.9 vs 172.0 bytes/frame)
• Latency: ZPE 6.4x lower on ContactPose (0.024 ms vs 0.154 ms mean)
• Fidelity: ZPE 9.5x better MPJPE (0.479 mm vs 4.554 mm)
• Packet-loss resilience: ZPE 9.5x lower pose error at 10% loss (0.399% vs 3.80% 5-seq mean)

Note on Photon: ZPE is smaller than the Photon articulation proxy on ContactPose real-data bytes (25.9 vs 38.0 bytes/frame), but the Photon row meters only the 19-byte-per-hand finger stream and does not include shared hand-root pose — semantics are narrower than ZPE's full two-hand position stream.

What We Prove

• The zpe-xr package and repo install surfaces are real.
• The current ContactPose rerun proves 23.90x compression vs raw, 0.057 ms mean encode+decode latency, and 0.479 mm mean position error on the selected five-sequence lane.
• Byte-identical replay is part of the carried transport surface.
• The cold-start audit and release-decision packet are present and live in the repo.
• Same-machine proxy lanes exist for Ultraleap VectorHand and Photon Fusion XR Hands, with Photon still narrower than the frozen full-position stream.
• The repo can tell the truth about strong transport behavior without pretending comparator superiority or runtime closure.

What We Don't Claim

• No public release readiness.
• No Unity or Meta runtime closure.
• No Photon displacement claim.
• No exact PRD-corpus closure claim.
• No broad hand-tracking superiority claim.

Commercial Readiness

Tests and Verification

Proof Anchors

Repo Shape

Quick Start

Install from PyPI:

pip install zpe-xr

Verify from source:

git clone https://github.com/Zer0pa/ZPE-XR.git zpe-xr
cd zpe-xr
python -m venv .venv
source .venv/bin/activate
python -m pip install "./code[dev]"
python ./executable/verify.py
python -m pytest ./code/tests -q

Read docs/ARCHITECTURE.md first, then docs/LEGAL_BOUNDARIES.md, then the Phase 5 and Phase 6 proof anchors above. LICENSE is the legal source of truth; the repo uses SAL v7.0.