zpe-mocap 0.1.0

Zero-label structural motion search and compression. ZPE-MoCap deterministic codec/search package.

ZPE-Mocap

---

What This Is

ZPE-Mocap applies the ZPE deterministic 8-primitive encoding architecture to motion-capture data — compression, search, and retrieval for spatial-temporal skeletal signals.

Real-data results (CMU fixture, 10 BVH clips): 18.77× compression, 32.45 mm MPJPE, 82.51° angle RMSE. Evidence: proofs/artifacts/2026-04-14_cmu_corpus_benchmark/.

Synthetic-corpus results (identity encoding): 85.19× compression, 1.19 mm MPJPE, 1.16e-07° angle RMSE, p@10 = 1.0 search, 26.14 ms p95 latency. Evidence: proofs/artifacts/2026-02-20_zpe_mocap_wave1/. The synthetic corpus is pre-tokenized from the codec's own alphabet — the encoder passes tokens through unchanged, so these numbers represent a theoretical ceiling, not operational performance.

Animation studios, game engines, and motion-data infrastructure teams evaluating deterministic mocap compression should benchmark against the CMU fixture numbers (18.77×, 32.45 mm). The gap to commercial readiness is real-world fidelity improvement and broader corpus validation.

Readiness: staged. Real-data compression demonstrated (18.77×) but fidelity is not production-grade (32.45 mm MPJPE). No Blender runtime pass, no CMU commercialization-safe closure, no clean-clone verification exist.

Not claimed: Production-grade fidelity on real data, real-data parity with synthetic benchmarks, fair competitive benchmarks (ACL comparison is circular), Blender runtime compatibility, clean-clone verification, commercialization-safe closure.

Proof anchor	Location
Wave1 evidence bundle	proofs/artifacts/2026-02-20_zpe_mocap_wave1/
Compression / fidelity / search / latency	wave1 bundle artifacts
Falsification results	wave1 bundle

Part of the Zer0pa family. Platform layer: ZPE-IMC.

---

What This Is

Deterministic mocap compression and retrieval. CMU fixture corpus: 18.77× compression, 32.45 mm MPJPE across 10 real BVH clips. Synthetic corpus (identity encoding): 85.19× compression, 1.19 mm MPJPE. Compress, search, and retrieve skeletal motion data without decompression.

ZPE-Mocap targets animation pipeline teams and mocap-data infrastructure where studios archive terabytes of BVH/FBX data that can only be queried after full decompression. This codec indexes motion during encoding — downstream search never touches the raw stream.

Question	Answer
What is this?	A deterministic mocap compression and retrieval reference stack backed by a synthetic corpus with preserved proof lineage.
What is the current authority state?	Imported 2026-02-20_zpe_mocap_wave1 synthetic-corpus proof bundle; no new run-of-record has been accepted inside this repo boundary.
What is actually proved?	CMU fixture-corpus compression (18.77×, 32.45 mm MPJPE, 82.51° RMSE) and synthetic-corpus metrics (85.19× compression, 1.19 mm MPJPE — identity encoding). The synthetic numbers are a theoretical ceiling due to pre-tokenized data bypassing quantization.
What is not being claimed?	No real-data parity with synthetic benchmarks (identity encoding inflates synthetic numbers). No fair competitive benchmark (ACL comparison is circular — fed synthetic data from ZPE tokens). No CMU-backed commercialization-safe closure, no Blender runtime pass, and no clean-clone verification. The bundle is historical and may retain machine-absolute paths.
Where should an outsider acquire and verify?	Clone https://github.com/Zer0pa/ZPE-Mocap.git, run the quick verify path below, and inspect proofs/artifacts/2026-02-20_zpe_mocap_wave1/ as the authority surface.

Field	Value
Architecture	SKELETON_MANIFOLD
Encoding	JOINT_ANGLE_V2

Key Metrics

CMU fixture corpus (10 real BVH clips):

Metric	Value	Notes
COMPRESSION	18.77×	vs raw BVH float32 (range 15.2×–23.0×)
MPJPE	32.45 mm	mean per-joint position error
ANGLE RMSE	82.51°	mean joint-angle RMSE

Source: results.json, summary.md

Synthetic corpus (pre-tokenized from codec alphabet):

Metric	Value	Notes
COMPRESSION	85.19×	identity encoding — synthetic tokens pass through encoder unchanged
MPJPE	1.19 mm	synthetic corpus only
ANGLE RMSE	1.16e-07°	synthetic corpus only
SEARCH	p@10 1.0	—
LATENCY	26.14 ms	p95 query latency

Source: mocap_compression_benchmark.json, mocap_position_fidelity.json, mocap_search_eval.json, mocap_query_latency.json

Why the gap? The synthetic corpus is pre-tokenized from the codec's own 8-direction alphabet (synthetic.py populates xy_tokens, xz_tokens, magnitudes_mm directly). When these fields are present, the encoder short-circuits quantization (codec.py:189-195) and passes tokens through unchanged — an identity encoding. Real BVH data must be quantized from continuous joint positions, which introduces quantization error. The synthetic numbers represent a theoretical ceiling, not operational fidelity.

Competitive Benchmarks

Circular methodology disclosure: The ACL comparison below was run on the synthetic corpus. The synthetic BVH was generated FROM ZPE tokens (the codec's own 8-direction alphabet), not from independent real BVH data. Both ZPE-Mocap and ACL were fed the same synthetic clips, but the data structurally favours ZPE-Mocap because it was generated from the ZPE token vocabulary. This comparison does not reflect real-world competitive performance. A fair comparison would require both codecs to compress the same independent BVH corpus (e.g., CMU).

Selected clips from the 10-clip direct comparator on the synthetic BVH corpus. The mean row covers all 10 compared clips.

Clip	ZPE-Mocap	ACL	Win ratio
walk_0000	61.0×	17.1×	3.6×
run_0000	60.9×	14.7×	4.1×
jump_0000	52.6×	12.3×	4.3×
fall_recover_0000	77.8×	15.2×	5.1×
Mean	57.0×	19.1×	3.0×

Source: acl_direct_comparator_table.json

Synthetic corpus — general-purpose compressor comparison:

Tool	Synthetic Corpus CR	Notes
ZPE-Mocap	85.19×	Synthetic benchmark (identity encoding)
gzip	69.70×	~22% behind ZPE on same synthetic corpus

ZPE-Mocap exceeds gzip by ~22% on the synthetic corpus. This margin is considerably narrower than the ACL comparison above. Real-world corpus validation (CMU, AMASS) is pending.

CMU fixture corpus (real BVH) — compression only:

Corpus	Mean CR	Range	Source
CMU fixture (10 clips)	18.77×	15.2×–23.0×	2026-04-14 benchmark

No ACL comparison has been run on the CMU fixture corpus.

What We Prove

Auditable guarantees backed by committed proof artifacts. Start at AUDITOR_PLAYBOOK.md.

• Synthetic-corpus compression at 85.19× (identity encoding — tokens pass through unchanged)
• Synthetic joint-angle fidelity at 1.16e-07° RMSE (identity encoding)
• Synthetic positional fidelity at 1.19 mm MPJPE (identity encoding)
• Synthetic search ranking at p@10 = 1.0
• Synthetic query latency at 26.14 ms p95
• CMU fixture-corpus compression at 18.77× mean across 10 real BVH clips
• CMU fixture-corpus positional fidelity at 32.45 mm MPJPE
• CMU fixture-corpus joint-angle fidelity at 82.51° RMSE

What We Don't Claim

• Real-data parity with synthetic benchmarks — the synthetic corpus is pre-tokenized from the codec's own 8-direction alphabet, producing an identity encoding. Real BVH data produces substantially different fidelity (32.45 mm vs 1.19 mm MPJPE; 82.51° vs 1.16e-07° angle RMSE; 18.77× vs 85.19× compression).
• Fair competitive benchmarks — the ACL comparison used synthetic data generated from ZPE tokens, not independent BVH. This is circular methodology.
• Release readiness
• Production motion-pipeline integration
• Blender or Maya plugin support

Commercial Readiness

Field	Value
Verdict	CONDITIONAL — synthetic only
Commit SHA	34d94f1f29b4
Confidence	Synthetic PASS; real-data fidelity FAIL (32.45 mm MPJPE, 82.51° angle RMSE)
Source	proofs/artifacts/2026-02-20_zpe_mocap_wave1/quality_gate_scorecard.json, proofs/artifacts/2026-04-14_cmu_corpus_benchmark/summary.md

Evaluators: Synthetic-corpus wave-1 PASS (identity encoding). CMU fixture corpus shows 18.77× compression but 32.45 mm MPJPE and 82.51° angle RMSE — not production-grade fidelity. Real-corpus validation is the remaining gate. pip install -e ./code to evaluate. Contact [e-mail] for real-corpus evaluation access.

Tests and Verification

Code	Check	Verdict
V_01	Synthetic compression benchmark	PASS
V_02	Joint-angle fidelity	PASS
V_03	Position fidelity	PASS
V_04	Search ranking	PASS
V_05	Query latency	PASS
V_06	Commercialization claim adjudication	CONDITIONAL — synthetic PASS, real-data fidelity gap open
V_07	Integration readiness contract	INC

Proof Anchors

Path	State
proofs/artifacts/2026-02-20_zpe_mocap_wave1/mocap_compression_benchmark.json	VERIFIED
proofs/artifacts/2026-02-20_zpe_mocap_wave1/mocap_joint_fidelity.json	VERIFIED
proofs/artifacts/2026-02-20_zpe_mocap_wave1/mocap_position_fidelity.json	VERIFIED
proofs/artifacts/2026-02-20_zpe_mocap_wave1/mocap_search_eval.json	VERIFIED
proofs/artifacts/2026-02-20_zpe_mocap_wave1/mocap_query_latency.json	VERIFIED
proofs/artifacts/2026-02-20_zpe_mocap_wave1/quality_gate_scorecard.json	VERIFIED
proofs/artifacts/2026-02-20_zpe_mocap_wave1/commercialization_claim_adjudication.json	VERIFIED
proofs/artifacts/2026-02-20_zpe_mocap_wave1/integration_readiness_contract.json	PARTIAL
proofs/artifacts/2026-04-14_cmu_corpus_benchmark/results.json	VERIFIED
proofs/artifacts/2026-04-14_cmu_corpus_benchmark/summary.md	VERIFIED

Repo Shape

Field	Value
Proof Anchors	10
Modality Lanes	1
Authority Source	proofs/artifacts/2026-02-20_zpe_mocap_wave1/quality_gate_scorecard.json

Quick Start

Quick Verify

Use the clone/install path below as repository verification guidance, not packaged public-release guidance.

git clone https://github.com/Zer0pa/ZPE-Mocap.git
cd ZPE-Mocap
python -m venv .venv
source .venv/bin/activate
python -m pip install -e ./code
python -m unittest discover -s code/tests -v
python - <<'PY'
from zpe_mocap.codec import decode_zpmoc, encode_clip
from zpe_mocap.synthetic import generate_clip

clip = generate_clip(
    clip_id="readme_smoke",
    label="walk",
    frames=120,
    fps=60,
    seed=20260220,
    noise_scale=0.0002,
)
enc = encode_clip(clip, seed=20260220)
dec = decode_zpmoc(enc.payload)
print(enc.compression_ratio, dec.clip_id)
PY

Expected outputs:

• python -m unittest discover -s code/tests -v completes locally after the editable install.
• The smoke snippet prints a compression ratio and returns readme_smoke as the decoded clip id.
• Evidence remains anchored in proofs/artifacts/2026-02-20_zpe_mocap_wave1/.

Shortest outsider path:

docs/README.md

docs/ARCHITECTURE.md

AUDITOR_PLAYBOOK.md

License Boundary

• Free tier boundary: annual gross revenue at or below USD 100M under SAL v6.2.
• SPDX tag: LicenseRef-Zer0pa-SAL-6.0.
• Commercial or hosted use above threshold must follow the contact and enforcement terms in LICENSE.

Ecosystem

Workstream	Route	Notes
ZPE-Mocap	github.com/Zer0pa/ZPE-Mocap	This motion-capture compression and retrieval workstream.
ZPE-IMC	github.com/Zer0pa/ZPE-IMC	Portfolio reference repo reused for documentation and structure alignment.
ZPE-XR	github.com/Zer0pa/ZPE-XR	Adjacent spatial-media workstream in the ZPE portfolio.
ZPE-Robotics	github.com/Zer0pa/ZPE-Robotics	Sibling workstream for robotics motion and control surfaces.
ZPE-Bio	github.com/Zer0pa/ZPE-Bio	Another proof-anchored codec workstream in the same portfolio.

Observability: Comet dashboard (public)

Who This Is For

Ideal first buyer	Animation pipeline or mocap-data infrastructure team evaluating deterministic compression with search for motion archives
Pain	Studios archive terabytes of BVH/FBX data that can only be queried after full decompression — storage costs climb, retrieval is slow
Deployment	Python reference implementation (pip install -e ./code). Public repo, not a packaged release
Family position	Proves ZPE encoding applicability to motion-capture and spatial-temporal signal domains. Staged/validation tier alongside Neuro, Prosody, and Bio

Runtime Proof (Wave-1)

The only promoted proof surface is the imported 2026-02-20_zpe_mocap_wave1 synthetic-corpus bundle. No clean-clone verification, Blender runtime pass, or CMU-backed closure is promoted beyond this evidence.

Evidence bundle 2026-02-20_zpe_mocap_wave1 Imported synthetic-corpus proof artifacts retained for lineage and current claims.

Runtime boundary python reference only No Blender runtime verification or clean-clone replay is promoted here.

Historical Authority Surface

Accepted authority bundle 2026-02-20_zpe_mocap_wave1 Imported synthetic-corpus evidence bundle. No later run-of-record is promoted.

Backend truth backend=python Python reference implementation; no compiled runtime authority is claimed here.

Performance authority CMU fixture (real): mean_cr=18.77, mpjpe_mean_mm=32.45, angle_rmse_deg=82.51 Synthetic (identity): mean_cr=85.19, mpjpe_mean_mm=1.19, query_latency_p95_ms=26.14 Synthetic metrics are a theoretical ceiling (identity encoding). CMU fixture metrics are the operational benchmark.

Surface	Locked value	Why it matters
Authority bundle	proofs/artifacts/2026-02-20_zpe_mocap_wave1/	Current proof surface for all promoted metrics.
Corpus type	synthetic	All current claims are synthetic-corpus claims; no CMU-backed closure is promoted.
Compression ratio	zpmoc_mean_cr=85.1893	Synthetic-corpus mean compression ratio from the wave1 benchmark artifact.
Joint-angle fidelity	joint_angle_rmse_deg≈1.16e-07	Synthetic joint-angle RMSE for wave1 fidelity tests.
Position fidelity	mpjpe_mean_mm=1.1901	Synthetic mean per-joint position error from wave1.
Search ranking	p_at_10=1.0	Synthetic search evaluation for the wave1 corpus.
Query latency	query_latency_p95_ms=26.1375	Synthetic query latency p95 from the wave1 benchmark.
ACL comparator	zpmoc_mean_ratio=57.0328, acl_mean_ratio_same_raw_bvh32=19.1487	Circular methodology: ACL comparator captured on synthetic BVH generated from ZPE tokens, not independent real data. This structurally favours ZPE-Mocap and does not reflect real-world competitive performance.
External acquisition surface	https://github.com/Zer0pa/ZPE-Mocap.git	Public clone target for this repo.

Authority Notes

The imported wave1 bundle is the current authority surface; no later run-of-record has been re-accepted inside this repo boundary.

Blender runtime verification remains unpromoted; existing compatibility notes are simulated only.

CMU-backed commercialization-safe closure and clean-clone verification remain gaps and are explicitly not claimed.

Proof Anchor Notes

proofs/artifacts/2026-02-20_zpe_mocap_wave1/mocap_compression_benchmark.json Compression ratio metrics for the synthetic corpus.	proofs/artifacts/2026-02-20_zpe_mocap_wave1/mocap_joint_fidelity.json Joint-angle RMSE evidence for the synthetic corpus.
proofs/artifacts/2026-02-20_zpe_mocap_wave1/mocap_position_fidelity.json MPJPE positional fidelity evidence for the synthetic corpus.	proofs/artifacts/2026-02-20_zpe_mocap_wave1/mocap_search_eval.json Search ranking evidence for the synthetic corpus.
proofs/artifacts/2026-02-20_zpe_mocap_wave1/mocap_query_latency.json Query latency p95 evidence for the synthetic corpus.	proofs/artifacts/2026-02-20_zpe_mocap_wave1/acl_direct_comparator_table.json ACL comparator table for the same raw-BVH32 baseline.
proofs/artifacts/2026-02-20_zpe_mocap_wave1/integration_readiness_contract.json Integration readiness contract captured in the bundle.	proofs/artifacts/2026-02-20_zpe_mocap_wave1/falsification_results.md Falsification results for the synthetic wave.
proofs/artifacts/2026-04-14_cmu_corpus_benchmark/results.json CMU fixture corpus benchmark: 18.77× compression, 32.45 mm MPJPE, 82.51° angle RMSE across 10 real BVH clips.	proofs/artifacts/2026-04-14_cmu_corpus_benchmark/summary.md Human-readable summary of CMU fixture corpus benchmark results and limitations.

Proof rung	Locked value	What it proves now
CMU fixture compression	mean_cr=18.77	Compression ratio on 10 real CMU BVH clips. Operational benchmark.
CMU fixture position fidelity	mpjpe_mean_mm=32.45	Mean per-joint position error on real CMU data. Not production-grade.
CMU fixture angle fidelity	angle_rmse_deg=82.51	Joint-angle RMSE on real CMU data. Not production-grade.
Synthetic compression	zpmoc_mean_cr=85.1893	Compression ratio on the synthetic corpus (identity encoding — theoretical ceiling).
Synthetic joint fidelity	joint_angle_rmse_deg≈1.16e-07	Joint-angle RMSE on the synthetic corpus (identity encoding — theoretical ceiling).
Synthetic position fidelity	mpjpe_mean_mm=1.1901	Mean per-joint position error on the synthetic corpus (identity encoding — theoretical ceiling).
Synthetic search ranking	p_at_10=1.0	Search evaluation at p@10 on the synthetic corpus.
Synthetic query latency	query_latency_p95_ms=26.1375	p95 query latency for the synthetic corpus.

Modality Status Snapshot

ZPE-Mocap is a motion-capture sector. The status below reports only the synthetic-corpus evidence that exists today and marks the missing Blender, CMU, and clean-clone gates.

Surface	Status	Proved now	Boundary and evidence
Compression	AMBER	CMU fixture: 18.77× (real). Synthetic: 85.19× (identity encoding).	Real-data compression is proven but substantially below synthetic headline. Synthetic zpmoc_mean_cr=85.1893 is identity encoding (theoretical ceiling).
Joint-angle fidelity	RED	CMU fixture: 82.51° RMSE (real). Synthetic: 1.16e-07° (identity encoding).	Real-data angle RMSE is not production-grade. Synthetic value is meaningless as an operational metric (identity encoding).
Position fidelity	RED	CMU fixture: 32.45 mm MPJPE (real). Synthetic: 1.19 mm (identity encoding).	Real-data MPJPE is not production-grade. Synthetic value is identity encoding (theoretical ceiling).
Search ranking	GREEN	Synthetic search evaluation in wave1.	p_at_10=1.0 in mocap_search_eval.json.
Query latency	GREEN	Synthetic latency p95 in wave1.	query_latency_p95_ms=26.1375 in mocap_query_latency.json.
Blender runtime	RED	No Blender runtime proof is promoted.	Compatibility notes remain simulated only.
CMU closure	RED	No CMU-backed commercialization-safe closure.	Workspace CMU clone lacks usable corpus files.
Clean-clone verification	RED	No clean-clone verification has been run from this repo boundary.	Evidence remains imported and unrerun in this repo.

Throughput

No throughput benchmark is promoted. Performance summary:

Compression ratio CMU fixture (real): mean_cr=18.77 Synthetic (identity): mean_cr=85.19 CMU fixture is the operational benchmark. Synthetic is identity encoding (theoretical ceiling).

Query latency p95 query_latency_p95_ms=26.1375 Synthetic query latency p95 from wave1.

Measure	Locked value	Meaning
Latency unit	ms (p95)	All latency values are p95 in milliseconds.
Compression ratio (CMU real)	mean_cr=18.77	Mean compression ratio on 10 real CMU BVH clips. Operational benchmark.
Compression ratio (synthetic)	zpmoc_mean_cr=85.1893	Mean compression ratio on the synthetic corpus (identity encoding — theoretical ceiling).
Query latency p95	query_latency_p95_ms=26.1375	Search query latency p95 on the synthetic corpus.

Public ML Workbooks

No public ML workbook is promoted for ZPE-Mocap at this time. All promoted evidence remains in the local wave1 proof bundle under proofs/artifacts/2026-02-20_zpe_mocap_wave1/.

Role	Run name	Workbook
Current promoted public twin	NONE	NOT_PUBLISHED
Historical lineage	2026-02-20_zpe_mocap_wave1	LOCAL_BUNDLE_ONLY

Go Next

If you need to...	Open this
Understand the runtime map and authority classes	docs/ARCHITECTURE.md
Navigate the documentation surface	docs/README.md
Read legal and lane-specific public boundaries	docs/LEGAL_BOUNDARIES.md
Audit historical compatibility and replay boundaries	AUDITOR_PLAYBOOK.md
Read public audit limits and explicit non-claims	PUBLIC_AUDIT_LIMITS.md
Inspect proof artifacts and logs directly	proofs/

Area	Purpose
README.md, CONTRIBUTING.md, SECURITY.md, SUPPORT.md, LICENSE	Root governance and release-facing metadata
code/	Installable package and codec implementation surface
docs/	Architecture, legal boundaries, support, and documentation routing
proofs/	Proof corpus, baselines, and falsification evidence

Open Risks (Non-Blocking)

• Synthetic benchmark circularity: All synthetic metrics (85.19× compression, 1.19 mm MPJPE, 1.16e-07° RMSE) are produced by identity encoding — the synthetic corpus is pre-tokenized from the codec's own alphabet, so the encoder passes tokens through unchanged. These numbers are a theoretical ceiling, not operational fidelity.
• ACL comparison circularity: The ACL direct comparator was run on synthetic BVH generated from ZPE tokens, not independent real data. The "3× win" is structurally inflated.
• Real-data fidelity gap: CMU fixture corpus shows 32.45 mm MPJPE and 82.51° angle RMSE — substantially worse than synthetic numbers and not production-grade.
• Blender runtime proof remains unpromoted; compatibility notes are simulated only.
• CMU-backed commercialization-safe closure is not available in this repo boundary.
• Clean-clone verification has not been executed from this repo.
• Historical artifacts can retain machine-absolute paths from the 2026-02-20 bundle.
• No public ML workbook has been published for this repo; evidence is local to the wave1 bundle.

Contributing, Security, Support

Contribution workflow: CONTRIBUTING.md	Security policy and reporting: SECURITY.md	User support channel guide: docs/SUPPORT.md
Documentation index: docs/README.md	Autonomous agents and AI systems using this repository are subject to Section 6 of the Zer0pa SAL v6.2.

Ecosystem Cross-Links

• ZPE-IMC — reference repo for shared repository structure and documentation alignment.
• code/README.md — installable package surface for the ZPE-Mocap workstream.
• docs/README.md — documentation router for architecture, legal boundaries, and support surfaces.
• proofs/README.md — proof-corpus entrypoint for the evidence carried inside this repo.