ZPE-Neuro

What This Is

Extracellular recording codec. Bounded spike-event extraction for electrophysiology signals, independent of other portfolio lanes. Install from PyPI: pip install zpe-neuro

The strongest CI-anchored result to date: deterministic encode-decode round-trip on DANDI 000034 with a 401x event ratio, 78.44 µV RMSE, and Gate C + Gate D both PASS on blind-clone replay from the current origin/main snapshot. No comparison baseline exists for this lane; the numbers stand on their own terms.

This front door promotes only claims backed by a tracked proof artifact and exercised in CI. Treat CURRENT_AUTHORITY_PACKET.md as the April 24 routing layer; the full proof archive goes deeper.

Codec Mechanics

Field	Value
Architecture	SPIKE_STREAM
Encoding	NEURO_DELTA_V1
Mechanics Asset	.github/assets/readme/lane-mechanics/NEURO.gif

Key Metrics

Metric	Value	Baseline
DANDI 000034 event ratio (window-scoped)	401x
DANDI 000034 RMSE	78.44 µV
Gate D embedded latency (p99)	850 ns
Determinism (identical-hash runs)	5 / 5 seeds

Source: April 24 blind-clone replay from origin/main. All four rows are CI-anchored or artifact-anchored per the proof packets in proofs/selected_artifacts/2026-04-24_zpe_neuro_blind_clone_replay/. Window-scoped metrics (6000-sample, 8-channel window at 30 kHz). Embedded latency uses a hardware-proxy cycle model at 80 MHz ARM-class clock; not a measured on-silicon result.

Repo Identity

Field	Value
Identifier	ZPE-Neuro
Repository	https://github.com/Zer0pa/ZPE-Neuro
Section	encoding
Visibility	PUBLIC
Architecture	SPIKE_STREAM
Encoding	NEURO_DELTA_V1
Commit SHA	7d30d52d704e
License	SAL-7.0
Authority Source	proofs/manifests/CURRENT_AUTHORITY_PACKET.md

Readiness

Field	Value
Verdict	PARTIAL
Checks	6/6
Anchors	6 display anchors
Commit	7d30d52d704e
Authority	proofs/manifests/CURRENT_AUTHORITY_PACKET.md

Honest Blocker

Blind-clone authority replay remains INC; no claim of lossless signal reconstruction; no claim beyond the extracellular lane.

What We Prove

• DANDI 000034 remains the positive public anchor with 41 events, 401.04x event ratio, and 78.44 µV RMSE. Artifact: public_corpus_eval_dandi_000034_mouse412804_ecephys.json. CI: tests/test_dandi_offline.py.
• Blind-clone replay from current repo truth closed with Gate C and Gate D both PASS. Artifacts: verification_summary.md, gate_c_summary.json, gate_d_summary.json. CI: Verify Package Surface / proof-import-smoke, gate-slice, tests/test_roundtrip.py, tests/test_wave1_determinism.py.
• Breadth adjudication records IBL as the counted second extracellular target and does not count the Tier 1 DANDI anchor as breadth closure. Artifacts: public_corpus_summary.json, public_corpus_ibl_waveform_eval.json. CI: tests/test_breadth_adjudication.py.
• DANDI 000003 was executed as the first next-family DANDI breadth probe and recorded FAIL. Artifacts: public_corpus_eval_dandi_000003_yutamouse20_ecephys.json, dandi000003_decision.md. CI: tests/test_breadth_adjudication.py.
• AJILE12 remains explicitly out of family and is excluded from counted breadth. Artifacts: ajile12_family_boundary_decision.md, public_corpus_summary.json. CI: tests/test_breadth_adjudication.py.

What We Don't Claim

• No claim of lossless signal reconstruction.
• No claim that the window-scoped event-encoding ratios are whole-recording compression results.
• No claim that DANDI 000003 closed new breadth.
• No claim of commercialization-safe closure or tagged public release.
• No claim beyond the bounded extracellular lane.

Verification Status

Code	Check	Verdict
T_01	tests/test_dandi_offline.py::test_fixture_reproduces_benchmark_metrics — DANDI 000034 event ratio and RMSE match proof artifact	PASS
T_02	tests/test_roundtrip.py — Encode-decode NWB roundtrip SHA256 bit-consistency	PASS
T_03	tests/test_wave1_determinism.py — 5/5 seed determinism across blind-clone replay	PASS
T_04	tests/test_breadth_adjudication.py — IBL counted breadth logic; DANDI 000003 FAIL recorded	PASS
T_05	CI gate-slice — Gate D embedded latency p99 < 900 ns; drift cliff at 20 µm	PASS
T_06	CI Verify Package Surface / proof-import-smoke — Package importable; proof artifacts present	PASS

Proof Anchors

Path	State
proofs/manifests/CURRENT_AUTHORITY_PACKET.md	VERIFIED
proofs/artifacts/dandi000034_benchmark/benchmark_summary.json	VERIFIED
proofs/selected_artifacts/2026-04-24_zpe_neuro_blind_clone_replay/verification_summary.md	VERIFIED
proofs/selected_artifacts/2026-04-24_zpe_neuro_blind_clone_replay/gate_c_summary.json	VERIFIED
proofs/selected_artifacts/2026-04-24_zpe_neuro_blind_clone_replay/gate_d_summary.json	VERIFIED
proofs/selected_artifacts/2026-04-24_zpe_neuro_blind_clone_replay/determinism_replay_results.json	VERIFIED

Repo Shape

Field	Value
Proof Anchors	6 display anchors
Modality Lanes	1
Architecture	SPIKE_STREAM
Encoding	NEURO_DELTA_V1
Verification	6/6 checks
Authority Source	proofs/manifests/CURRENT_AUTHORITY_PACKET.md

• src/zpe_neuro/: installable extractor package.
• tests/: repo-local verification slice.
• tools/: gate runners and operator scripts.
• proofs/: current authority packet plus April 24 replay and breadth packets.
• docs/: architecture, legal boundaries, release status, and dataset-scope notes.

Extended Metrics

Detailed metric rows retained from the previous ## Key Metrics section. The public product page uses the four-row summary above.

DANDI 000034 Tier-1 Authority Anchor

Metric	Value	Proof artifact	CI test
Event ratio (window-scoped)	401x	benchmark_summary.json	tests/test_dandi_offline.py::test_fixture_reproduces_benchmark_metrics
RMSE	78.44 µV	benchmark_summary.json	tests/test_dandi_offline.py::test_fixture_reproduces_benchmark_metrics
Encode latency (mean / max)	0.089 ms / 0.208 ms	benchmark_summary.json	artifact only — no pytest bound asserted
Decode latency (mean / max)	0.474 ms / 0.686 ms	benchmark_summary.json	artifact only — no pytest bound asserted

These are window-scoped metrics (6000-sample, 8-channel window at 30 kHz). They are not whole-recording compression results.

IBL Second-Target (Tier-2 Breadth, Counted PASS)

Metric	Value	Proof artifact	CI test
Event ratio (window-scoped)	224x	public_corpus_ibl_waveform_eval.json	artifact only — tests/test_breadth_adjudication.py tests logic, not this metric value
RMSE	38.16 µV	public_corpus_ibl_waveform_eval.json	artifact only — tests/test_breadth_adjudication.py tests logic, not this metric value

Gate D: Embedded Latency and Drift Resilience

Metric	Value	Proof artifact	CI test
Modeled latency (mean / p99)	612.5 ns / 850 ns	neuro_embedded_latency.json	CI gate-slice
Latency threshold	< 900 ns	neuro_embedded_latency.json	CI gate-slice
Drift accuracy at 0–15 µm	100%	neuro_drift_resilience.json	CI gate-slice
Drift cliff	at 20 µm	neuro_drift_resilience.json	CI gate-slice

Embedded latency uses a hardware-proxy cycle model at 80 MHz ARM-class clock plus Python reference timing. It is not a measured on-silicon result.

Determinism

Metric	Value	Proof artifact	CI test
Identical-hash runs	5 / 5 seeds	determinism_replay_results.json	tests/test_wave1_determinism.py, tests/test_roundtrip.py
NWB roundtrip SHA256	bit-consistent	neuro_nwb_roundtrip.json	tests/test_roundtrip.py

Competitive Benchmarks

Two non-commensurable metrics, reported in two separate sections so they cannot be conflated. Reproduce with python scripts/comp_benchmark/run_neuro_comparison.py. Full numbers in proofs/artifacts/comp_benchmarks/neuro_codec_comparison.json.

Lossless raw-channel comparison

Operating on the raw int16 samples in tests/fixtures/dandi_000034_mouse412804_ecephys_scan_6000x8.npz (8 channels x 6000 samples = 96000 bytes), with no information loss.

Codec	CR on int16 fixture	Notes
gzip (level 6)	2.20x	stdlib gzip.compress
lz4 (frame, default)	1.42x	lz4.frame.compress
zstd (level 3)	2.17x	zstandard.ZstdCompressor(level=3)

ZPE-Neuro is not in this table. It is not a lossless general-purpose compressor and is not commensurable with these baselines.

ZPE event-extraction ratio (lossy by design)

ZPE-Neuro's 401x ratio is a LOSSY event-extraction operation: it drops non-event samples and retains spike events (41 events kept from 48000 input samples; 768000 raw bits -> 1915 encoded bits on the same window). Window-scoped fidelity at this operating point is RMSE 78.44 µV with roundtrip_exact=False and roundtrip_fidelity=0.0792 per benchmark_summary.json. This is not comparable to the lossless CRs above; reading "401x vs gzip 2.2x" as "ZPE is ~180x better than gzip" is a category error - they are different operations on different outputs. For lossless raw-channel storage, gzip/lz4/zstd remain the appropriate baselines.

Quick Start

git clone https://github.com/Zer0pa/ZPE-Neuro.git
cd ZPE-Neuro
python3.11 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e '.[dev]'
python -m pytest tests

For the bounded gate slice:

python -m pip install -e '.[gate,proof]'
python tools/run_gate_c.py --artifact-root artifacts/manual_gate_c --seed 20260220
python tools/run_gate_d.py --artifact-root artifacts/manual_gate_d --replay-seeds 20260220,20260221,20260222,20260223,20260224

Upcoming Workstreams

This section captures the active lane priorities — what the next agent or contributor picks up, and what investors should expect. Cadence is continuous, not milestoned.

• Second passing DANDI corpus — Active Engineering. Pure data-ingestion work; closes the breadth gap (currently 1 passing corpus) and unlocks the commercialization-safe release gate.
• DANDI 000003 FAIL diagnosis — Research-Deferred — Investigation Underway. Determine whether the failure is a corpus property (announce out-of-scope) or a primitive gap (close it).