zpe-neuro 0.1.0

Private staged repo for the ZPE Neuro Wave-1 package and curated proof surface.

ZPE-Neuro

---

What This Is

Deterministic spike-event encoding for extracellular neural recordings. Anchored on DANDI 000034 with IBL second-target PASS. Compression ratios are window-dependent (see Key Metrics).

ZPE-Neuro targets neurotech research-infrastructure teams and academic neuroscience platforms where non-deterministic encoding pipelines make spike sorting, population decoding, and cross-session comparison irreproducible. The codec is a lossy spike-event extractor — it preserves detected spike events with deterministic replay, not full-signal content. This repo carries the Wave-1 neural signal package, curated proof corpus, and release-surface documentation. Scoped strictly to extracellular data — not broad neural generality.

AJILE12 out-of-family handling is explicitly documented rather than silently excluded. Unresolved blind-clone and commercialization gates remain open.

Field	Value
Architecture	SPIKE_STREAM
Encoding	NEURO_DELTA_V1

Key Metrics

Metric	Value	Scope / Baseline
DANDI_CR	401.04×	0.2 s peak-density window, 8 ch, rank 1 of 9 candidates
IBL_CR	224.30×	0.2 s peak-density window, 8 ch, rank 1 coarse search
RMSE (DANDI)	78.44 µV	Full-window reconstruction error vs. original signal
RMSE (IBL)	38.16 µV	Full-window reconstruction error vs. original signal
DETERMINISM	2/2	Gate C + Gate D synthetic replay

Reading note: Compression ratios are event-encoding ratios measured on the highest-event-density 0.2 s window selected from each recording. They reflect spike-event compactness, not full-signal lossless compression. Whole-recording ratios would be substantially different. RMSE is root-mean-square reconstruction error in microvolts against the original signal.

Source: public_corpus_eval_dandi_000034_mouse412804_ecephys.json | public_corpus_ibl_waveform_eval.json

What We Prove

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

• Deterministic spike-event extraction and encoding on DANDI 000034 extracellular data
• IBL second-target PASS under bounded refinement conditions (0.2 s window scope)
• NWB container round-trip: original samples survive HDF5 write/read with bit-consistent SHA-256 (this tests container fidelity, not codec reconstruction)
• DANDI compatibility: the codec output can be packaged in standard NWB containers
• Both datasets are real public archives with auditable lineage
• AJILE12 out-of-family handling explicitly documented

What We Don't Claim

• Lossless reconstruction -- this is a lossy spike-event extractor. Signal-level RMSE (78 uV on DANDI, 38 uV on IBL) reflects intentional trade-off for spike-event preservation, not full-signal fidelity
• Full-signal compression parity -- the 401x headline applies to a 0.2 s peak-density window (rank 1 of 9 candidates by event count); whole-recording ratios are substantially lower and have not been published
• Bit-identical signal reconstruction -- the NWB round-trip test verifies that original samples survive HDF5 container serialization, not that the codec reconstructs the original signal
• No claim of blind-clone verification
• No claim of commercialization-safe closure
• No claim of tagged release
• No claim beyond extracellular lane

Commercial Readiness

Field	Value
Verdict	PRIVATE_STAGED
Commit SHA	da657d0e12a2
Signal fidelity	NOT_CLAIMED -- codec is a lossy spike extractor; full-signal reconstruction fidelity is not a design goal
Compression scope	WINDOW_ONLY -- headline ratios measured on 0.2 s peak-density windows, not full recordings
Source	proofs/manifests/CURRENT_AUTHORITY_PACKET.md

Evaluators: Available for qualified evaluation. pip install -e . in a clean venv. Contact [e-mail].

Authority Detail

Field	Current truth	Authority
Authority snapshot date	2026-03-21	proofs/manifests/CURRENT_AUTHORITY_PACKET.md
Repository / acquisition surface	Canonical GitHub repo: https://github.com/Zer0pa/ZPE-Neuro for authorized readers. Clone surface for authorized readers: git clone https://github.com/Zer0pa/ZPE-Neuro.git.	AUDITOR_PLAYBOOK.md, docs/ARCHITECTURE.md
Repo posture	PRIVATE_STAGED. Package, install, docs, and proof surfaces are aligned for the current repo state, but this is not a public release and not a clean-clone-closed authority packet.	RELEASING.md, PUBLIC_AUDIT_LIMITS.md
Top unresolved gate	A fresh clean-clone replay of the authority packet remains the top unresolved acceptance gate.	RELEASING.md
Gate status	OPEN for blind-clone and public-release gates. PASS for the current clean packaged baseline and the tracked release-alignment gate slice.	proofs/selected_artifacts/2026-03-21_zpe_neuro_release_alignment/
Current lane / scope lock	Lane 1 is a narrow extracellular recording lane. AJILE12 is explicitly OUT_OF_FAMILY for the current codec; broader human or intracranial coverage is not claimed here.	proofs/selected_artifacts/2026-03-21_zpe_neuro_ibl_refinement/public_corpus_summary.json
Primary positive public anchor	DANDI 000034 remains the strongest positive public waveform anchor in the current repo surface.	proofs/selected_artifacts/2026-03-21_zpe_neuro_ibl_refinement/public_corpus_summary.json
Counted breadth verdict	PASS in the current bounded local evidence packet after the March 21 IBL refinement. This does not upgrade blind-clone status or broader release claims.	proofs/selected_artifacts/2026-03-21_zpe_neuro_ibl_refinement/public_corpus_summary.json
Family-boundary decision	OUT_OF_FAMILY for AJILE12 in the current lane.	proofs/selected_artifacts/2026-03-21_zpe_neuro_ibl_refinement/public_corpus_summary.json
Packaged clean baseline	PASS for the packaged ., .[gate], .[public], and .[proof] surfaces that are actually declared for clean install/import. IBL chunked-waveform tooling and Allen parity remain operator-only.	proofs/selected_artifacts/2026-03-21_zpe_neuro_release_alignment/verification_summary.md, pyproject.toml, docs/ARCHITECTURE.md
Blind-clone verification status	OPEN	PUBLIC_AUDIT_LIMITS.md, RELEASING.md
Release status	NO_PUBLIC_RELEASE. The current repo surface is documented and internally coherent, but no public-release verdict is claimed.	RELEASING.md
Commercialization status	OPEN. Allen parity and commercialization closure remain unresolved.	docs/LEGAL_BOUNDARIES.md, RELEASING.md
Primary authority artifacts	proofs/manifests/CURRENT_AUTHORITY_PACKET.md, proofs/selected_artifacts/2026-03-21_zpe_neuro_release_alignment/, proofs/selected_artifacts/2026-03-21_zpe_neuro_ibl_refinement/	proofs/README.md
Audit route	Start with the short replay path, then read the limits note before widening any claim.	AUDITOR_PLAYBOOK.md, PUBLIC_AUDIT_LIMITS.md

Tests and Verification

Code	Check	Verdict
V_01	DANDI 000034 public anchor	PASS
V_02	IBL second-target bounded refinement	PASS
V_03	AJILE12 family boundary memo	PASS
V_04	Breadth adjudication summary	PASS
V_05	Release-alignment Gate C summary	PASS
V_06	Release-alignment Gate D summary	PASS
V_07	Blind-clone authority replay	INC

Runtime Proof And Package Truth

Surface	Current state	Why it matters
Core package build/import	PASS	The repo now ships a truthful Python package surface rather than relying on undeclared runtime assumptions.
Repo-local tests	PASS on the current shipped unit slice	The staged code surface has a small but real regression check that ships with the repo.
Synthetic gate baseline	PASS for sequential Gate C and Gate D replay in the current clean packaged gate stack	This is the strongest current shipped replay baseline in the repo.
Public replay import surface	PASS for the declared .[proof] import stack	This keeps the docs honest about what the packaged replay stack actually installs cleanly.
IBL and Allen operator paths	OPERATOR_ONLY	Those paths currently require manual dependency or toolchain work around ONE-api, ibl-neuropixel, llvmlite or numba, or allensdk conflicts and are intentionally not shipped as clean extras.
Release automation	PASS for static verification coverage	The repo now has a verification workflow that checks package, build, and install truth without implying a live publish pipeline or automated publish step.

Proof Anchors

Path	State
proofs/manifests/CURRENT_AUTHORITY_PACKET.md	VERIFIED
proofs/selected_artifacts/2026-03-21_zpe_neuro_release_alignment/verification_summary.md	VERIFIED
proofs/selected_artifacts/2026-03-21_zpe_neuro_release_alignment/gate_c_summary.json	VERIFIED
proofs/selected_artifacts/2026-03-21_zpe_neuro_release_alignment/gate_d_summary.json	VERIFIED
proofs/selected_artifacts/2026-03-21_zpe_neuro_ibl_refinement/public_corpus_eval_dandi_000034_mouse412804_ecephys.json	VERIFIED
proofs/selected_artifacts/2026-03-21_zpe_neuro_ibl_refinement/public_corpus_ibl_waveform_eval.json	VERIFIED
proofs/selected_artifacts/2026-03-21_zpe_neuro_ibl_refinement/ajile12_family_boundary_decision.md	VERIFIED

Repo Shape

Field	Value
Proof Anchors	7
Modality Lanes	1
Authority Source	proofs/manifests/CURRENT_AUTHORITY_PACKET.md

Proof Corpus

Packet	Class	How to read it
proofs/manifests/CURRENT_AUTHORITY_PACKET.md	CURRENT	The routing manifest for what is current, what is historical, and which packet owns each claim layer.
proofs/selected_artifacts/2026-03-21_zpe_neuro_release_alignment/	CURRENT	Tracked summaries for the March 21 release-alignment technical pass.
proofs/selected_artifacts/2026-03-21_zpe_neuro_ibl_refinement/	CURRENT	Current bounded local extracellular breadth packet, including the counted IBL second-target PASS.
CHANGELOG.md and runbooks/README.md	SUPPORTING	Chronology, receipts, and operational history. Use them to understand how the current repo state was reached, not as current proof authority.

Open Risks

Surface	Class	Current truth
Blind-clone authority pack	OPEN	The repo does not yet prove a fresh blind-clone authority replay.
Public release	OPEN	No tagged or public release readiness is claimed from this repo state.
IBL / Allen operator paths	OPERATOR_ONLY	These remain outside the clean packaged release surface because the upstream dependency chains are not currently truthful for a clean install.
Commercialization closure	OPEN	Allen parity and commercialization-safe closure remain unresolved.
Broader neural scope	PARKED_BY_SCOPE	Broader human or intracranial coverage is outside the current lane.
Historical path residue	KNOWN_RESIDUE	Some tracked runtime artifacts still contain machine-absolute paths inside captured traces. They are evidence lineage, not current filesystem instructions.
Pickle files in proof tree	KNOWN_RESIDUE	Eight .pkl files (pca_model.pkl, provenance.pkl) ship in the proof artifact tree under proofs/selected_artifacts/. These are SpikeInterface sorter outputs retained as evidence lineage. Loading untrusted pickle files carries arbitrary code execution risk. Future proof surfaces should migrate to safer serialization (JSON, safetensors).

Quick Start

Acquire from the private GitHub repo if you have authorized access, then verify the shipped package surface from a checkout.

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 .
python -c "import zpe_neuro"

Repo-local test slice:

python -m pip install -e '.[dev]'
python -m pytest tests

Repo-local synthetic gate slice:

python -m pip install -e '.[gate,dev]'
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

Repo-local clean public replay import surface:

python -m pip install -e '.[proof]'

The tools/ runners are repo-local scripts, not installed console entry points. LICENSE is the legal source of truth. Read docs/LEGAL_BOUNDARIES.md before turning any repo result into a wider legal or commercial claim.

Ecosystem

Surface	Link	Role
Zer0pa portfolio namespace	github.com/Zer0pa	Parent portfolio namespace for the workstream
ZPE-IMC reference repo	github.com/Zer0pa/ZPE-IMC	Shared docs and README reference line for structure and quality, not for inherited proof claims
Proof surface registry	proofs/README.md	Canonical proof surface entry points for this repo
Reproducing guide	REPRODUCING.md	Offline verify, public DANDI download, and benchmark replay commands for this repo
Open dataset surfaces	docs/OPEN_DATASET_SURFACES.md	Verified next-step public breadth targets without widening the current extracellular lane claim
KiloSort4 operator note	docs/KILOSORT4.md	Benchmark-only comparator guidance and current operator install posture

Observability: Comet dashboard (public)

Who This Is For

Ideal first buyer	Neurotech research-infrastructure team or academic neuroscience platform needing deterministic spike-event encoding with proof lineage
Pain	Neural recording pipelines lack reproducibility guarantees -- re-analyses produce different spike-sorting outputs across environments, eroding audit trust
Deployment	Python package (pip install -e .), private staged. Not currently distributed as a public package
Family position	Validates ZPE encoding generality to neural signal domains. Staged/validation tier alongside Mocap, Prosody, and Bio

Where To Go

Need	Start here
Shortest honest audit path	AUDITOR_PLAYBOOK.md
Current authority routing	proofs/manifests/CURRENT_AUTHORITY_PACKET.md
Architecture and package boundaries	docs/ARCHITECTURE.md
Limits and caveats	PUBLIC_AUDIT_LIMITS.md, docs/LEGAL_BOUNDARIES.md
Release gate	RELEASING.md
Support and contact routing	docs/SUPPORT.md
Canonical doc registry	docs/README.md