Measured product lead custody receipt.
github.com/Zer0pa/ZPE-NeuroRESEARCH-READY
LIVE
BREADTH OPEN
Event Horizons of Neural Spikes
Extracellular spike-event codec · ZPE-Neuro · PyPI zpe-neuro v0.1.1 · github.com/Zer0pa/ZPE-Neuro
A neuron fires when the voltage crosses a threshold — not before, not after. At that exact moment, a spike exists. Extracellular recordings catch millions of those crossings and return them as gigabytes of raw voltage.
ZPE-Neuro works at the threshold: it extracts spike events on DANDI 000034 at 481× event-ratio compression and 74.44 µV residual, then replays them bit-identical across five seeds. The window is bounded. The crossing is precise. Latency is a modeled cycle figure, not measured on silicon. Breadth beyond 000034 stays open.
01 · THE GAPBOUNDED EXTRACTION
Brain recordings come back as raw voltage. the spike events inside them have no standard archive.
02 · MARKETSADJACENT FORECASTS
Electrophysiology and neurotech are buying the recorders. The spike events those recorders produce still have no shared archive format. ZPE-Neuro is tested on DANDI 000034 only.
03 · VALUE
$33.6B
The 2031 electrophysiology market; a spike-event codec is the unpriced layer beneath it.
04 · INSIGHT
A spike is not the whole recording. it is the event the archive can keep.
05.1 · CURRENT TECHRAW VOLTAGE FIRST
Extracellular pipelines usually keep voltage first and derive spike events later. That preserves the full trace, but it leaves every threshold crossing buried inside large files, local preprocessing choices, and dataset-specific scripts.
05.2 · OUR TECHEVENTS, HELD
ZPE-Neuro extracts threshold crossings into a deterministic event stream. On the declared DANDI 000034 window, it reports 481× event-ratio compression, 74.44 µV residual, and 5/5 bit-identical replay across random seeds. Breadth beyond that window remains open.
05.3 · BENCHMARKSDANDI 000034 WINDOW
Event ratio481×
Spike RMSE74.44µV
Seeds5/5exact
Releasev0.1.1stale
Scope: public extracellular spike-event extraction on DANDI 000034. Latency is modeled, not silicon-measured.
06 · MEASUREMENTDECLARED WINDOW
Every claim stays inside DANDI 000034. breadth beyond the window is still open.
06.1 · COMPARATIVE PERFORMANCE · SPIKE EVENT EXTRACTION
Measured on the declared DANDI 000034 window. DANDI 000003 failed; IBL remains unclosed. Check D latency is a modeled proxy, not an on-silicon claim.
07 · KEY METRICSDANDI 000034 · EVENT EXTRACTION
07.1 · EVENT RATIO
481×
Window-scoped · DANDI 000034 only
07.2 · SPIKE RMSE
74.44µV
Voltage residual after replay · DANDI 000034 window
07.3 · SEEDS
5 / 5
Bit-identical event replay · across five random seeds
07.4 · CHECKS
C & D
Residual and timing checks pass · latency modeled, not silicon
07.5 · RELEASE
v0.1.1
PyPI STALE · PR #52 draft
08 · DETERMINISM5-SEED BIT-EXACT
Five random seeds, five bit-identical spike-event streams.
08.1 · WHAT THE CHECKS MEASURESCOPED CLAIM
Deterministic means one narrow thing here. On DANDI 000034's declared window, the spike-event extractor emits bit-identical event streams across five random seeds — 5/5 SEEDS measures that. Extracted events are the threshold crossings; the voltage between them is gone. Latency is a modeled cycle figure at 80 MHz ARM-class clock, not measured on silicon. Check D modeled mean 612.5 ns, p99 850 ns, against a 900 ns proxy threshold. No on-silicon determinism claim follows.
08.2 · THE FIDELITY GAP
Honest Blocker ·
DANDI breadth MISS: 1 of 2 counted targets; DANDI 000003 ran and failed, IBL remains unclosed. Latency is hardware-proxy modeled, not silicon. PR #52 is draft, not on main. PyPI v0.1.1 is stale. README still carries a private staged badge. No clinical or diagnostic claim is made.
09
ONE SPIKE EVENT opens five paths.
09.1 · THIS REPO'S AMBITION
The aim is not a general neural decoder. It is one thing well: a threshold crossing that survives encode, compress, and replay unchanged. Applied across datasets, that discipline turns extracellular archives from raw voltage dumps into objects two labs can point at and mean the same event.
09.2 · WHAT WORKS NOWEXTERNAL
Working today: 481× event extraction with 74.44 µV residual and bit-identical replay on DANDI 000034.
09.3 · WHAT'S STILL OPENEXTERNAL
Still open: DANDI 000003 failed, IBL breadth unclosed; silicon latency and release state remain open.
09.4 · ARCHIVES · NEAR-TERM (12–24 MO)
Labs keep full sessions, not samples
A lab that compresses a recording 481 times can keep entire sessions on the same disks that used to hold representative slices. The conversation about what to discard from a neural archive ends.
09.5 · PROBES · NEAR-TERM (12–24 MO)
Spike packets fit a probe’s timing budget
A modeled 850 ns p99 encode places this packet inside the time budget of a real electrode interface. A probe-firmware architect can plan around it now and validate on silicon later, rather than wait for both at once.
09.6 · CALIBRATION · MID-TERM (24–48 MO)
Pass and fail across datasets calibrate the field
DANDI 000034 passes. DANDI 000003 failed. IBL is next. Each result, kept and named, tells a neuroscience research group exactly which kinds of extracellular recordings this codec covers and which it does not — before they commit budget.
09.7 · CROSS-DEVICE · MID-TERM (24–48 MO)
One spike event reads the same everywhere
Five seeds already produce bit-identical replays on one machine. The next step is the same identity across different runtimes and hardware — the condition for a BCI platform to trust that a spike captured in one rig is the same object when a partner lab opens it.
09.8 · REPLICATION · PARADIGM (48 MO+)
Neural archives acquire chain-of-custody
When a threshold crossing has a deterministic identity, two labs can point at the same spike across time, hardware, and protocol revisions. Replication in extracellular neuroscience starts to mean the same thing twice — not a similar plot, the same event.