Bidirectional peripheral neural interface — research orchestration round 01

CONCEPT STUDY · RESEARCH ONLY · NO HUMAN TESTING · NOT AVAILABLE · NO DEVELOPMENT PROGRAM EXISTS

Historical source verdict: REVISION. Continue the research. Do not select a product. The source said the scientific question was legitimate but the evidence was not sufficient for visualization, architecture selection, biological work, hardware work, or a development program.

Record date 12 Aug 2026Round 01Verdict REVISIONProduct selection None

Preserved source identity

Source PDF: output/pdf/novel-biotech-bpni-research-round-2026-08-12.pdfPrivate repository source

SHA-256: b885bc850ce472a44263976d66e2dad1bf420e38b786457e468966ca30fcac6b

A hash establishes byte identity, not truth or current authority. All nine pages were rendered and inspected; this edition preserves the historical claims, categories, uncertainty, dates, and stated limits without updating them.

Source page 1 of 9 · integration report

Evidence before design

The round recorded seven active lanes reviewed — Science, Product Design, Backend, Business, QA, Legal, and Brand — and 2/10 Pro requests used. Both Pro outputs were retained as advisory; neither was accepted without source reproduction.

Historical operating window
Point	Planned state
Now	Evidence
Q4 2026	Simulation preparation
Q1 2027	Comparator narrowing
Mid-June 2027	Research closeout

Source page 2 of 9 · system

The layered idea is valid — but gain cannot recover lost information

Forward path

Nerve (natural ENG) → contact (mixed voltage) → front end (dynamic range) → digitize (captured samples) → cancel (estimated artifact) → classify (bounded state) → computer (decision log).

Information-loss zone: clipped, saturated, blanked, or never-digitized samples cannot be recovered downstream.

Return path

Computer (abstract request) → policy (allow / inhibit) → electronics (command) → interface (delivery state) → afferent nerve (reported effect) → central nervous system (unknown meaning).

What layers could and could not do in the historical source
What layers can do	What layers cannot do
Improve gain and dynamic range Reduce predictable artifact Represent uncertainty and abstain Separate requested versus delivered state Fail closed when inputs are invalid	Create signal-to-noise ratio that never existed Recover clipped or blanked samples Prove recovered ENG is natural Turn EMG or evoked compound action potential into ENG Establish biological safety

Core open question: can natural ENG remain decision-relevant while the same human peripheral nerve is being stimulated?

Source page 3 of 9 · evidence

What the human evidence actually establishes

Historical evidence matrix
Evidence object	Route / sample	Direct ENG	Nerve return	Chronic	Duplex	Human
Short direct nerve read + stimulate	N=2 / 4–5 weeks	FACT	FACT	N/A	OPEN	FACT
Limited closed-loop task	S3 only / neural + EMG	FACT	FACT	N/A	OPEN	FACT
Long direct neural recording	N=3 / one to 502 days	FACT	N/A	OPEN	N/A	FACT
Implanted muscle sensing + nerve return	iSens / e-OPRA	N/A	FACT	FACT	OPEN	FACT
Same-nerve natural ENG during stimulation	Same human interface	OPEN	OPEN	OPEN	OPEN	OPEN
Biohybrid two-way intermediary	Living tissue + electronics	OPEN	OPEN	OPEN	OPEN	OPEN

0 target demonstrations: none of the seven reproduced Pro sources showed same-interface human natural ENG preserved during stimulation.
6/7 arrays lost recording yield: in one N=3 percutaneous USEA study, six of seven arrays lost functional recording electrodes within two months.
OPEN — chronic direct two-way: negative search is not proof of absence; the source called it a frontier requiring narrower evidence.

Source page 4 of 9 · options

Three architecture families remain research branches — none is selected

Historical option frame
Family	Status	Strongest support	Strongest evidence against
A — Direct electronic	OPEN	Strongest human component evidence	Recording yield, artifact, fibrosis, infection, lead, and revision failures
B — Biohybrid	HYPOTHESIS	Human regenerative peripheral nerve interface (RPNI) EMG to 1,054 days	It is muscle EMG, not direct ENG; no human afferent return path
C — Combined	HYPOTHESIS	Four-week rat induced-pluripotent-stem-cell myocyte interface	No human evidence; coupled drift and attribution multiply

Common platform layer the source called defensible: schemas, provenance, uncertainty, calibration lifecycle, and fail-closed states.

Not defensible as common: biological meaning, nerve map, return code, clinical endpoint, or universal calibration.

Source page 5 of 9 · stress test

Two visual warnings: chronic yield and source-category drift

Chronic recording yield / human USEA

One participant reached 502 days. Six of seven arrays lost functional recording electrodes within two months. The source emphasized that long duration in one channel or participant could not substitute for stable population-level yield.

Biohybrid yield / rat iPSC-myocyte study

Electrode yield fell from 97% to 25% over four weeks; longitudinal awake signal-to-noise ratio N=2. The source described this as a promising preclinical signal route, not evidence of a stable human two-way biological intermediary.

Pro advisory / seven-source reproduction
Count	Category	Limit
2	Reviews	Not primary experiments
2	Concurrent other modality	Human EMG / primate cortical local field potential
3	Animal component studies	Not concurrent natural ENG duplex
0	Target demonstrations	Same human nerve + natural ENG + stimulation

Source page 6 of 9 · program

Research plan through mid-June 2027

Now — source contract. Correct canonical Science records; keep route, N, duration, and conflicts of interest adjacent.
Next — comparator narrowing. Test seven candidate problems against the strongest external or temporary alternative.
After — synthetic parameter pack. Science defines dimensionless assumptions and thresholds; no patient data.
Then — offline simulation. Only after separate implementation approval; no hardware, actuator, network, or biology.
June 2027 — research closeout. Decide whether any branch deserves a separate product-development gate; not automatic.

Always blocked in this phase: implantation / biological protocols / animal or human work / product claims / deployment / outreach / availability.

Source page 7 of 9 · governance

Lane and QA scoreboard

All rows were marked PASS in the historical round
Lane	Record	Source statement
Science	`SCI-BPNI-03C`	Transfer accepted only as bounded Science evidence; research remains open.
Product Design	`PD-BPNI-02R1`	Corrected provenance, G0–G10, and ownership; visualization blocked.
Backend	`BPNI-SIM-01`	Offline plan only; implementation blocked pending Science parameter pack.
Business	`BUS-BPNI-01R1`	Comparator frame retained; papers relabeled routing-only.
Legal / Ethics	`LRE-BPNI-01R1`	Neurodata classification wording narrowed; no classification issued.
Brand	`BRAND-BPNI-01`	Internal factual / aspirational / speculative framework only.
Independent QA	`ROUND 01`	All P2 corrections reproduced; no P0 or P1 remains in this round.

Integrator verdict: REVISION. Round 01 was accepted as a research-control package, not product approval. Release, visualization, architecture selection, simulation implementation, biological work, and development remained unauthorized.

Source page 8 of 9 · next

The next three bounded work packages

Science — comparator baseline review. Inspect candidate sources across seven problem classes; record population, comparator, endpoint, N, duration, limitation, and one kill threshold. No ranking yet.
Science + Backend — synthetic parameter contract. Define dimensionless event, noise, artifact, drift, confidence, and failure assumptions. The simulator must refuse a pass when any threshold lacks provenance.
Integrator + user — research-only indication decision. After packages 01–02, retain one clinical problem for deeper research or keep all branches open. This selects a question, not a product or development program.

Earliest product-development decision: mid-June 2027, only if evidence and governance gates close. The source warned that time does not close a gate, and a simulation, paper, or prototype does not itself create authorization.

Source page 9 of 9 · sources

Decision-bearing records and source limits

Wendelken et al. — PMID 29178940: N=2; 4–5 weeks; S3-only closed-loop task; no simultaneous duplex claim.
George et al. — PMID 33045689: N=3; one participant to 502 days; 6/7 arrays lost functional recording electrodes within two months.
Rochford et al. — PMID 36947608: Four-week rat iPSC-myocyte/electronic forward-only analogue; longitudinal awake SNR N=2; yield 97% to 25%.
Vu et al. — PMID 37023743: Human RPNI N=2; long-lived muscle EMG, not direct ENG.
Earley et al. — DOI 10.3389/fnhum.2022.1030207: Human implanted EMG plus nerve-cuff stimulation; blanked and artifact-subtracted, not natural ENG duplex.
WAND — DOI 10.1038/s41551-018-0323-x: N=1 rhesus macaque cortical local field potential; coincident samples interpolated; not peripheral ENG.
EU ATMP Regulation — Regulation (EC) 1394/2007: Biohybrid classification remained conditional; no Novel Biotech classification had been issued.
FDA AI PCCP — guidance issued 2025-08-18: Bounded change-control guidance; did not authorize unbounded adaptation.

Evidence rule: reviews, search snippets, model summaries, screenshots, and prior PASS labels were routing evidence only. Claims became canonical only after primary-source reproduction and Independent QA.