FILE 00 / WORKING SYSTEM CONCEPT

Stimulate the nerve at the right point in the stride.

Norysr is a pre-prototype research programme for pediatric cerebral-palsy foot drop. Its first engineering question is whether one lower-leg motion sensor and a small causal model can detect toe-off and heel strike reliably enough for a clinician-gated stimulation device to act on, and abstain when they cannot. There is no finished device, no gait dataset and no device performance result.

FIG. 01 / PROPOSED CONTROL LOOP
6-AXIS IMUTA / GM sEMGEDGE CLASSIFIERINT8 / 1D-CNNFES DRIVERDEFAULT OFFLIMITS TBDLATE STANCEPRE-SWINGMID-SWINGCOMPUTE BUDGET ≤23 ms
Proposed architecture, not measured prototype performance. This unmeasured sensor-to-command budget is not a stimulation window.
PROJECT STATE

Specification v0.1 and software scaffold

Log-only package, 281 synthetic tests. No hardware, no clinical use.

INITIAL SCOPE

Pediatric CP

Unilateral spastic CP with foot drop, ages 6 to 17, provisional.

EVIDENCE STATUS

Audit in progress

Paper withdrawn; partial phase-1 audit recorded.

NORYSR PARTICIPANTS

0

No original participant data. The side study used public data.

01 / CONTROL LOOP

Five operations. One timing decision to prove.

The v0.1 contract is narrow on purpose. It does not diagnose cerebral palsy, prescribe stimulation or replace a clinician. It defines a causal signal-to-event path whose timing, abstention and failure behavior can be measured before any output exists.

01

Read the shank

One calibrated 6-axis inertial sensor on the affected distal shank at 100 Hz. Surface EMG is deferred to later evidence-gated ablations and is not a v0.1 input.

02

Validate every frame

Sequence, timestamp, calibration identity and range checks run before any transform. A gap, reset or stale frame clears the 32-frame window; nothing catches up.

03

Estimate gait events

A 1,603-parameter causal temporal convolutional network reports the probability of toe-off within 80 ms, heel strike within 80 ms and stance now. A transparent rule-based detector runs beside it as the required comparator.

04

Supervise deterministically

An explicit state machine owns every inhibit, timeout and request. Output stays disabled or simulated through the shadow stages; the model never chooses a stimulation setting.

05

Seal the run

Immutable run bundles record code, configuration, split and model identities, and append-only logs keep every decision and fault. The control loop never depends on the cloud.

02 / EVIDENCE BOUNDARY

The evidence map is being rebuilt from primary sources.

Release review found citation identifiers that did not match the works they were said to identify. The earlier paper and its quantitative summaries stay withdrawn. A partial audit has since tagged the high-impact claims, found one conflated citation, and moved every literature value out of the requirements.

Identity
Match every DOI, PMID and preprint identifier to the correct title, authors and year.
Claim
Read the source and confirm that it reports the attributed population, method and result.
Boundary
Keep published findings separate from unmeasured Norysr requirements and from the measured side study.
A polished review is not evidence until its sources survive verification.

03 / MEASURED SIDE STUDY

One result has been measured. It is not a device result.

EEG-Clear, a 360,081-parameter network that cleans one channel of scalp EEG at a time, was trained on public healthy-adult motor-movement recordings and tested on five people it never saw. It beat the standard band-pass filter for every one of them, passed clean signal through almost untouched, and exposed a real limitation: waveform accuracy alone did not preserve the movement rhythms until the training objective was told to.

+14.4 dB

Held-out SNR gain, filter +7.1 dB

5 of 5

Held-out people improved, 5.0 to 7.8 dB over the filter

0.70

Band-power fidelity after stage 2, from 0.57; untouched 0.83

04 / OPEN WORK

The next useful proof is a gait model that beats a rule on held-out children.

The software that would train and judge it exists and is tested. The data that would make the result meaningful do not. Nothing participant-facing follows until they do and until qualified reviewers say so.

A

Finish the evidence audit

Verify primary identifiers, remove every claim that cannot be traced to a source, and keep literature separate from requirements.

B

Obtain data that can support a claim

License or collect a participant-grouped gait dataset with pediatric CP, a distal-shank sensor and an independent gait-event reference. The workspace has none.

C

Train Model 1 against the baseline

Only on data at the required tier, with participant-level splits, frozen thresholds and predeclared paired gates. Retain the rule method if the model shows no justified advantage.

D

Obtain qualified review

Clinical, ethics, electrical-safety and regulatory review precede any hardware output or participant-facing plan.