Fall Barometer — Gait & Biomechanical Signal Pipeline¶

Date: 2026-06-30 · Author: Gene Lang, PharmD — Director of Pharmacy Operations (research + drafting: multi-agent Workflow, adversarially citation-verified) For: Mike (AI/wearable architect) + engineering Grounded in: Fall_Risk/Balance_Meter_Canonical_Model_2026-05-20 · Product_Specs/Balance_Meter_Three_Versions_Spec_2026-06-07 · Product_Specs/Balance_Meter_Sensor_DataFlow_Spec_2026-06-08 · Engineering_Reference/Hardware_Truth_Synthesis_2026-05-04. Citation discipline: every specific claim below is backed by a real, independently-verified source (Apple's own published methodology, peer-reviewed gait-biomechanics literature, and competitor product pages) — 50 citations gathered, 42 passed adversarial verification unchanged, 8 were corrected (misquotes, a unit error, an unsupported bundled detail, or an overstated characterization) before shipping. No citation ships unfixed. [NEEDS VERIFICATION] marks anywhere a fact was wanted but not found in a real source — never filled with an invented number. Status: 🟡 Draft technical grounding — not yet engineering-reviewed or pharmacist-signed-off. The numeric drift thresholds (delta gates, %) remain the trade-secret canonical model's; this doc is about how the underlying gait metrics get computed, not the scoring weights.

1. Purpose¶

No existing Rx360 document specifies how Layer 2 of the Balance Meter actually turns raw sensor samples into the gait metrics it claims to compute. The current spec set names outputs (gait speed, cadence, walking asymmetry, step-timing variability, double-support time, sway, sit-to-stand transition time, turn smoothness) and assigns them a relative signal-weight tier (Mark II = 100%/"unqualified" confidence, Mark I = ~60–70%/"moderate," No-Wearable = 0%), but none of the source documents specify a sampling rate, a windowing scheme, a gait-event or bout-detection method, or a sit-to-stand measurement algorithm. The bout-detection gap in particular is flagged in the current-state grounding as "the single most important unbuilt prototype."

This document exists to close that gap with engineering-ready technical grounding drawn from (a) how Apple computes the closest published equivalents on iPhone/Apple Watch, and (b) the peer-reviewed gait-biomechanics and fall-risk literature. It is a technical grounding reference, not a finalized implementation. Every number below is either a competitor/reference-implementation data point or a published research threshold — none of it is an Rx360-validated number, and per the Sensor/DataFlow spec and Canonical Model, actual delta-gate thresholds remain "mechanism-locked, numbers pending the pilot." Nothing here overrides that pilot-calibration requirement; it narrows the engineering search space ahead of it.

All Mark I/Mark II hardware references below are constrained to what the current-state grounding confirms: Mark I = 6-axis IMU (LSM6DSV320, 320 Hz accel+gyro), GPS/GNSS, BLE, LTE-M, NTN satellite, mic+speaker, capacitive touch, magnetometer (AK09940A) — no barometer, no PPG, no dedicated wear-detection, no display. Mark II adds barometric pressure, PPG, dedicated wear-detection, and the Pill Bottle accessory, carrying the same 320 Hz IMU forward with no stated rate change.

2. Signal-by-signal breakdown¶

Gait speed / speed-delta from personal baseline¶

Sensor(s): Mark I 6-axis IMU (LSM6DSV320, 320 Hz accel+gyro) only. No barometer or PPG is needed or available on Mark I for this metric. Mark II carries the same IMU forward — no confirmed sampling-rate change.
Standard computation method: For wrist-worn daily-living sensing, direct heel-strike event detection (the foot/ankle-sensor approach) is not viable — wrist motion is confounded by non-gait arm movement, so "gait must be inferred indirectly (as 'gait windows' rather than discrete heel-strike events)... modern approaches use deep learning (CNN/U-Net, self-supervised learning) trained on windowed data (10-second non-overlapping windows in ElderNet)." ElderNet standardized its sampling rate to 30 Hz ("we standardized the sampling rate of all datasets to 30 Hz to align with the frequency used in the pre-trained UK Biobank model. This relatively low sampling rate allowed for the efficient use of long-duration recordings"), well below Mark I's native 320 Hz, meaning downsampling is a legitimate option, not a hardware constraint. A separate real-world wrist validation study (Kluge et al.) used the same 100 Hz / ±8g configuration as its co-located INDIP reference system, with accelerometer-only algorithms, and reported sensitivity ranging "between 0.55 (SD 0.29) and 0.81 (SD 0.09) across disease groups; specificity ranged between 0.95 (SD 0.06) and 0.98 (SD 0.02)." [NEEDS VERIFICATION: which specific gait-window classifier architecture — ElderNet-style self-supervised model vs. a lighter threshold heuristic — is feasible on Mark I's compute budget; not addressed in any source reviewed.]
Published fall-risk association / threshold: The World Falls Guidelines recommend "a cut-off value of <0.8 m/s on the basis of its predictive ability and simplicity" for fall-risk stratification. Studenski et al.'s pooled meta-analysis (~35,000 older adults, 9 cohorts) found a "pooled hazard ratio per 0.1 m/s, 0.88... P < .001" for survival, and that "predicted 10-year survival across the range of gait speeds ranged from 19% to 87% in men and from 35% to 91% in women."
Apple HealthKit reference implementation (external precedent, not Rx360's number): Apple's iPhone Mobility metrics derive walking speed from phone-based accelerometer/gyroscope during "periods of flat overground walking," validated against a pressure-mat gold standard (ProtoKinetics Zeno Walkway) with iPhones "at the hip (hip clip), in a front or back pocket, or in a waist bag" — a body-coupling assumption that does not transfer directly to a wrist-worn device. Apple's own validation reported strong reliability for walking speed: "Reliability Comparison of pressure-mat reference and iPhone walking-speed estimate (ICC(A,1)) 0.93 [Design] 0.92 [Validation]." Independent academic validation (Werner et al.) found "highest [agreement] for gait speed, with clinically acceptable PEs (11.6–14.1%) and good ICCs ranging from 0.85 to 0.86 in all three age groups" when comparing Apple Health app output to the APDM Mobility Lab reference system.

Stride length¶

Sensor(s): Mark I 6-axis IMU only (same hardware constraint as gait speed — no additional sensor is confirmed to support this metric specifically).
Standard computation method: Direct inertial-navigation approaches integrate acceleration twice with Zero-Velocity Update (ZUPT) correction during stance, but this is documented specifically for foot-worn IMUs: "A few studies perform a direct INS integration-based walking speed estimation algorithm... with IMUs attached to feet with a reported accuracy of about 3%." Rx360's sensor is wrist-worn, not foot-worn, so this ZUPT/double-integration approach is not a direct transfer — it is cited here as the gold-standard mechanism, not a validated wrist-applicable method. [NEEDS VERIFICATION: no source reviewed specifies a wrist-worn stride-length estimation method or its accuracy; this is an open engineering gap, not a citable number.]
Published fall-risk association / threshold: Not independently covered in the verified citation set for stride length specifically — the World Falls Guidelines and Studenski threshold above are gait-speed metrics, not stride-length metrics. [NEEDS VERIFICATION for a stride-length-specific fall-risk threshold.]
Apple HealthKit reference implementation: Apple's walking speed and step length algorithms are both derived from the same iPhone accelerometer/gyroscope pipeline and validated against the same Zeno Walkway pressure-mat study design described above ("All participants completed proctored overground walking tasks across an instrumented pressure mat... while carrying two iPhone devices"). Apple does not publish a stride-length-specific ICC in the verified quote set (only walking speed 0.92–0.93 and double support 0.53–0.59 are quoted numerically here); treat as directionally similar to step length, not independently verified. [NEEDS VERIFICATION for Apple's stride-length-specific ICC.]

Step length¶

Sensor(s): Mark I 6-axis IMU only.
Standard computation method: Same wrist-worn gait-window inference constraint as gait speed applies — step length is not independently derivable from a discrete heel-strike detector on the wrist the way it is on foot-worn sensors described in the biomechanics research (G-STRIDE's "1.6 m/s² upper threshold and 0.8 m/s² lower threshold" acceleration-magnitude approach is a foot-sensor method, cited here as the mechanical reference, not a wrist-transferable one). [NEEDS VERIFICATION: no wrist-specific step-length algorithm or accuracy figure was found in the verified citation set.]
Published fall-risk association / threshold: Not covered by a verified fall-risk-specific threshold for step length in isolation.
Apple HealthKit reference implementation: Apple computes step length as one of the four core iPhone Mobility metrics, validated in the same pressure-mat study referenced above, with iPhones carried at hip/pocket/waist (not wrist). Apple's independent validation (Werner et al.) found step length reliability weaker than gait speed and inconsistent across age groups, in the same study that reported double-support time "PEs were clinically acceptable in children (27.7%) and adults (18.4%) but not in seniors (31.6%)" — the seniors cohort is the directly relevant population for Rx360, and it is the weakest-performing group in Apple's own third-party validation.

Cadence¶

Sensor(s): Mark I 6-axis IMU only.
Standard computation method: Cadence is derived from step-timing series once gait windows/steps are segmented (see gait speed above for wrist-worn windowing approach — ElderNet 10-second windows at 30 Hz, or Kluge et al.'s 100 Hz accelerometer-only approach).
Published fall-risk association / threshold: Free-living wrist-accelerometer cadence data (STURDY cohort) found "every 10 steps/min higher cadence was associated with a 13.2% lower fall rate (p=0.036)" — notably, this is a wrist-accelerometer study, making it the most directly analogous published result to Rx360's own sensor placement of any metric in this document.
Apple HealthKit reference implementation: Not one of Apple's four published core Mobility metrics (walking speed, step length, double support time, walking asymmetry) in the verified citation set — Apple does not appear to publish a standalone cadence algorithm/validation in the sources reviewed. [NEEDS VERIFICATION for an Apple-specific cadence validation figure.]

Double-support time / double-support %¶

Sensor(s): Mark I 6-axis IMU only.
Standard computation method: Defined biomechanically as the "contact period (lasting 18% of the GC, from HS to Foot Flat (FFL))"-adjacent portion of the gait cycle where both feet are on the ground; computed from segmented stride/step timing series requiring "≥50 steps or ~1 minute of walking for stable variability estimates" per the gait-biomechanics synthesis. This is a foot-contact-timing metric that is mechanically harder to infer from a wrist sensor than from foot/waist sensors — flagged explicitly by Apple's own data below.
Published fall-risk association / threshold: Identified in the fall-risk/frailty IMU literature survey as one of the "most significant frailty gait parameters," alongside gait speed, stride time, and step time — but the same survey notes fall risk specifically (as opposed to frailty) is better discriminated by trunk-stability parameters, not double-support time. No absolute double-support-% cutoff tied to fall risk was found in the verified citation set. [NEEDS VERIFICATION for a numeric double-support fall-risk cutoff.]
Apple HealthKit reference implementation: Apple's white paper defines the metric directly: "The double support time metric provides a measure of the percentage of the gait cycle... that a user spends on two feet (double support)... Typical walking behavior ranges between 20 and 40 percent, with lower values indicating better balance." Critically, this is Apple's weakest-validated core metric: "Reliability Comparison of pressure-mat reference and iPhone double support–time estimate (ICC(A,1)) 0.59 [Design] 0.53 [Validation]" — versus 0.92–0.93 for walking speed. Independent validation (Werner et al.) confirms this weakness is worse in older adults specifically: "The lowest level of agreement was observed for double support time... ICCs ranged from poor in seniors (0.42) to moderate in children (0.54) and adults (0.58)." This is a direct engineering flag: double-support time is the metric where a phone/wrist-based competitor is weakest, and where Rx360 has the least room to simply copy Apple's approach and expect comparable accuracy in its target senior population.

Walking asymmetry¶

Sensor(s): Mark I 6-axis IMU only.
Standard computation method: Requires reliable left/right step-timing segmentation from the same gait-window/stride-segmentation pipeline described for gait speed above (hierarchical HMM segmentation "significantly outperformed the other methods with an accuracy of 96%" for heterogeneous/pathological gait per the Parkinson's segmentation-methods comparison — that study used a shin/foot-worn Shimmer IMU at 102.4 Hz, not wrist).
Published fall-risk association / threshold: No walking-asymmetry-specific numeric fall-risk cutoff was found in the verified citation set. Related but distinct: step-width CV (not step-time asymmetry) had "a cut-off value... calculated to be 21.7%" in a 446-participant cohort with AUC 0.715 — the strongest single gait-variability discriminator in that study, but it is a width-variability metric, not the timing-asymmetry metric Rx360's Layer 2 currently names. [NEEDS VERIFICATION: no source directly validates a left/right step-timing asymmetry threshold against fall outcomes.]
Apple HealthKit reference implementation: Walking asymmetry is one of Apple's four core Mobility metrics, computed and validated via the same iPhone pressure-mat methodology described above. Per the current-state grounding, Apple's own numeric ICC for asymmetry specifically was not among the two figures captured in the verified quotes (only walking speed and double-support ICCs are quoted numerically here); treat Apple's asymmetry validation as directionally similar to double-support time's weaker performance rather than walking speed's stronger performance, pending a source with the exact number. [NEEDS VERIFICATION for Apple's walking-asymmetry ICC figure specifically.]

3. Sit-to-stand (chair-rise) time/power¶

Sit-to-stand carries distinct fall-risk relevance and a categorically different sensor/computation profile from ambulatory gait metrics — it is a discrete transition event, not a continuous walking-bout signal, and the clinical/research literature instruments it differently (thigh-mounted, not wrist-mounted, in the highest-fidelity validation study reviewed).

Sensor(s): Mark I 6-axis IMU only — worn at the wrist, not the thigh. This is the critical caveat for this metric: the current-state grounding itself flags as an unresolved open question whether sit-to-stand (and turns) "can even be derived passively (no prompted test)" from Mark I's confirmed sensor set. No accelerometer/gyroscope specification in the current-state grounding places any Mark I or Mark II sensor at the thigh, waist, or lower back.
Standard computation method: The highest-fidelity validated method uses "a single IMU placed on the lateral face of the thigh" at 500 Hz, validated against motion-capture + force-plate gold standard, finding "an extremely large correlation for the total time trial, mean concentric time, and mean concentric force (r = 0.99, r = 0.93, and r = 0.97, respectively) and a very large correlation for mean concentric velocity (r = 0.76), and mean concentric power (r = 0.79)," concluding the thigh IMU is "a promising practical alternative to the gold standard methods." This is a thigh-worn method and does not transfer to Rx360's wrist placement without independent validation. Separately, a systematic review of sensor-based fall-risk assessment notes wrist placement is rare and unproven for this specific transition: "Only one study group used wrist-worn sensors for detection of sit-to-stand transitions, but the performance was comparable to studies using waist-worn devices" — a single-study data point, not a validated consensus method, and the specific accuracy figures from that study were not captured in the verified citation set. [NEEDS VERIFICATION: no wrist-specific sit-to-stand detection algorithm or accuracy number beyond this single unquantified reference was found.]
Published fall-risk association / threshold: The World Falls Guidelines cite a related test (Timed Up and Go) at ">15 seconds, although evidence for fall risk stratification is mixed." For the Five-Times-Sit-to-Stand test specifically, "the commonly cited fall-risk cutoff for the clinical (stopwatch) FTSTS is >15 seconds for community-dwelling older adults," with IMU-literature-reported typical total times "comparable to those found in the literature (from ~8 s to ~15 s, for not fallers and fallers of both sexes, 65–90 years old)." In the 446-participant cohort comparing multiple fall-risk predictors, 5TSTS had "an AUC of 0.692," meaningfully lower than TUG (0.708) or step-width CV (0.715) in the same dataset — i.e., even with gold-standard thigh-mounted instrumentation, sit-to-stand is a secondary, not primary, fall-risk discriminator relative to gait-based metrics.
Apple HealthKit reference implementation: Not applicable — sit-to-stand transition timing is not one of Apple's published Mobility metrics or e6MWD algorithm inputs in the verified citation set. Apple's e6MWD algorithm does use "accelerometer, gyroscope, barometer, and GPS" plus derived metrics like "flights climbed, steps, exercise minutes, walking distances, estimated step length, and walking speed," validated against in-clinic 6MWTs in adults 65+ with "ICC was 0.926" — cited here only as the nearest Apple analog for a composite mobility-endurance metric, not as sit-to-stand precedent. [No Apple sit-to-stand reference exists in the verified set — do not imply one.]

Bottom line for engineering: sit-to-stand is the one Layer 2 metric where the best available published validation (r=0.93–0.99) comes from a sensor location and mounting Rx360 does not have, and the one wrist-worn data point available is a single unquantified study reference, not a validated method. This is the highest-risk metric in the current Layer 2 signal list to ship without additional validation work or a scope reduction (see Section 6).

4. Competitive positioning¶

Platform	What it measures	How (method/sensor)	Validation posture
Rx360 (Mark I / Mark II)	Gait speed-delta from personal baseline, cadence, walking asymmetry, step-timing variability, double-support time, sway, sit-to-stand time, turn smoothness	Wrist-worn 6-axis IMU (LSM6DSV320, 320 Hz), continuous/passive; Mark II adds barometer + PPG + wear-detection for the broader Balance Meter but these are not confirmed gait-metric inputs	Mechanism-locked, numbers pending pilot calibration — no independent validation yet (per current-state grounding)
Apple Watch / iPhone HealthKit	Walking speed, step length, double support time, walking asymmetry (iPhone Mobility metrics, hip/pocket-worn); Walking Steadiness (OK/Low/Very Low) fall-risk-insight classifier; e6MWD; binary hard-fall detection (Watch, wrist-worn)	Phone accelerometer/gyroscope with pressure-mat (Zeno Walkway) validation for the four core metrics; Watch fall detection uses a "next-generation accelerometer and gyroscope, which measures up to 32 g-forces" and works "by analyzing wrist trajectory and impact acceleration"	Four core metrics have a published white paper + ICC statistics (0.53–0.93 depending on metric); Walking Steadiness itself has no published algorithmic validation — outside researchers say Apple discloses only "a narrow glimpse"; fall detection sensitivity is highly mechanism-dependent (4.7% for wheelchair falls in one study) and Apple states plainly "Apple Watch cannot detect all falls"
Zibrio (Balance Score)	Static postural sway / center-of-pressure, output as a 1–10 fall-risk score	Passive pressure sensors under the feet in a bathroom-scale form factor, ~1 minute standing assessment, descended from NASA JSC posturography research	Peer-reviewed sensor accuracy validation vs. force platform (mean COP error ≤0.5mm); prospective field validity claimed ("users who weren't told their score were found to experience falls at the expected rate") but the strongest field anecdote is explicitly uncontrolled
Clario Precision Motion (Opal V2C, formerly APDM)	Gait, balance, and actigraphy — "pre-configured standard assessments of instrumented tests embedded in software with more than 140 validated outcome measures," incl. a 2-minute walk test	Multiple wearable Opal sensors, protocol-driven research/clinical assessment, not continuous passive wear [sensor sampling rate and body-placement locations: NEEDS VERIFICATION — not found in sources checked]	Research/clinical-grade gold standard: "more than 1,000 researchers worldwide have used Opal technology... 700+ peer-reviewed scientific publications... over 45,000 scientific citations"; "validated across many indications, including Parkinson's Disease (PD), Multiple Sclerosis (MS), Spinocerebellar Ataxia (SCA), and more"
BioSensics LEGSys	Gait and fall-risk assessment — "35+ independent parameters of gait and balance"	Body-worn IMU sensors, protocol-driven clinical assessment [sensor count: NEEDS VERIFICATION — "5 sensors" claimed in some secondary sources but not confirmed on BioSensics' own product page]	"The first FDA-listed wearable-based medical device for objective assessment of gait and fall risk"
Traditional PERS (Life Alert, Medical Guardian, Bay Alarm Medical)	Binary hard-fall event, not a trending gait/balance score — Medical Guardian's own materials describe a multi-variable evaluation ("body position, physical activity, acceleration or movement smoothness") rather than a pure single-axis threshold, but it is still a per-event trigger, not a personal-baseline gait signal	Triaxial accelerometer measuring "vibrations in three directions: X: Across the waist, Y: From head to toe, Z: From the posterior upward"; Life Alert has no automatic detection at all (manual button press only)	No gait/balance validation published by any of the three; Medical Guardian's own materials describe slower falls (e.g. "sliding off a sofa or bed onto the floor") as harder to detect, and frame a brief evaluation window as a mitigation for drop-triggered false alarms

5. Recommendation — what Rx360 should adopt vs. where we differ¶

The following is an engineering recommendation for discussion, not a settled architecture decision. Nothing below should be read as locking a threshold, sensor claim, or algorithm choice — those remain pilot-gated per the current-state grounding.

Adopt the wrist-appropriate gait-window paradigm, not the foot/waist-sensor event-detection paradigm. Mark I is wrist-worn only. The literature is explicit that wrist gait sensing is a different, harder problem than foot/waist sensing ("Gait detection from a wrist-worn accelerometer is more challenging... due to the non-gait related hand movement"). Recommend building on the ElderNet-style windowed/self-supervised approach (10-second windows, 30 Hz-equivalent processing) or the Kluge et al. 100 Hz accelerometer-only real-world validation as the nearer analogs, rather than foot-worn ZUPT/double-integration or G-STRIDE-style stance-phase thresholding, which assume a sensor location Rx360 does not have.
Treat cadence as the strongest early wrist-worn fall-risk signal to prioritize. It is the one metric in this document with a fall-risk association published from a wrist-accelerometer study specifically (STURDY cohort, 13.2% lower fall rate per 10 steps/min). Recommend it be a first-tier signal in the Layer 2 scoring weight, ahead of double-support time and asymmetry where wrist-based accuracy is far less proven.
Do not over-invest early engineering effort in double-support time and walking asymmetry precision. These are Apple's own weakest-validated metrics even from a pocket-worn phone (double support ICC 0.53–0.59, worse in seniors specifically — the target population), and Rx360's wrist placement is a harder sensing problem than Apple's pocket/hip placement. Recommend flagging both as lower-confidence contributors to the Layer 2 band until pilot data says otherwise, and being conservative in any external claims about their precision.
Sit-to-stand needs a scope decision before further build. The best available literature validation for sit-to-stand power/timing comes from a thigh-mounted sensor Rx360 does not have; the only wrist-worn precedent found is a single unquantified reference ("performance was comparable to... waist-worn devices," no numbers). Recommend either (a) scoping sit-to-stand out of Mark I's near-term Layer 2 outputs until a dedicated validation pass is run, or (b) explicitly labeling it lowest-confidence in the UI/scoring band language, consistent with the "not-yet-numerically-locked" status already assigned to gait thresholds generally.
Positioning differentiator: continuous ambulatory signal, not static sway or post-hoc impact. The competitive table in Section 4 shows a real gap between (a) research-grade multi-sensor systems (APDM, BioSensics) that require protocols and dedicated hardware, (b) Zibrio's static once-in-a-while sway score, and (c) PERS/smartwatch post-hoc impact detection. A continuously-worn wrist IMU deriving a trending, personal-baseline-relative gait signal is a genuine positioning niche distinct from all three — but this differentiation claim should stay internal/engineering-facing until Rx360 has its own validation data; it is a market-structure observation, not evidence Rx360's algorithm outperforms any named competitor.
Do not claim barometer- or PPG-based fall/gait signal on Mark I. Per the current-state grounding, barometric pressure and PPG are Mark II-only. Any recommendation implying height-loss detection, orthostatic stand-check, or wear-verification as part of the Mark I gait pipeline would misstate confirmed hardware — Mark I's entire Layer 2 gait signal must be described as IMU-only.

6. Open items for engineering / Mike¶

Bout/walking-window detection criteria. No windowing or filter specification exists for the gait DSP today ("DSP per walking bout" with no window length, filter type, or sample-rate-to-feature detail). Recommend evaluating ElderNet-style fixed 10-second non-overlapping windows vs. a lighter onboard heuristic, given Mark I's compute/power budget — decision needed before any gait metric can be implemented at all. This remains, per the current-state grounding, the single most important unbuilt prototype.
Sample-rate/downsampling decision. Mark I's IMU natively samples at 320 Hz; ElderNet's reference model was trained at 30 Hz, Kluge et al.'s wrist validation used 100 Hz. A decision is needed on whether to downsample on-device (power/compute savings) or process at native rate, and whether Mark II's unchanged IMU spec means the same DSP pipeline can be shared across both hardware tiers.
Gait-event/segmentation algorithm choice. Peak detection, DTW template matching, and hierarchical HMM segmentation are the three documented options, with hHMM outperforming others (96% F-score) specifically for heterogeneous/pathological gait — but that result comes from a shin-worn sensor, not wrist. Needs a wrist-specific bake-off, not a direct port.
Sit-to-stand build/no-build decision. Per Section 3, no validated wrist-worn method exists in the literature reviewed. Mike needs to decide whether to (a) commission a dedicated validation study, (b) scope sit-to-stand out of near-term Layer 2 output, or (c) ship it pilot-labeled low-confidence.
Delta-gate/drift-threshold reconciliation. The two conflicting numbers already flagged in current-state grounding (Sensor/DataFlow spec's ~18%/25% vs. Canonical Model's ~12%/15%) are not resolved by anything in this document — neither figure has independent published support in the verified citation set, and both remain pilot-calibration-gated. This spec does not adjudicate between them.
On-device compute budget. No confirmed MCU/DSP compute-budget figures were available in the current-state grounding or research summaries for this pass — needed before committing to any specific segmentation algorithm's runtime cost (hHMM vs. simpler peak detection vs. a learned model). [NEEDS VERIFICATION — outside the scope of the six source files and research reviewed here; flag to firmware/hardware team directly.]
Validation plan. Recommend a phased plan: (1) internal bench validation of gait-window detection against a reference wearable (e.g., research-grade IMU) before pilot; (2) pilot-phase comparison of Rx360's gait-speed/cadence output against self-reported or observed walking bouts to sanity-check direction/magnitude, given no in-house gold-standard force-plate or pressure-mat access is indicated anywhere in the source documents; (3) explicit exclusion of sit-to-stand and walking-asymmetry from any external accuracy claim until each has its own validation pass, given the weak/absent precedent found in this research.
Confluence reconciliation carryover (not new to this doc, but relevant context): the capacitive-touch part-number conflict, the magnetometer confirm-flag inconsistency, the PPG tier-status inconsistency, and the Mark I fall-detection-path history all remain open per the current-state grounding and are outside this document's scope — noting them here only so engineering doesn't assume this spec resolves them.