Your "Soul Signature" — biometric identity from passive WiFi, no camera, no wearable

Project: RuView / WiFi-DensePose · Date: 2026-05-24 · Spec: docs/research/soul/ in the repo

The pitch in one paragraph

Your body is constantly bouncing radio waves around the room you're in. WiFi routers fill that room with 2.4 / 5 / 6 GHz signals; you reflect, absorb, and re-scatter them in a pattern that's measurably unique to you — your heartbeat, your breathing, your posture, your gait, your skeletal proportions, the way your chest cavity flexes when you inhale. RuView captures those reflections from a $9 ESP32 sensor and fuses them into what we call a Soul Signature — a single per-person electromagnetic fingerprint. No camera. No microphone. No fingerprint reader. No wearable on your wrist or in your ear. The system just notices you in the room and knows who you are — through walls, in the dark, while you sleep.

The spec is open. The implementation runs on $9 hardware. The privacy story is what makes it interesting.

What's actually in a "soul signature"

Seven separable channels, all measurable today with shipped code in RuView:

Channel	What it captures	Why it's distinctive
AETHER embedding (128-dim)	Learned contrastive fingerprint from CSI subcarrier patterns	Like FaceID's "face vector" but for your RF backscatter
Cardiac profile	Heart rate baseline + heart-rate variability shape	Your HR shape is as personal as a signature
Cardiac waveform morphology	Wavelet decomposition of your filtered chest-wall motion	The micro-shape of your heartbeat differs per person
Respiratory pattern	Breathing rate + depth + paradoxical-breathing index	Asthma, COPD, fitness, posture — all leak into this
Gait timing	Cadence + stride period variance + asymmetry	Forensic gait analysis is established; this is the contactless version
Skeletal proportions	Torso/limb ratios from 17-keypoint pose	Bone proportions don't change after adolescence
Subcarrier reflection profile	Per-subcarrier amplitude + phase shift when you stand still	Your body's bulk dielectric properties

The seven channels are fused into one RuVector Format (RVF) graph — a single content-addressed file (signature-<sha256>.rvf) that's the canonical "this is you" object. Matching across signatures is a weighted cosine similarity that gracefully degrades when channels are missing (e.g., you're seated and we can't see your gait — fine, match on the other six).

How you'd "enroll" your signature

A structured 60-second scan, just once per ~3 months for adults:

T+0–10s   Empty-room calibration (you step aside)
T+10–25s  Stand in scan zone, deep-breathe 5x      → captures respiratory pattern
T+25–35s  Sit at rest                              → captures cardiac baseline + HRV
T+35–50s  Walk 2m back-and-forth, twice            → captures gait + dynamic skeleton
T+50–60s  Three 90° static rotations              → captures multipath reflection
                                                     (AETHER embedding accumulates throughout)

Output: one .rvf file. ~50 KB. Cryptographically signed with the Ed25519 witness chain RuView already uses for ESP32 firmware attestation.

After enrollment, the system gets BETTER at recognizing you over weeks of normal occupancy — it converges, it never drifts away. A 90-day re-scan refreshes the cardiac/respiratory channels which can change with fitness. A clinical-event flag triggers immediate re-scan in healthcare deployments.

What this is NOT

This is the section where most "magic biometric" announcements deserve their skepticism. Here's what the Soul Signature explicitly does NOT claim:

Not unbreakable. A motivated attacker with a phased-array vest can spoof a stored signature. Mitigation is in the threat model.
Not equal in false-accept rate to FBI-tier fingerprints. Open research; baseline TBD. We have AETHER's published numbers at small N=5 (~80% mAP) but the household-scale N=20–100 false-accept rate has never been measured. That's the blocking research question before any commercial deployment threshold can be calibrated.
Not legal evidence in court without orthogonal corroboration.
Not a replacement for explicit consent in regulated contexts (GDPR / HIPAA / California CMIA). The spec has a hard-enforced enrollment-vs-recognition distinction: ambient WiFi capture cannot create an enrollment. Bystanders get a 24-hour TTL track that auto-prunes.
Not metaphysics. The word "soul" is marketing. The thing under it is a 1,000-dimensional vector backed by 7 measurement channels, all of which physics review boards already accept individually.

If we couldn't defend a claim to an EE PhD reviewer, it's not in the spec.

Why this is interesting (the "SOTA hook")

What fingerprints were to identity in 1900, what iris scans were in 1990, what FaceID was in 2017 — a passive-at-distance, through-walls, no-line-of-sight biometric is what 2026+ is asking for. The smart-home market is begging for "the lights / thermostat / alerts behave differently when it's me vs my partner vs a guest" without making every resident carry a tag, scan a face, or remember a code.

The RuView stack is the first one I know of that can actually do this on $9 of hardware. The just-merged ADR-115 (Home Assistant + Matter integration) shipped 10 inferred semantic primitives (someone-sleeping, possible-distress, fall-risk-elevated, elderly-inactivity-anomaly, etc.) — but those primitives are per-room, not per-person. The soul signature is the layer that turns them per-person. That's the unlock.

How the security actually works

Five layers, all referenceable in the repo:

Encryption at rest — .rvf files encrypted with ChaCha20-Poly1305, key derived via Argon2id from a passphrase or FIDO2 hardware key. The Cognitum Seed appliance never stores the decryption key.
Ed25519 witness chain — every enrollment event is cryptographically attested using the same chain ESP32-C6 firmware updates use (ADR-110).
Field-model replay break — the signature incorporates the room's persistent field eigenvectors (ADR-030). A replay attack in a different room produces a different cross-validation against the field model. Replays don't work.
Adversarial-CSI detection — RuView's existing adversarial.rs module detects physically-impossible signal patterns (e.g., phased-array vest output). Enrollments route through it.
Differential privacy at training — when re-training the AETHER backbone on enrolled-person data, DP-SGD (ADR-106) prevents membership inference from leaked model weights.

Plus --privacy-mode (already in ADR-115): biometric data NEVER leaves the local device when this flag is set. The Home Assistant integration can still publish the inferred states (someone-sleeping, fall-risk-elevated) because those are derived states, not raw biometric values.

Where to read the actual code + spec

What	Where
The repo	https://github.com/ruvnet/RuView
Soul signature research spec (5 files, ~1,450 lines)	`docs/research/soul/`
Just-merged HA + Matter integration	ADR-115 + release notes
The 128-dim AETHER embedding (foundational ID channel)	ADR-024 — contrastive CSI embedding model
The cross-room generalization layer	ADR-027 MERIDIAN — cross-environment domain generalization
The room's persistent field model (replay defense)	ADR-030 — persistent field model + drift detection
The cryptographic witness chain (Ed25519)	ADR-110
The vital-sign pipeline (HR + BR extraction from CSI)	ADR-021
The RVF binary container format	`v2/crates/wifi-densepose-sensing-server/src/rvf_container.rs`
The 17-keypoint pose ground-truth pipeline	ADR-079
The published pretrained model on Hugging Face	`ruvnet/wifi-densepose-pretrained`
The $9 ESP32-S3 firmware	`firmware/esp32-csi-node/` in the repo

What the use cases actually look like

Aging-in-place / AAL — the system knows it's grandma vs the caregiver vs a stranger. Fall detection is tuned per-resident; baseline HR is per-person; the carer's phone only pings on grandma's distress, not the carer's own brief stumble.
Smart home that doesn't need to be told who's home — lights / climate / music adjust by who's actually in the room, even if you walked in without your phone.
Healthcare passive monitoring — clinically meaningful vitals tagged to the right patient automatically, even when two are in the same room.
Frictionless authentication for low-stakes actions — "unlock the kitchen tablet for the resident, never for guests" without typing a PIN.
Two-person automations — "when both partners are home AND someone's sleeping, drop hallway lights to 10%."

What it explicitly does NOT do well: replace something high-stakes like bank-account authorization. The false-accept rate at the relevant N is open research; it's an additive signal, not a sole authority.

The honest unknowns

N=20–100 household-scale FAR/FRR curve — never measured. Blocking item before any deployment threshold.
Child re-enrollment cadence — kids' skeletal proportions change; spec says 30 days but that's a guess.
Identical twins — open research; conventional wisdom says cardiac waveform morphology differs but no one's measured at WiFi-CSI fidelity.
Pets — the spec treats them as nuisance signal; whether they could be fingerprinted similarly is an interesting follow-on but not in scope.
Adversarial robustness against trained ML attackers — the existing adversarial.rs is heuristic; an ML-vs-ML arms race is foreseeable.

These unknowns are catalogued in docs/research/soul/security.md and docs/research/soul/references.md with explicit "open research; baseline TBD" markers.

tl;dr

A real, measurable, electromagnetic-only biometric fingerprint of you — derived from how your body interacts with the WiFi signals already in your home — fused into a single signed container, encrypted at rest, scannable in 60 seconds, runnable on $9 of hardware, with hard privacy enforcement at the wire boundary. Spec is open. The household-scale accuracy number is the open research item.

If "your house knows it's you the moment you walk in, no phone, no scan, no key" sounds interesting, the repo is at https://github.com/ruvnet/RuView and the spec is in docs/research/soul/.

🤖 Generated by an autonomous agent working in the RuView repo, 2026-05-24.

ruvnet/soul-signature-gist.md

Select an option

No results found