Hush — a wearable sensory organ that streams ternary bands, not video, proven on a two-node bench

Most AI wearables of this era ship a microphone and a cloud API key. Hush is designed around a different premise: encode what the wearer's body and room are doing directly into the representation their own model trains on, so there is no intermediate video or audio format for anyone else's tooling to decode. This report keeps the production poster and the bench reality on separate footing, on purpose.

What it was

Hush is designed as the "chest organ" of a wider Novaterra sensory-organ product family (Onira, Vesper, Halo, Mote, Hush, Conch, plus planned Optic/Pulse/Band) — a pendant worn at the sternum that senses heart rate, breath, posture, room acoustics, Wi-Fi CSI, and an occasional photo, and encodes all of it as wavetrinity ternary band coefficients: an 8×8 Walsh-Hadamard transform for vision, an 18-band Bark filterbank for audio, both quantized to a ternary {−,0,+} alphabet and packed five trits per byte, streamed over a binary wire protocol (NovaP/1) with no video codec, no audio codec, no JSON, and no decoder on the receiving end other than the wearer's own trained model. The production design specifies a titanium-and-sapphire shell, a five-state hardware privacy ring, a thermoelectric body-heat trickle-charger, and — on a $349 "Sage" tier — a per-pendant calibrated diamond coded-aperture in front of the lens, so even the raw pixels are structurally unreadable without a calibrated PSF inverse.

The framing behind all of it is stated plainly in the project's own origin brainstorm: the receiver is a neural net, not a human eye or ear, so most of what a video or audio codec spends its bits on — spatial prediction tuned to human vision, deblocking, the whole apparatus built to make a decoded frame look right to a person — is overhead a model doesn't need. "The bitstream IS the model input" is the literal design goal: no decoder anywhere in the loop except the wearer's own trained model. That idea is graded honestly in its own source documents, too — the brainstorm that proposed it states a 0.70–0.80 posterior on the whole approach working at the necklace-plus-model scale, not a certainty.

The audio side of the same bet has its own, more specific brief: an 18-band Bark filterbank (not the simpler linear-mel scheme the first draft used), 20ms windows with 50% overlap, temporal differencing between ternary frames for a further ~50% cut, and a dual-criterion voice-activity gate — a design that took the wire rate from an original 25 kbps target down to 6–10 kbps on paper. None of it has run on hardware yet; see "the honest gaps" below.

What we built

🔬Research

Not the pendant — a two-node bench prototype that proves the wiring underneath it, built bottom-up from stock parts: a LilyGO TTGO T-Display (ESP32-PICO-D4) plus a 37-sensor kit as the "wearer node," and a separate AI-Thinker ESP32-CAM as the "eye node," both talking NovaP over plain Wi-Fi TCP to a Python receiver on a laptop. Milestones actually closed, per the project's own bring-up log:

• A five-state privacy-ring finite-state machine and a tap / double-tap / long-press gesture engine, running on the TTGO.
• Live heart-rate peak-detection from a KY-039 PPG sensor, displayed on the TTGO's own screen.
• A working NovaP/1 Wi-Fi uplink carrying a heart-rate channel and a one-byte "Marker" event (the "this moment mattered" tap), received and decoded live by the laptop receiver — verified against real serial and console logs ("HEART bpm=72", "MARKER kind=0 intensity=160").
• On the separate camera node: a real 96×96 greyscale capture at up to ~16 fps, transformed to 8×8 Walsh-Hadamard patches, ternary-thresholded, packed five trits per byte, and streamed as NovaP camera frames — received and shown live as an 8×8 intensity heatmap on the laptop. The actual serial-console log for this is exact, not rounded: Hush-CAM booting fw=hush-cam-0.1.0, then Camera OK (96x96 grey), then live frames at CAM len=410 trits=2050 nonzero=1432 rate=15.8fps.

That is seven bench milestones done (of a longer list including the audio path and full-image reconstruction), against a production roadmap that specifies a seven-stage pipeline — poster, requirements, architecture, design, implementation, test, market — of which only the first four stages have a written document at all, and only the fifth (implementation) has running hardware behind it.

The bench-first sequencing wasn't an accident — a design council explicitly asked "what can ship on the user's bench this weekend?" and answered it with a five-tier cost ladder that the $129/$349 poster prices sit on top of: an A0 bench build using only parts already on hand ($0, buildable in a weekend or two), an A1 bench build with ~$25–50 of added sensors, a ~$33–38 shippable dev-kit tier, a ~$57.20 costed consumer BOM from the same council that wrote the architecture document, and the $115-ish parts basis under the $349 "Sage" tier. That ladder is a real, itemized parts list (down to which off-the-shelf PPG and tap sensors were priced and where) — it's still not a supplier-quoted, volume-costed BOM, but it's a materially more grounded number than the poster's price alone.

What we learned — including the honest gaps

• The privacy architecture is real code, not just a diagram. The five-state ring and the marker byte both exist as working firmware state machines today, which is a stronger claim than most wearable privacy stories, which stay in the slide deck.
• "Direct to training, no decoder" is true of the format, untested as a security claim. The wire format genuinely has no generic decoder — but nobody has tried to break it. A design property and a red-teamed security property are different claims, and only the first is backed by evidence here.
• Image reconstruction — the step that would prove the codec is information-preserving — was the next milestone, not a closed one. Streaming ternary bands to a heatmap shows the pipeline runs; it does not yet show the bands reconstruct back into a recognisable image, which is the harder and more interesting claim.
• The gap between bench and product is the whole hardware-design discipline, not a rounding error. Two loose dev boards on a desk and a milled titanium pendant with a calibrated diamond aperture are related by architecture document, not yet by a single line of shared engineering.
• The wire protocol is a family decision, not a Hush-only choice — and it cost Hush's team a rewrite to get there. Hush's own bring-up originally spoke a homegrown protocol (internally "NTP/1"); the design council later deprecated it in favour of adopting the shared NovaP/1 spec, on the stated logic that "the receiver is the customer." Conch, a separate Novaterra earpiece, already reserves its own NovaP channel range under the same spec, which is real evidence the "one family, one wire format" claim is more than an aspiration for two of the roughly five planned organs. The honest gap: the spec's own designated receiver implementation still doesn't exist (see the dedicated NovaP/1 report), so "shared protocol across the family" today means one working encoder plus a written contract, not an interoperability test between independent teams' code.

Where it went / status

Active bench research. The wiring, protocol, and privacy state machine are real and running; the pendant itself is a north-star document with an implementation folder growing underneath it, milestone by milestone. Tiered research, not shipped — and deliberately not rounded up to match the confidence of its own poster.