# Aero device system · engineering preview 02

## What exists now

Aero can draft code from a typed request or supported dictation using its browser model. The display window mirrors draft text locally within the same browser profile; it does not carry the rest of the thread, execute code, or expose a remote endpoint. Move the window to any external display the operating system already recognizes. Both device packages include dimensioned CAD records, named STEP assemblies, and printable fit gauges. The two routed boards and Nordic firmware are bench fixtures, not a manufactured wearable. No battery life, ingress protection, optical field of view or physiological decoding performance has been measured on these designs.

## View PB01 · everyday display frame

- Frame front: 135.2 mm; bridge: 19.2 mm. Assembly envelope in the shipped CAD: 141.29 × 43.10 × 138.26 mm. The clear prescription insert and the 40.8 × 30.3 × 0.82 mm waveguide gauge are separate replaceable concepts.
- Each optical cartridge reserves a nominal 9.70 × 6.25 × 5.80 mm rectangular envelope. That is a packaging allowance only. Final cartridge geometry, pupil alignment, eye relief, coatings, cabling and thermal paths depend on supplier drawings. The fit kit checks nominal solids and the clear insert stack; a full fold, contact tolerances and supplier part interfaces remain open.
- Prototype materials to validate: PA12 or TR90 body; abrasion resistant clear polymer lens gauges; matte 6061 aluminium cartridge trim with an RF transparent temple section. Use ±0.20 mm for early gauge features and an adjustable gauge at the optical interface until actual tolerance and pupil-position sweeps set the datum requirements.
- Two display paths are deliberately separate: quiet glanceable status and selected source lines for a compact heads-up display, and full text through an OS-recognized external display. The first is a hardware target. The second works with the separate browser window today; compatibility depends on the host and the display. Bluetooth LE is for controls and fixtures, not full video. A production external-display implementation would need a verified DisplayPort path, its own power budget and optical module.
- Reference points: [JBD’s compact VGA microdisplay](https://www.jb-display.com/product_des/3.html) lists 640 × 480 and 60 mW typical for the display alone; the complete optical engine, controller and radio consume additional power. [XREAL’s One series specifications](https://tutorials.xreal.com/docs/glasses/one-series/spec/) illustrate a separate, larger external-display class. [Dispelix](https://dispelix.com/) is a waveguide sourcing candidate. None of these companies supply or endorse the current CAD.

First physical gate: a headform fit and hinge/cable sweep, then an optical bench with a supplier-provided module. Compare center and edge contrast, usable eyebox for different headforms, stray light, color, text legibility, surface temperature and mass distribution. Stop before a wearable trial until optical safety and ergonomics have been reviewed. Do not equate emitter specifications with useful brightness at the eye.

## Signal R02 · modular EEG research fit

- The inert CAD assembly fits in a 199 × 183.5 × 82.7 mm envelope. Eight upper contact shoes cover central scalp zones; four positions per ear on removable carriers provide an optional comparison array. Two reference sites and one bias site are packaging candidates. The coordinates in the fit kit are headform targets, not anatomical coordinates for use on people.
- Proposed constructions: soft TPU ear carrier, removable PA12 crown rails, closed-cell pad, electrically inert blank contact plugs for fitting. The shipped rails represent the integration envelope. Tune travel, pressure, cable strain relief, cleaning, connector retention and tolerances on a headform. Do not wire the inert fit plugs or use the fixture with a person.
- An eventual acquisition platform would need synchronized front ends, channel/reference configuration, documented input protection and compliance testing. The [TI ADS1299](https://www.ti.com/lit/ds/symlink/ads1299.pdf) supports eight simultaneous channels at rates starting at 250 SPS. A two-device concept would require clock synchronization and a qualified safety design. No such body-connected board is part of this release. [OpenBCI’s Cyton format](https://docs.openbci.com/Cyton/CytonDataFormat/) documents a separate reference acquisition platform with 125 Hz live multiplexed Daisy data; that stream is not interchangeable with this 250 Hz fixture.
- The fixture streams four rows per 216-byte ASG2 notification. Each row contains sixteen signed 24-bit sample values. A 24-byte header holds version, sample index, fixture clock, contact bitmap, motion value and CRC-32. At 250 rows/s these consume 12,000 B/s raw or 13,500 B/s at the application layer, plus radio overhead. The prototype requests a 247-byte ATT MTU; transmission requires an effective MTU of at least 219. Verify negotiated data length, connection interval, delivery jitter and disconnect behavior on each real phone and laptop. No runtime radio throughput measurement is available yet.
- The UI starts disarmed. Two consecutive 2-second windows with the same test cue are required for one navigation step. Missing or replayed data, stale timing, a failed checksum, mains pickup, excessive motion, or an open central contact resets the gate. Optional ear loss alone does not disable the central fixture. The gate cannot submit prompts, run source, or make decisions outside the display.

### What the EEG experiment measured

The [PhysioNet EEG Motor Movement/Imagery Dataset](https://physionet.org/content/eegmmidb/1.0.0/) provided three runs for each of 12 subjects. Per-person runs were split for training, calibration and testing, rotating which run was held out. The task was rest, left imagined fist, or right imagined fist. The run-level comparison used 1,080 held-out windows of existing scalp data. This was a small, exploratory study; it did not use the Aero headset, validate ear placement or test on unseen people.

| Regularized filter-bank estimate | Balanced accuracy |
| --- | ---: |
| 2 central scalp channels | 43.02% |
| 8 central scalp channels | 45.17% |
| 16 scalp channels | 41.54% |

The 8-channel gate rejected 298 windows for quality. It still marked 7.04% of rest windows as a movement, and recalled only 12.22% of intended cues. Those rates are too weak to justify live control. We keep the 8 extra ear positions as an optional experiment. They have not been compared in this dataset. Published [ear EEG work](https://doi.org/10.1038/srep16743) addresses different tasks, and [Brain2Qwerty](https://arxiv.org/abs/2502.17480) involved participants typing. Neither supports inferring arbitrary instructions or writing code from unrestricted thoughts. The demonstration uses a generated fixture solely for cursor navigation. Use speech, the keyboard or touch to supply instructions.

The study retains results, split protocol, hashes and repro source. The kit does not redistribute the raw EEG recordings. Cite G. Schalk, [EEG Motor Movement/Imagery Dataset v1.0.0](https://doi.org/10.13026/C28G6P), PhysioNet (2009).

## Fabrication and bench acceptance

RX-AS-P02-R01 (48 × 63 mm) connects a known artificial source through sixteen 10 kΩ series resistors to a bench ADC and routes a separate test reference. RX-AV-P01-R01 (46 × 43 mm) is a 3.3 V logic breakout with pull defaults, not a projector driver. Both include KiCad source, footprint libraries, board labels, BOMs, copper/mask/silkscreen/profile Gerbers and drill maps. A standard 1.6 mm two-layer FR-4 stack is the preliminary fabrication assumption. Use the board drawings for connector orientation and verify fit against the actual headers. [JLCPCB](https://jlcpcb.com/capabilities/pcb-capabilities) and [PCBWay](https://www.pcbway.com/capabilities.html) are candidates; no order has been placed.

1. Before fitting headers, compare drill and edge dimensions with the board profile. Measure expected through resistance and check every lane against opens, bridges, polarity and cross-talk using calibrated artificial sources. A common reference on the bench is not isolation.
2. For the glasses logic fixture, measure idle pull levels before clocking. Check reset sequencing, logic edge timing, connector retention and supply current on a dummy receiver. Do not connect a bare eye-facing emitter to the breakout.
3. Flash the provided development firmware only to the supported [Nordic nRF52840 DK](https://www.nordicsemi.com/Products/Development-hardware/nRF52840-DK) with no wearables attached. Pair with the serial-console passkey, press the physical fixture button and verify the host’s stop conditions. Confirm the browser never advances the display without the arm checkbox.
4. Only after these gates: supplier optical coupon, adjustable headform, radio coexistence/EMC, measured current and real thermals. A future battery design needs protection, fault tests and lifecycle review. A future human-contact design needs a separate safety and research protocol.

**Why use both.** Aero can keep generating a code draft on the host while a separate display presents one part of that draft. The input and display paths are replaceable. Device failures can stop one input without discarding your request or silently executing it. This architecture, an offline model and explicit controls are the differentiators under test; performance comparisons with commercial glasses or EEG products need equivalent hardware and independent measurement.
