compass

Repository · Roadmap · Decision log · graph.json

The thesis sentence in the repo reads: understand my world better, don’t miss things, get insight from the wealth of knowledge plus me. The last two words carry the design. The model holds breadth. The person holds the problem and the check.

Most smart-glasses products get this backwards. They listen all day, speak uninvited, and spend a week’s worth of attention in an afternoon. compass is built on the opposite premise: it speaks when invoked, it remembers across sessions, it never records continuously, and it never surveils anyone but the wearer’s own working memory.

The harder problem underneath is the one this page exists for. Building the pipeline is cheap now. Knowing whether a given piece of it is right is the work, and it has to be done in domains where the usual geometry instruments don’t apply: speech recognition, vision model answers, a HUD a person has to read in a glance.

Retro mode came up on the laptop in June. Press a key, the last five minutes of the mic buffer go through a local Whisper model, the transcript goes to the coach, a directive lands on the simulated HUD. It worked on the first try, which is the moment this project is about.

The STT model was base.en, chosen because it was fast. Said aloud at one desk, “facade panels” came back as “sign tables.” That was caught by ear, on one machine, in one session. The fix was a bigger model, small.en, and the fix was right. But the check that found the error was a person listening once, and that check does not run again. The next model swap, or the next shop vocabulary, gets no verdict unless somebody happens to be listening.

That is the skip in the figure below. The pipeline went from candidate to “looks right” with no reference in between. The reference is a phrase set: the vocabulary the coach has to hear correctly, recorded once, scored against every model. It is on the instruments list and not yet built.

Every instrument on this project has the same shape, whatever the domain. A candidate (a transcript, a vision answer, a driver) goes against a reference that neither the author nor the model produced. A comparator turns the pair into a residual, a number rather than an impression. A threshold, set before the run, turns the residual into a verdict. The verdict can be no.

The dashed edge is the alternative: candidate straight to verdict on the strength of looking right. It is the fast path and the one that costs understanding, because nothing from it comes back. The solid return edge is why instruments are worth building: an instrument outlives the candidate it was built for, and running it again is free.

The structure in the figure is also published as data at graph.json, with notes the drawing leaves out. The SVG is rendered from that file. Change the file and both faces change.

The hardware constraint is the one that shaped the build order. The Halo was ordered in May and has not arrived. Rather than wait, the three modes were built against a laptop mock (webcam, mic, a window standing in for the HUD) and the roadmap was restructured in June to match: software modes complete on the laptop first, the hardware wire-in last. Two of three modes run today on no glasses at all.

All three feed a layered memory store: rolling buffer, session transcript, session summary, daily digest, project memory. The same question can be asked about five minutes ago or five months ago. The rollup layers past the transcript are scaffolded, not finished.

This is the list the page exists to keep. Each entry says what the instrument can say no to. Entries get added as they are built; none get removed.

Hardware: Brilliant Labs Halo, with the SDK and firmware at docs.brilliant.xyz and github.com/brilliantlabsAR/halo-firmware. What the repo’s own inspection of that firmware established: a single-eye color microOLED, a camera that returns complete JPEG bytes at 640 px to a Lua app on the device, which relays them to the host over Bluetooth. A photograph of the device on this page will be one I took, once it is on the desk.

Where it sits

compass is the glance-paced member of the same family as Directive Engine. Same pattern, capture to reconcile to directive to log, at a different rhythm of attention: seconds to minutes during an install, milliseconds to seconds in a glance. It is also the first instrument here pointed at the person rather than the panel.