The numbers, warts and all.

Every measured claim we make, in one place — the flattering and the unflattering together. Two rules govern this page: population results are never merged with single-user results, and cold start is never merged with personalized. The short story is on the about page; this is the long one.

The technology, measured

The short form. The unabridged record — every number with every caveat, as we state it to ourselves — is public at /llms.txt, and the raw data is available under NDA.

• The recognizer. A compact quantized convolutional neural network over accelerometer and gyroscope, running on a Nordic nRF54LM20B with its on-chip neural accelerator. 25–28 ms per letter; zero accuracy loss from float to INT8. We don’t publish the architecture or the model size.
• Letters, cold (population). 67.5% per-character top-1 for a first-time stranger, across a 300+-subject paid program. Attune adds 8–10 points at population level and halves false rejections; on the physical device it has lifted a known hand into the 90s (single user). A practiced single writer through the full pipeline lands 97%+ of words exactly.
• Words, cold (population). 85% of a first-time stranger’s words come out exactly right — 317 held-out paid writers, leave-one-subject-out, phrases assembled from their real recorded letters. The median writer is at 95.5%; one in ten is under 66%. The spread is the finding: most people need no personalization at all.
• The re-test (August 2026). 28 people wrote twice, on different days. The twelve the model found hard went 82% → 93.5% of words exact after Attune — all twelve improved. The sixteen it already read well moved 95.6% → 96.5% — real but small, and four came out slightly worse. Attune is a fix for the tail, not a general upgrade — and after it, nobody was left above 20% word error. One narrow cohort (US men, 50–80); the full accounting, corrections included, is in the unabridged record.
• Speed. About 15 words a minute in letter mode in regular use — one letter at a time, on the wand itself, where the letter arrives as a keystroke while you write. The 5 standardized words per minute we published until 31 August is the same mode measured on the study rig: a prompt screen between every letter, a network round-trip on each one, and a pause to read the phone. About 70% of that clock is not writing — press, release, wait, read, reposition. Both figures are ours and they measure different things. Treat the 5 as the floor: it is what the conservative path measures, it matches the published Georgia Tech air-writing benchmark of 5.43, and the TV D-pad on our own figure is 5–9. On that same rig the founder’s draw-only bound is ~11 and the paid cohort’s median at real rhythm is 3.9; one early fleet user writes close to 12, letter by letter. The 15 is the founder’s own rate on his own hardware, self-timed — nobody but him has yet been timed on the device, and 15 words a minute is about 12 on the standardized ÷5 convention the other figures here use.
• Staccato, in the field. A whole word drawn in one continuous hold: early real-user sessions pool to ≈8.7 standardized words per minute on the shared model, no personalization (the founder’s median ≈12). Small samples, draw-time estimates — and we judge the mode on correct words per minute, not raw speed. That ≈12 is not comparable with the 15 above: it counts drawing only, and at the drawing tier staccato is in fact slightly slower than one letter at a time. What it buys is writing a whole word in one flow, and at 15 words a minute in that lane our recognizer cannot yet read people reliably — which is why staccato is a claim about flow and letter mode is the one carrying a speed number.
• Hardware status, stated plainly. The recognizer runs on production silicon in hand-built development hardware; prototype wands are in the field with outside users, updated over the air. The schematic is complete and electrically clean, the 27-line bill of materials is closed, the radio is a pre-certified module, the firmware has bench-passed the factory flow end to end — over-the-air update included — and manufacturing is quoted, not hypothetical. The first production-design boards (Rev A) are fabricated; assembly and bench bring-up follow in September and October 2026, and the production inertial sensor is validated on them. Working hardware at CES in January 2027 is the plan; nothing is on sale — the plan for 2027 is a crowdfunded first run of the device alongside licensing the recognizer to device makers.
• How we measure. Paid strangers only, scored leave-one-subject-out: the model reading a writer has never seen that writer’s handwriting. Nobody coached, nobody hand-picked; the results that went against us are published beside the ones that did not. What we lack is a third party re-running it — the university collaboration is the route, and the standing invitation is the better test: bring your own words.

Research & data

• The corpus: 490+ paid, consented participants across 28 studies. The largest set is the Atlas — 349 adults aged 18–78, 52,191 letters with age, sex and handedness. The free games added ~2,000 letters from nine countries; adding them changed accuracy by nothing we could measure, twice.
• Independent evaluation: US university researchers, IRB-approved — a nine-participant pilot on our hardware in spring 2026, a larger study on the latest hardware running fall 2026, peer-reviewed publications in preparation, data open-sourcing planned.
• The Air-Writing Atlas (August 2026): the first public analysis of the corpus. Men and women write the same in the air; older hands don’t shake more, they slow down after 50. Pre-registered, replicated on both phone platforms. The write-up · the release.

Diligence, welcomed

Serious parties can go past this page. Available on request: hands-on device demos — bring your own words, test the runtime offline, put new or left-handed writers on it — and, under NDA, raw anonymized study data, protocols, per-user accuracy distributions and confusion matrices. It is all there to be picked over. Write to [email protected].