Personal app · Alibaba Qwen3-4B · Multiverse LittleLamb

A personal coach that knows its owner. Better advice, eight times faster.

A fitness-coach app on small open models, governed by CA3 on one person's own records: watch, music and what they told it. It beat the same model handed the app's summaries at every question.

Illustration of the coach app on a phone and a watch; the conversation and the map are invented

Advice preferred69 of 87judged blind; the same model without CA3: 16

Crossing two records19 of 21right; without CA3: 3

Time per answer0.51 swithout CA3: 4.03 s

Tokens per answer1,685without CA3: 3,546

Answers

From its owner's own records.

Questions that only have an answer when two records are crossed, or that need where they train. Handed the app's summaries, the same model rarely gets them.

19 of 21crossings of two records, right; without CA3: 3

5 of 5questions about places, right; without CA3: 0

Advises

Plans that fit the week.

87 pieces of advice, each judged blind by three independent AI readers against the same model given the app's summaries. The readers preferred CA3's in 69, the rival's in 16; two were ties.

69 of 87advice preferred

70, 66, 71of 87, reader by reader

24 of 25morning notices preferred

Costs less

Eight times faster, half the tokens.

Median of 56 shared questions, every model call counted.

0.51 sper answer; without CA3: 4.03 s

1,685tokens read per answer; without CA3: 3,546

Blocks

Another person's days stay out.

Five days of someone else's records were planted among the owner's. CA3 left them out. On the smallest model, LittleLamb, it also kept them out and forgot a sport on request; that model was too small to give useful advice.

5 of 5days of another person left out

How the test was run

The test, step by step.

One person's records, six kinds of question, morning notices and one incident planted.

RecordsWatch, music, notesYears of one person's own records.

Advice87 plans and pieces of adviceJudged blind by three readers, majority of two.69 of 87 preferred

Data21 crossings, 5 on placesPlus 6 questions on what the app does not record.19 of 21 · 5 of 5

Mornings25 mornings of noticesAt most one notice a day.24 of 25 preferred

IncidentAnother person's daysFive days planted in the records.Left out

Compared with the same model and no CA3, given the app's summaries at every question: the last seven days, recent listening, notes and a look-up for the question. One pre-registered run, 1 November 2026.

Measured	Asked	With CA3	Same model, app's summaries
Advice preferred (2 ties)	87	69	16
Crossings, right	21	19	3
Places, right	5	5	0
Morning notices preferred	25	24	1
Tokens read per answer, median	56	1,685	3,546
Seconds per answer, median	56	0.51	4.03

Knowing its owner, the coach gave better advice than the same model reading summaries.

It answered eight times faster and read half the tokens.

Someone else's days never reached it.

Test it on your model.

A pilot inside your own app: you choose which of your users' data it may use and what it must never say, and you measure the answers with your own users.

Contact All twelve use cases

Model and data

Qwen3-4B-Instruct-2507 (Alibaba Cloud) and LittleLamb, 0.3 billion parameters (Multiverse Computing).
One person's own records: watch and fitness data, music history and notes. None of them is shown here.
Advice judged by three independent blind AI readers, majority of two of three.

Sources

Pre-registered round 22 and its reading, 1 November 2026.

Limits

One person and one pre-registered run (round 22, 1 November 2026).
The phone and watch screens are illustrations of the app; the conversation, the song and the map on them are invented.
The models ran on a server; the demonstration app ran on a Mac with the phone as the screen.
With LittleLamb (0.3B) the control worked but the model was too small to give useful advice; the figures above are Qwen3-4B's.

Results from recorded, pre-registered tests. The mechanism is not shown. Model and dataset names are for identification only. Not produced or endorsed by Apple, Spotify, Multiverse Computing or Alibaba Cloud. CA3 is developed by Alberto Acedo, PhD, at Biome Makers Inc.