Research · Evidence card

Neural Imprint — Evidence

Subject
On-device learning with Neural Imprint, and the runtime it depends on. Six evaluations on devices we own.
Models
Qwen3.5-4B · Qwen3.5-9B
Devices
iPhone Air · iPhone 17e · iPhone 17 Pro · iPad Air M3
Data
Simulated personal records: 3,283 entries across 8 areas of life. Real records: 112 entries from one person.
Published
Principle, limits, evaluation results.
Not published
Internal structure of the learning state, activation rules, thresholds, recipes, raw conversations.

1Summary

#EvaluationDeviceModelDataResult
E1Answer from learned state · exact-figure lookupiPhone 17e · iPhone 17 Pro9B 4-bitsimulated, 3,28316/16 · 16/16
E2Learned-state consistency checkMac, Apple Silicon (host)4B 6-bit · 9B 4-bitsimulated, 8 areas16/16
E3Profile from real recordsiPhone Air4B 6-bitreal, 112, one person48.5 s · 5/5
E4200-turn sessioniPhone Air · iPhone 17 Pro9B 4-bit200/200 · 200/200
E5LimitsiPhone Air · iPad Air M39B 4-bit100/100 @ 2,048 · 262,144 ctx
E6Compression · exact-fact recalliPad Air M39B 4-bit4/4 → 2/4 · 5/5

2Learning

E1Answer from learned state · exact-figure lookup

DevicesiPhone 17e (A19) · iPhone 17 Pro (A19 Pro)
ModelQwen3.5-9B, 4-bit
DataSimulated personal records, 12 months, 3,283 entries
Areasfinance · health · reading · journal · travel · cooking · music · work
Profile question, answered from learned state16/16 · 0 lookups · 10.2–17.5 s
Exact-figure question, looked up16/16 correct tool, result returned
Control: same model, no learned state0/16 within time limit

Simulated data, not real users. · No third setup with lookup tools and no learned state; lookup behaviour is not attributed to the learned state alone.

E2Learned-state consistency check

MachineMac, Apple Silicon · the check runs where the learned state is built, before it is loaded onto a phone
ModelsQwen3.5-4B 6-bit · Qwen3.5-9B 4-bit
DataSame 8 simulated areas as E1
ChecksTraits are distinct from one another · each trait carries signal
Result16/16 pass (8 areas × 2 model sizes)

Measures internal consistency of a learned state, not answer quality. · Pass lines fixed; not published.

E3Profile from real records

DeviceiPhone Air
ModelQwen3.5-4B, 6-bit
Data112 real spending records, one person
Extraction time48.5 s
Profile components selected5
Reference-behavior check5/5 pass

One person; no statement about other users. · Record contents not published.

3Runtime

E4200-turn session

iPhone AiriPhone 17 Pro
Device class12 GB · A19 Pro12 GB · A19 Pro
ModelQwen3.5-9B 4-bitQwen3.5-9B 4-bit
Workload200 × 1,024 tokens200 × 1,024 tokens
Turns completed200/200200/200
Tokens generated204,800204,800
Duration7.06 h5.05 h
Tokens / s, average8.1711.45
Tokens / s, minimum7.2310.31
Peak memory5,501.6 MB5,505.7 MB
Memory growth, first to last turn+166.1 MB+165.9 MB
Failed turns · out-of-memory events0 · 00 · 0

Each turn forced to exactly 1,024 tokens; output quality not scored.

E5Limits

Working point1,024 tokens / turn · verified in E4
Higher load, 100 turns, iPhone Air2,048 tokens / turn · 100/100 · 204,800 tokens · peak 5,299 MB · 0 OOM
Context ceiling262,144 tokens · reached 261,425 on iPad Air M3 before overflow

Beyond the ceiling a session must be compressed or restarted; flat memory does not extend it.

E6Compression · exact-fact recall

Device · modeliPad Air M3 · Qwen3.5-9B 4-bit
Turns completed72/72
Facts planted at the start9: five names and places (a person, a city, three code names) · four exact figures (a token count, a device count, an item count, a date)
Compressions, at turn32 · 56
Asked back at turn 28, before any compression5/5 names and places · 4/4 exact figures
Asked back at turn 40, after one compression5/5 · 3/4 (the date lost)
Asked back at turn 72, after two compressions5/5 · 2/4 (the date lost; the item count answered as “about 12”)
Peak memory5,345.8 MB
Tokens / s, average8.83
Comparison run, smaller memory window — asked back at turn 721/5 · 0/4

After compression the model keeps names and places and loses exact figures and dates. Exact facts need to be looked up from a local tool or store rather than relied on from context memory. · One device.

4In validation

ItemDone so farCompletes when
Answer quality gain from learningFour-way comparison in progress; pass line locked before the run.Paired real-device result over the line.
Multiple usersReal data from one person (E3).Same evaluation on real data from more users.
Non-Apple chipsAll six evaluations on consumer Apple hardware.Same evaluations on other chips.
Model improvement from user correctionsCorrection path built and audited; disabled by default.Enabled after the comparison above passes, then evaluated.

5Method

  1. 01Held-out test set the model has not seen.
  2. 02Leak check between training data and test set.
  3. 03Paired runs on two real devices.
  4. 04Blind scoring.
  5. 05Compared against a pass line set before the run.

6Provenance

Each evaluation is stored internally as an evidence package: configuration snapshot, raw logs, per-file SHA-256 checksums, device class, model, software build.

Not published: raw conversations, receipts, real records.

Back to research