Technical reference · for existing users
What model validation proves: revision, asset SHA, loading and warm-up differ
Pinned asset identity is not accuracy; installation, loading and warming are separate gates
Qwen3ModelStore.isInstalled checks a fixed revision marker and listed asset sizes/SHA-256, not directory existence alone. New loading still has an integrity gate, while CoreML load/warm-up can fail separately. validateInstalledModel can reuse a loaded/warmed instance rather than reread every file each call. File evidence does not establish every word’s accuracy or scenario readiness.
Different from trial, voiceprint and material digests
Trial checks the whole chain; voiceprint is another convenience filter; digests identify bytes. Here pinned ASR assets, installation, memory loading and warming have separate evidence scopes. No download, private voice access, removal or inference ran.
| Check | Verified scope | Manual review | Not established |
|---|---|---|---|
| Pinned manifest | Repository revisions/per-asset bytes/SHA | Actual fixed-source manifest | Any same-name weights compatible |
| isInstalled | Marker and listed sizes/digests | Cancellation/missing/mismatch causes | Directory presence suffices |
| Manifest scope | Listed assets checked | Reviewed files versus other material | Every directory file authenticated |
| Publish tree | Verified staging then replace/move | Actual install/error | Progress means loaded |
| Load | loadIfNeeded checks memory first | New load integrity/generation | Every call rescans all files |
| Generation | installGeneration guards suspended old load | Old preparation/current model | Identity signature |
| Warm-up | validate can reuse warmed instance | Actual results/errors | Real speech accuracy accepted |
| Retention evidence | Version/manifest/actual checks | Device/permissions/audio/content separately | Security certification/semantic truth/restore SLA |
Synthetic validation note
Known: assets match one pinned manifest. Unknown: this device’s CoreML load, warming and speech quality. Report file integrity only, not recognition performance. Reusing a loaded instance cannot be described as a just-completed full disk recheck.
Do not falsify revision markers/manifests or download/remove real models for demonstration. Digests match expected bytes, not release trust, usage permission, runtime compatibility or speech accuracy. None was benchmarked here.
FAQ
Does isInstalled check only folder existence?
No. Fixed revision, listed sizes and SHA are checked.
Do matching hashes guarantee every sentence?
No. Integrity differs from runtime, device/audio and meaning.
Does validate reread all files every time?
Not necessarily. Loaded/warmed instances can be reused; new loading still needs integrity.
Can loading fail after installation?
Yes. CoreML load/warm-up differ from download status.
Does false prove malicious modification?
No. Cancellation, missing/read/size/hash/version conditions need separate review.
Was a model run/downloaded?
No. Source/evidence review only, without performance/device acceptance.
Reading and source
Whole-chain trial · Voiceprint distinction · Material/version mapping · Preview/final · Start · Privacy
Manifest/install/verification · Load/validation · Existing preparation
