📦 RFC: should apps be able to bring their own models?

[rufolangus]

Today JibarOS ships platform-default models per capability, OEMs override via /vendor/etc/oir/oir_config.xml, devs sideload via /data/local/oir/. Apps cannot supply their own model.

Should they be able to?

Arguments for — specialty domains (medical imaging, music transcription, niche languages) where the platform default is wrong and waiting for an OEM bake-in isn't realistic. Apps already ship llama.cpp/ONNX themselves to work around this; OIR's runtime is more efficient than each app re-implementing.

Arguments against — the whole point of OS-level capabilities is sharing. Every app shipping a 500MB–2GB model multiplies storage + RAM pressure. New attack surface (malicious GGUF loaded into a shared daemon). Quality variance hurts the OIR brand.

A middle path worth debating:

• BYOM allowed only via a namespaced variant: app:com.myapp.medical_qa, never as override of the platform text.complete default.
• App's UID is charged for the resident memory; lazy-load + aggressive eviction when other apps need budget.
• Models live at /data/app/<pkg>/oir-models/; app declares SHA in its manifest, OS verifies on load.
• OEMs can disable BYOM globally via a knob (apps.byom.enabled = false).

Open questions:

1. Is BYOM the right pattern, a leaky abstraction, or fundamentally at odds with the share-the-runtime model?
2. If yes — should apps register models per-capability-shape, or as opaque "use this model for inference X"?
3. Should the OS gate BYOM on a privileged permission?