Jordan-Hall / Jordan-Hall/browser

[P1][LOCAL-02] Role-specific model packs

Open
#62 2 comments 0 reactions 0 assignees View on GitHub
Dominant language
No language data
Stars
0
Forks
0
PR merge metrics
No merged PRs in 30d

Description

Programme: #1
Epic: #21

## Objective
Package local models by product role and measured hardware support instead of treating one large model as the answer to speech, retrieval, UI and coding.

## Scope
- Model-pack manifest: role, model/runtime IDs, hashes, tokenizer/templates, license, quantization, context limits, hardware requirements and evaluation results.
- Roles: coding/planning, fast extraction/UI planning, embeddings/reranking, speech-related support where applicable, and visual grounding.
- Signed downloads, resumable installation, disk management, rollback and version pinning.
- Disable loaders that execute untrusted repository/custom model code.
- Compatibility matrix per OS/accelerator/runtime.
- Benchmark/regression gate before default-pack promotion.

## Product rules
- Installed weights and active parameters are distinct concepts; publish realistic storage/RAM/VRAM requirements.
- Candidate model names are not permanent product promises.
- Privacy mode cannot change silently because a pack/runtime changes.

## Acceptance criteria
- [ ] Every installed pack exposes hashes, license, hardware/context limits and measured task results.
- [ ] Corrupt/unsigned/untrusted-code model packages are rejected.
- [ ] Rollback restores the last known-good compatible pack.
- [ ] Unsupported hardware receives a useful smaller-pack/degraded path rather than crash/OOM loops.
- [ ] Pack updates run regression suites before becoming defaults.
- [ ] Storage cleanup never deletes an actively required model without explicit state transition.

## Dependencies
- LOCAL-01
- SEC-03

**First phase:** P1
**Maturity target:** P5
**Owner:** local-ai-speech

Contributor guide

No contributing guide indexed for this repository

Research direction

No files, tests, or entry points are named. Start by reading dependencies LOCAL-01 and SEC-03, then break the scope into the manifest, package verification, rollback, hardware compatibility, storage state, and regression-gate requirements. Done means all listed acceptance criteria are met without unsafe packages, crashes on unsupported hardware, or silent privacy changes.

Written by the indexing model from the issue text.

Assessment

Domain
ai, machine-learning, security
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Active
Clarity
Mostly clear
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.