Academy teacher requires cloud API key — train local teacher adapter

[joelteply]

Problem

Academy teacher pipeline requires --model and --provider (cloud API like DeepSeek). This means:

• Users need an API key to use academy at all
• Violates the zero-API-keys principle
• System isn't self-sufficient

Solution

1. Short term: resolveTeacherLlmConfig should auto-detect best available provider (cloud if available, fail clearly if not)
2. Medium term: Train a continuum:role=teacher LoRA adapter specialized in curriculum design, exam generation, and grading
3. Long term: Bootstrap with cloud teacher → capture training data → train local teacher → publish as adapter → fully self-sustaining

The teacher adapter learns:

• Curriculum design (progressive topics, difficulty scaling)
• Exam generation (scenario-based questions, rubrics)
• Grading (score + detailed feedback + weak area identification)
• Remediation (targeted training data synthesis for weak areas)

Depends on

• Academy pipeline working end-to-end (Academy training pipeline: real models, real benchmarks, real results #372)
• Adapter publish to HuggingFace (Academy training pipeline: real models, real benchmarks, real results #372)

[joelteply]

Superseded by #417 (Qwen3.5 MoE) + #430 (MoE training). The reasoning-distilled model could BE the teacher without cloud API keys.

[joelteply]

Reopening — this is not superseded by #417. Compacting the 35B MoE is the path to a local teacher, but it hasn't been attempted yet. The Academy can't run fully locally until this is solved. Plan: compact Qwen3.5-35B-A3B on 5090, train with Academy data, publish as local teacher model.