Factory → Academy → Sentinel: end-to-end model lifecycle pipeline

[joelteply]

Summary

The full lifecycle of a model in Continuum. Each stage feeds the next. This is the master issue that ties the factory, academy, and sentinel systems together.

The Pipeline

1. FORGE (Factory)
   Input: base model + recipe
   Output: forged model (pruned, trained, published)
   
2. EVALUATE (Factory + HF Leaderboards)
   Input: forged model
   Output: benchmark scores, leaderboard position
   Submit to HF leaderboards automatically after forge
   
3. DEPLOY (Grid)
   Input: forged model + adapters
   Output: running inference on grid nodes
   Genome paging loads/unloads adapters on demand
   
4. LEARN (Academy)
   Input: deployed model + curriculum
   Output: LoRA adapters, skill improvements
   Continuous training while deployed
   
5. RE-FORGE (Factory)
   Input: deployed model + learned adapters + usage data
   Output: improved base model
   Fold learned adapters back into base weights
   
6. ORCHESTRATE (Sentinel)
   Input: pipeline definition
   Output: automated execution of any pipeline stage
   Sentinels can trigger forge, evaluate, deploy, learn cycles

What Exists

• Factory widget with forge controls (Factory — forge production floor with SCADA monitoring and AI foreman #576, PR Factory + Grid: end-to-end forge pipeline with live controls #654)
• model/forge command — starts forge jobs on grid nodes
• model/list-published — live HF API data
• model/forge-status — polls status.json
• Academy session system (Academy: no full training session proven end-to-end #377 — needs e2e proof)
• Sentinel pipeline system (runs multi-step workflows)
• Grid nodes discoverable via Tailscale
• Genome paging (adapter load/unload)

What's Missing

• Auto-submit to HF leaderboards after forge completes
• Re-forge from existing model — load our model card, offer deeper controls
• Adapter → base folding — merge learned LoRA weights back into base
• Sentinel forge recipe — sentinel pipeline that runs the full lifecycle
• Usage-driven re-training — capture interaction data, feed back to academy
• Right-panel leaderboard widget — show HF rankings for our models
• Forge queue UX (Factory widget: wire START FORGE, live status polling, forge queue UX #653) — queue, priority, estimated wait times

Related Issues

• Factory — forge production floor with SCADA monitoring and AI foreman #576 (Factory SCADA)
• Factory job queue — RTOS-style task scheduling across grid nodes #638 (Job queue)
• Factory recipe: context window extension via RoPE rescaling #648-651 (Recipe types: context, vision, audio, composition)
• Academy: no full training session proven end-to-end #377 (Academy e2e)
• Python↔Rust bridge: protobuf schema for forge events (like ts-rs for Rust↔TS) #646 (Python↔Rust bridge)
• Low-latency sensory pipeline — sub-100ms vision + real-time audio for personas #652 (Low-latency sensory pipeline)