Adapter registry — semantic search, auto-forge, publish (npm for intelligence)

[joelteply]

Vision

LoRA adapters as packages. HuggingFace as the registry. Cosine similarity as the search.

System needs Fortran expertise → searches registry by embedding → 
  finds "legacy-systems" at 0.87 cosine → pages it in → good enough.
If not good enough → auto-forges closer match → publishes → cached forever.

Architecture

Base model = CPU

One forged Qwen3.5 base per hardware tier. Always loaded. General purpose.

Adapters = PCIe cards

Hot-swappable LoRA adapters. Few hundred MB each. Domain-specific.
Paged in/out by the genome paging system based on current task.

Registry = npm

• Search by embedding, not keyword. Cosine similarity over capability vectors.
• Capability vectors derived from: training data domain, exam scores, user ratings.
• Close counts. 0.7 match that's cached beats 1.0 match that needs forging.
• Anyone publishes. Researcher, hobbyist, company. Quality proven by exam scores.
• Auto-forge on miss. No adapter? Sentinel generates curriculum, forges one, publishes.
• Coverage densifies organically. Every user's need that triggers a forge fills a gap.

Quality signal = exam scores

Not perplexity. Not downloads. RealClassEval-style exam transcripts.
The 95% adapter gets pulled over the 60% one.

Implementation

Phase 1: Local registry

• adapters.json manifest listing available LoRA adapters with metadata
• Metadata: domain embedding, exam scores, base model, size, author
• genome/adapter/search command: cosine similarity over embeddings
• genome/adapter/load command: download from HF if not cached

Phase 2: Auto-forge

• genome/adapter/forge command: generate curriculum + forge + publish
• Triggered automatically when search returns no match above threshold
• Uses sentinel pipeline for curriculum generation
• Publishes to continuum-ai HF org

Phase 3: Grid distribution

• Nodes advertise available adapters
• Forge jobs distributed to nodes with spare compute
• Adapters replicated across grid for availability

Dependencies

• Python forge pipeline — ship Qwen3.5 forged models now #89 — Python forge pipeline (produces the adapters)
• Sentinel-synthesized training data — academy generates forge curriculum #91 — Sentinel-synthesized training data (generates curriculum for auto-forge)
• model/forge and model/compact as jtag commands — not scripts #83 — model/forge as jtag command
• Genome paging system in continuum (runtime adapter swapping)