Roadmap entry — Phase 2. Surfaced 2026-05-01 during PC-056 architectural conversation.

Why

ATLAS is currently model-locked because the Geometric Lens has model-specific artifacts (metric tensor + embeddings). To make ATLAS truly model-agnostic, users need the ability to train Lens artifacts for new models locally.

Scope

`atlas lens build <model_path> [--task-set ...] [--candidates N]` — long-running pipeline that produces Lens artifacts for a new model:

• Run benchmark task set against the model (sandbox-labeled pass/fail candidates)
• Extract C(x) embeddings from each candidate
• Train metric tensor (G(x)) on the labeled embedding dataset
• Write artifacts to local `models//lens/{cost_field.pt, metric_tensor.pt}`
• Update local registry with `lens_status: supported`

UX expectations

This is not interactive — expect hours-to-days of GPU compute per model depending on candidate count. The CLI should:

• Resume from interruption (checkpoint candidates + intermediate state)
• Show clear progress (candidate N/M, training epoch)
• Estimate time-to-completion based on first-N timing
• Default candidate count to a sensible minimum that still produces a usable Lens

Dependencies

• PC-057 must land first (compat probe — don't start a multi-hour build against an incompatible model)
• Reuses existing benchmark infrastructure (`benchmark/v3_runner.py`, sandbox)
• Reuses existing Lens training code (`geometric-lens/training/`)

Out of scope

Publishing artifacts to a registry (that's PC-059 / PC-060). Distributed/cloud training.