Expand the model catalog

Require authorization to modify the repository and an explicit candidate/model budget. Use discover-papers to search arXiv and Hugging Face Papers, normalize identities, and deduplicate against uv run tsf model list --json. No candidate clearing the gate is a successful no-change outcome.

For each retained candidate, verify the primary paper, forecasting task, required inputs, official source, revision, and license before selecting at most the authorized number. Extract paper structure, inspect pinned official code when it exists, and integrate with add-model. Use implement-model for an ordinary paper architecture or integrate-foundation-model for a released pretrained foundation runtime. Never use search relevance or a shape-only test as implementation evidence.

Build a reuse-existing / extract-new / model-local component decision map before writing code. Reuse an existing component whenever its semantics and runtime contract match; keep material variants model-local. Preserve the flat model layout, useful paper comments, and every known difference. Each added model must pass focused tests, unified verification, strict runtime, model audit, component audit, and repository audit.

Do not publish, push, open issues, or dispatch more tasks unless separately authorized. Stop when the model budget is exhausted or paper, license, data, or runtime ambiguity would require inventing a claim.