docs: surface the model-card standard in README + CLAUDE.md

CLAUDE.md

@@ -50,6 +50,15 @@ The package is small and self-documenting, so there's no separate `src/` guide.
 - **Gates (all required, blocking PR checks):** `ruff check .`,
   `ruff format --check .`, `mypy` (strict), `pytest --cov` at **100% coverage**.
 
+## Publishing model checkpoints to HuggingFace (any agent session)
+
+Any checkpoint published to the robocurve HF org **must** follow
+[docs/model-cards.md](docs/model-cards.md): fill in
+[docs/model-card-template.md](docs/model-card-template.md) top-to-bottom, then pass the
+gating checklist (any unchecked REQUIRED box blocks publish). Common traps it prevents:
+license inheritance from the base model (never default apache-2.0), missing action/observation
+I/O contracts, and unverifiable eval claims. Reference implementations are linked in the doc.
+
 ## Adding a benchmark to the catalog
 
 Append a `Benchmark(...)` entry to `src/worldevals/catalog.py` with its name,

README.md

@@ -85,6 +85,14 @@ validates every entry (unique name, well-formed repo URL, ≥1 task key). See
 [KitchenBench](https://github.com/robocurve/kitchenbench) as the reference
 implementation.
 
+## Publishing model checkpoints
+
+Benchmarks compare *policies* — and a benchmark number is only interpretable when the
+checkpoint behind it is documented. Every model checkpoint published to the
+[robocurve HuggingFace org](https://huggingface.co/robocurve) must follow the
+**[model-card standard](docs/model-cards.md)** (start from the
+[template](docs/model-card-template.md); the standard includes a gating checklist).
+
 ## Development
 
 > **Dependency changes:** after editing dependencies in `pyproject.toml`, run

docs/model-card-template.md

@@ -32,6 +32,11 @@ inference: false
 
 # <Model name — robot + method in plain words>
 
+> **Evaluate this checkpoint** with [Inspect Robots](https://github.com/robocurve/inspect-robots) —
+> the Inspect-AI-style eval framework for physical AI — using the
+> [<embodiment> adapters](https://github.com/robocurve/inspect-robots-<embodiment>); benchmarks
+> are catalogued in [WorldEvals](https://github.com/robocurve/worldevals).
+
 <≤3 sentences: [base model](link) + method (LoRA/full FT; what's frozen) + robot +
 [dataset](link). Adapter/merge status and where raw adapters live.>
 

docs/model-cards.md

@@ -93,6 +93,10 @@ model-index:
 
 At most three sentences: base model + method (LoRA/full FT; what's frozen) + robot/embodiment
 + dataset. State whether adapters are merged and where raw adapters live if published.
+**Immediately after the H1**, add an eval-path blockquote linking
+[Inspect Robots](https://github.com/robocurve/inspect-robots), the embodiment's adapter repo
+(e.g. inspect-robots-so101 / inspect-robots-yam), and
+[WorldEvals](https://github.com/robocurve/worldevals) — see the template.
 Link every **named artifact**: base model, each dataset, training-code repo, method libraries,
 papers. (Ordinary nouns like "camera" need no link.)
 
@@ -201,6 +205,7 @@ Frontmatter
 
 Body
 - [ ] Header ≤3 sentences; every named artifact linked; adapter/merge status stated
+- [ ] Eval-path blockquote right after the H1 (Inspect Robots + embodiment adapters + WorldEvals)
 - [ ] Intended use + out-of-scope + validation status (with e-stop/supervision language)
 - [ ] Training table: data counts + filter/split rule (level, %, seed)
 - [ ] Training table: embodiment config + state/action key layout