Templates

The curated garden. In an era of slop, the value is a small set of pages you can trust — each one a frontier technique with working code, real data, and a reference run whose numbers you can reproduce.

What a template is

A template is a public koyu project: project files (code + playbook), runs (including at least one completed reference run with checkpoints, metrics, and eval results), and linked public manifests (the data). Templates are fully public — reading and replicating one requires no account, no token, nothing. Auth appears only if you push results to your own cloud.

Trust is earned mechanically, not editorially. A template enters the garden only if: its golden tests are green, its reference run is attached and reproducible, its INTERFACE block is present, and every manifest it links is public. Agents can lint all four.

The three verbs

Every template declares which it supports:

verb	you bring	you get
TRAIN	a conforming dataset (manifest)	a checkpoint + metrics
EVAL	a policy implementing the contract	success rates + eval episodes
INFER	nothing (checkpoint ships)	actions over the wire

The README anatomy

Every template README has the same seven sections, in order:

1. WHAT — three sentences and the reference-results table.
2. VERBS — which of train/eval/infer this template supports.
3. INTERFACE — the exact data contract: feature keys with dtype, shape, units and frames (joint angles vs EEF, radians, gripper range), rates, chunking semantics, and the stats rule. Plus the conformance test: your data conforms iff the vendored dataloader golden test passes on it. Robot learning's gnarliest thorn — action spaces, tokenizers, image transforms — is handled by making it explicit and executable, never implicit.
4. COST — measured throughput on the reference GPU and a scaling formula, so the price of a fine-tune is known before a GPU is rented.
5. EXECUTE — per-verb staged commands. Every step ends in a named artifact; verify it before proceeding. Known failure modes are documented inline at the step where they occur.
6. ACCEPTANCE — expected artifacts and numbers ± tolerance, and the optional final step: push your run and it appears in the template's public clone lineage. Replication ends with a receipt.
7. ANNEX — links to format, policy-contract, agents.md, and the repo AGENTS.md files.

Replication

The template page's Replicate button copies a short, frozen prompt: fetch the README, follow EXECUTE for the goal stated at the bottom, verify each artifact, report against ACCEPTANCE. The prompt contains no credentials and no platform knowledge — everything lives in the versioned README.

The README opens with a triage table routing by situation: replicate as-is (pure recipe, no questions) · train on your data (recipe + conformance check) · adapt to your own runtime/ROS stack (fetch agents.md, interview the user, adapt the contract files) · integrate with koyu-workspace (the user clones; the agent works on the resulting directory as plain files). One-verb simplicity for the common case; explicit permission to leave the recipe for the tails.

Templates own their runtime

A template ships its own services.yaml: simulator, policy service, and the data recorder run as supervised processes on the same cells. Eval rollouts are therefore recorded by the same recorder that captures real-robot data — eval episodes carry collection_mode: eval, source_checkpoint, and a reward verdict, land in an eval manifest whose success_rate rolls up from them, and are indistinguishable in provenance quality from teleop data.