v0.2.0 - Informed and capability-faithful embodied evaluation

Highlights

• Adds source-backed model and benchmark capability discovery before execution.
• Adds informed mode, output, resource, and claim selection through run-contract v2.
• Enforces end-to-end capability activation across checkpoint, wrapper, adapter, writer, validator, and visualizer.
• Adds truthful T+1 replay, predicted-vs-observed separation, and human-readable behavior analysis.
• Reframes G8 as an approved goal-driven study rather than an automatic batch expansion.
• Hardens mutable API boundaries, result-blind pilot selection, repair precedence, manifests, resumability, failure taxonomy, and delivery safety.

Compatibility

• v0.2.0 creates embodied-run-contract/v2 contracts by default.
• Existing v1 contracts receive explicit migration guidance.
• A v1 contract never gains implicit capability, media, study, training, or permission authority.
• Full training remains outside the default evaluation workflow and requires a separate handoff and approval.

Validation

• 35 unit and regression tests passed.
• GitHub Actions passed on Windows and Linux with Python 3.10 and 3.12.
• Repository and Agent Skill validators passed.
• Five isolated forward-test scenarios passed: action-only, required future prediction, resource-constrained study design, training-boundary handoff, and append-only repair lifecycle.
• CI evidence: https://github.com/Yinzhanqing/embodied-eval-automation/actions/runs/30804419172

Installation

Install the tagged skill directory with Codex skill-installer:

https://github.com/Yinzhanqing/embodied-eval-automation/tree/v0.2.0/skills/embodied-eval-automation

The tagged URL is recommended for reproducible use. The main branch may contain unreleased changes.

Scope

This release contains workflow instructions, templates, and standard-library validation tools. It does not include model weights, benchmark code, datasets, generated episodes, videos, or private infrastructure details.

Full Changelog: v0.1.0...v0.2.0

v0.1.0 - Initial public release

Highlights

• Provider-neutral G0-G9 workflow for producing validated embodied-AI episode datasets.
• Permission-aware onboarding for SSH/cloud servers, GitHub, Hugging Face, local workspaces, and storage boundaries.
• Existing benchmark/model/checkpoint inventory and reuse checks before installation.
• Isolated model and benchmark environments connected through a loopback adapter.
• Three-representation episode workflow: native official output, current unified schema, and candidate schema.
• Paired evaluation contracts with pair_key, T+1 observations, policy_queries, and raw/executed action separation.
• Staged acceptance gates from single request through full runs, with monitoring, resumability, transfer verification, and retention controls.
• Standard-library validators, unit tests, examples, templates, and GitHub Actions CI.

Scope

This release packages an automation skill and workflow contract. Model- and benchmark-specific adapters remain explicit integration work and must be validated against pinned upstream revisions before production runs.