RLE-Bench/robocasa-openfridge-astra-low-eval

OpenFridge: Kinex and Codex evaluation

Watch the evaluation website · Dataset repository · Browse all files · Website source

Ten recorded OpenFridge episodes, evaluated on September 18, 2026 with GPT-6 Astra / low, standard service tier, fast mode disabled. Five fresh conversations per harness; seeds 0–4; existing cross-episode tool/skill/memo sharing retained.

Harness	Native successes	Outcomes E1–E5	Harness execution
Kinex	1/5	fail, success, fail, fail, fail	Four episodes hit the default 100-model-turn cap
Codex	3/5	fail, success, fail, success, success	No execution exceptions; two voluntary incomplete stops

These are native outcomes from an integration evaluation. Kinex has an additional model-loop cap, so the results are not an isolated model-capability comparison. Artifact acceptance passed for all ten episodes; campaign execution validity did not pass because of Kinex exceptions. No claim of reflection improving success rate is made.

What is included

• Full available native-format sessions, readable transcripts, model request/configuration records and tool call histories.
• Every episode's saved tool/memo versions, workspace archive, file hashes, changes and text diffs. No independent skill was generated in this run.
• Evidence-based per-episode failure/success analysis, with read-only native door measurements and reproducible analysis code.
• Ten edited dual-view 1536×896 website videos, rendered natively at 768×768 per view. Clips run at 4× simulation speed, with captions and initial/final holds. Per user request, original camera streams and unedited HD videos are not published, including inside evidence archives; the private source evaluation retains them.
• Control commands, timelines, observations, native verdicts, model XML, joint-layout metadata and precomputed door diagnostics. Raw states.f64, large model.mjb scene binaries and duplicate owner evidence archives are excluded by user request; original local evaluations retain them.
• Protocol, execution exceptions, setup incidents, validation records, source commit identities and documented log-recovery provenance.

Read the results

Data format and analysis guide · CSV episode index · JSON episode index

Full analysis · Structured results · Native door audit · Effective protocol · Artifact acceptance

Files are organized as episodes/kinex/episode_01/ through episode_05/, and likewise for codex. Stable original run IDs remain in the metadata. Each episode directory includes analysis.md, session.txt, agent/, artifacts/workspace/, verifier/, edited.mp4, replay/, and owner/.

Interpretation and provenance

The common native budget is 5000 control intervals per episode; simulation is synchronous at 20 Hz. Each accepted device write advances one interval. Reads and model reasoning do not advance physics. The safety timeout is 28800 seconds. Seeds match between harnesses and the five paired native initial states were verified identical.

Kinex's builtin reflect skill was not invoked. The protocol asks agents to update persistent resources within their task conversations; it does not guarantee a separate post-episode reflection stage. Saved resources were actually reused, but some notes overgeneralized between fixture types.

Native success uses RoboCasa's refrigerator predicate: every target fridge door must reach normalized opening ≥0.90 (81 degrees for these joints). Freezer opening alone is insufficient. Agent prose, process exit status and visible motion do not replace this predicate. Owner diagnostics are post-evaluation evidence and were not exposed to the agents.

The original Codex run suffered Harbor overredaction of the non-secret auth selector value 1. Published session texts use the documented recovery copy. Recovery details retain hashes and the ambiguity of originally literal [REDACTED] strings; pre-recovery text is also retained. Binary native evidence was unaffected. Do not describe these texts as untouched original logs.

Publication only replaces private host path prefixes in text and archive members, neutralizes archive ownership metadata and omits generated GPU shader caches and the excluded unedited videos. Session events are not abridged. Publication transformations document every changed source file/member. Frozen experiment manifests retain their historical hashes; use the publication manifest to validate the public copy. All ten retained edited MP4s and observation images are byte-identical to the source review.

The historical replay restored recorded native states and verified them; it was not deterministic command resimulation. This compact public edition preserves the precomputed native diagnostics and verdicts, but excludes raw physical state files and large scene binaries. Independently recomputing the state audit or rerendering the scene requires the retained local source evaluation. Model XML, observations and rendered videos originate from third-party simulator assets: their upstream terms remain applicable. No blanket license is asserted for those assets; see third-party notices.

Local use

hf download RLE-Bench/robocasa-openfridge-astra-low-eval --repo-type dataset --local-dir openfridge-eval
cd openfridge-eval
python3 tools/verify_publication.py
python3 -m http.server 8000

Open http://localhost:8000/ for local browsing. The published Static Space serves a compact website mirror without a GPU service and links to this dataset for complete evidence. The dataset retains all records and supports local browsing. The included native-door audit reports all-state measurements computed before publication. The audit scripts are provided for transparency, but require the intentionally omitted source states (and, for layout export, scene binaries); the public download alone cannot rerun that audit.