| { |
| "recipe_id": "robometer-4b", |
| "generated_at": "2026-07-07T18:42:12+00:00", |
| "bundle": "build/robometer-4b/robometer-4b.aimodel", |
| "conversion_manifest": null, |
| "gate_a": { |
| "status": "passed", |
| "checks": [ |
| { |
| "name": "bundle_exists", |
| "status": "passed", |
| "detail": "build/robometer-4b/robometer-4b.aimodel" |
| }, |
| { |
| "name": "bundle_files_present", |
| "status": "passed", |
| "detail": "3 expected file(s) present" |
| }, |
| { |
| "name": "metadata_json_parses", |
| "status": "passed", |
| "detail": "metadata.json parses (3 top-level keys)" |
| }, |
| { |
| "name": "metadata_matches_recipe", |
| "status": "passed", |
| "detail": "1 key(s) match" |
| } |
| ] |
| }, |
| "gate_b": { |
| "metric": "graph_output_cosine", |
| "threshold": 0.999, |
| "tolerance": 0.0005, |
| "value": 0.999999999996286, |
| "status": "passed", |
| "measurement_source": "graph-output-parity-measured.json", |
| "min_cosine": 0.999999999996286, |
| "median_cosine": 0.9999999999982424, |
| "mean_cosine": 0.9999999999979249, |
| "cosine_ci95": [ |
| 0.999999999996286, |
| 0.9999999999980718 |
| ], |
| "min_progress_cosine": 0.999999999998956, |
| "min_success_cosine": 0.999999999996286, |
| "per_obs_cosine": [ |
| 0.9999999999981611, |
| 0.9999999999984405, |
| 0.9999999999980718, |
| 0.9999999999986486, |
| 0.9999999999986552, |
| 0.999999999996286, |
| 0.9999999999983237, |
| 0.9999999999968127 |
| ], |
| "n_obs": 8, |
| "hidden_dim": 2560, |
| "progress_bins": 10, |
| "frames": 8, |
| "reference_dtype": "float32", |
| "runner": "coreai-fabric-parity-runner/0.1.0", |
| "environment": { |
| "platform": "darwin-arm64", |
| "accelerator": "apple_silicon", |
| "runtime_version": "1.0.0b2", |
| "coreai_torch": "0.4.1", |
| "torch": "2.9.0", |
| "transformers": "4.57.6" |
| }, |
| "reference": "Torch Robometer reward heads (progress + success) vs the Core AI .aimodel over identical seeded per-frame prog-token hidden states; non-autoregressive single forward (the Qwen3-VL-4B backbone is host-owned, not in the asset). The metric is the worst per-obs cosine across both output heads." |
| }, |
| "overall": "passed" |
| } |
|
|
