TWO_EVIDENCE_LINE_RESULT_SUMMARY.md

# Two Evidence-Line Result Summary

Generated: `2026-06-22T09:56:30+00:00`.

Source matrix: [`docs/data/task_method_20_result_matrix.json`](docs/data/task_method_20_result_matrix.json)

Interpretation rule: Read the 1-episode line as the inspectable task lab. Read the 128-episode line as the selected comparison surface for metadata/raw baselines, Qwen3-Omni v6 LoRA, Cosmos3-Super, and Cosmos3-Nano.

## Read This First

The suite has two public reading lanes. Line 1 is the fully inspectable one-episode task lab. Line 2 is the 128-episode comparison surface for aligned baselines, the Qwen3-Omni series, and the Cosmos3 series. Compare scores within the same lane first.

Score formula: 2 single-episode methods x 20 tasks = 40 records; 7 selected-128 methods x 20 tasks = 140 records; total public matrix = 180/180 scored records.

| Line | What the scores mean | Best use | Read separately from |
| --- | --- | --- | --- |
| 1 sample episode | 40/40 direct scores from Minimal and Neural MLP heads on the same 20 task contracts. | Inspect the raw sample, understand file organization, reproduce the 20 task targets, and compare Minimal vs Neural MLP behavior inside one episode. | The selected-128 comparison rows and broader held-out model behavior. |
| 128 selected episodes | 140/140 selected-128 scores across seven methods: 134 direct scores plus 6 documented compact-proxy scores. | Compare same-split metadata/raw baselines, Qwen3-Omni v6, Cosmos3-Super, and Cosmos3-Nano while keeping the 6 compact-proxy cells visible. | Direct raw-target interpretation for the proxy-marked cells. |

## Public Score Totals

- Lines: 2
- Tasks per method: 20
- Methods: 9
- Scored records: 180/180
- Direct scores: 174
- Compact-proxy scores: 6 documented cells

## Line Ledger And Entry Points

| Line | Methods | Tasks | Scored records | Direct scores | Proxy scores | Primary visuals | Source artifacts |
| --- | --- | --- | --- | --- | --- | --- | --- |
| 1 sample episode | 2 | 20 | 40/40 | 40 | 0 | docs/assets/charts/two_evidence_line_map.svg<br>docs/assets/charts/single_episode_task_model_radar.svg | docs/data/single_episode_task_model_radar.json<br>docs/data/two_evidence_line_result_summary.json<br>results/episode_task_suite/summary_report.json<br>results/episode_task_suite/feature_manifest.json<br>docs/single_episode_explorer.html |
| 128 selected episodes | 7 | 20 | 140/140 | 134 | 6 | docs/assets/charts/two_evidence_line_map.svg<br>docs/assets/charts/episode128_task_model_radar.svg<br>docs/assets/charts/unified_task_model_radar.svg | docs/data/episode128_task_model_radar.json<br>docs/data/two_evidence_line_result_summary.json<br>docs/data/xperience10m_128_episode_feature_index.json<br>docs/data/omni_model_comparison.json<br>docs/data/qwen3_omni_run_lineage.json<br>docs/data/task_method_20_gap_audit.json |

## Method Blocks By Evidence Line

| Line | Method block | Methods | Scored records | Direct scores | Proxy scores | Evidence type | Read as |
| --- | --- | --- | --- | --- | --- | --- | --- |
| 1 sample episode | Task-head baselines | Minimal, Neural MLP | 40/40 | 40 | 0 | Direct target metrics on the public sample windows. | Task construction, local reproducibility, and Minimal-vs-Neural behavior. |
| 128 selected episodes | Aligned baseline heads | 128ep Aligned Simple, 128ep Aligned NN, 128ep Raw Simple, 128ep Raw NN | 80/80 | 74 | 6 | Direct processed-target metrics where available; compact proxies for documented raw-target gaps. | Same-split metadata/raw-feature baseline comparison. |
| 128 selected episodes | Qwen3-Omni series | Qwen3-Omni v6 LoRA | 20/20 | 20 | 0 | Verified selected-128 Qwen3-Omni v6 LoRA plus source-linked task-specific probes. | Trainable Qwen3-Omni diagnostic baseline on the selected-128 surface. |
| 128 selected episodes | Cosmos3 series | Cosmos3-Super Reasoner, Cosmos3-Nano Future Window | 40/40 | 40 | 0 | Verified Cosmos3-Super Reasoner and Cosmos3-Nano Future Window public-safe artifacts. | Cosmos3 reasoner and future-window diagnostics on the selected-128 surface. |

## Method Detail By Line

| Line | Method | Method detail | Scored records | Direct scores | Proxy scores |
| --- | --- | --- | --- | --- | --- |
| 1 sample episode | Minimal | Single-episode simple heads over the public sample split. | 20/20 | 20 | 0 |
| 1 sample episode | Neural MLP | Single-episode compact PyTorch MLP heads on the same 20 task contracts. | 20/20 | 20 | 0 |
| 128 selected episodes | 128ep Aligned Simple | 128-episode aligned simple baselines: JSONL metadata/text tasks plus staged sensor-block tasks where the processed target exists. | 20/20 | 19 | 1 |
| 128 selected episodes | 128ep Aligned NN | 128-episode aligned MLP baselines: JSONL metadata/text tasks plus staged sensor-block tasks where the processed target exists. | 20/20 | 19 | 1 |
| 128 selected episodes | 128ep Raw Simple | 128-episode 4430-dim sensor NPZ simple heads; tasks 15/19 use compact proxies. | 20/20 | 18 | 2 |
| 128 selected episodes | 128ep Raw NN | 128-episode 4430-dim sensor NPZ MLP heads; tasks 15/19 use compact proxies. | 20/20 | 18 | 2 |
| 128 selected episodes | Qwen3-Omni v6 LoRA | Verified held-out Qwen3-Omni v6 LoRA metrics, plus task 16 and any completed private-GPU future/retrieval/sensor-target probes scored from task-specific JSON. | 20/20 | 20 | 0 |
| 128 selected episodes | Cosmos3-Super Reasoner | Verified Cosmos3-Super base-weight Reasoner JSON-task evaluation, plus task 5/8/9/10/11/12/13/14/16/17/18/19/20 probes where public metrics exist. | 20/20 | 20 | 0 |
| 128 selected episodes | Cosmos3-Nano Future Window | Verified Cosmos3-Nano future-window compatibility metrics, plus model-output probes for tasks 2/5/7/8/10/11/12/13/14/15/16/17/18/19 and a derived task-20 boundary timing probe scored from held-out future-window artifacts. | 20/20 | 20 | 0 |

## Related Model Artifacts

| Artifact | Role | Link or path |
| --- | --- | --- |
| Qwen3-Omni v1-v6 run lineage | Explains the LoRA/evaluation version ladder; v6 is the current 20-task matrix row, v5 remains the pinned prior release, and v1-v4 are lineage/ablation evidence. | docs/data/qwen3_omni_run_lineage.json |
| Cosmos3-Super Forward-Dynamics LoRA | Separate fine-tuned adapter artifact for forward-dynamics loss metrics; published with weights/results but not counted as a 20-task matrix method row. | https://huggingface.co/cy0307/ropedia-cosmos3-super-forward-dynamics-lora-128ep |

## Proxy-Scored Cells

| Task | Task label | Method | Metric | Reason |
| --- | --- | --- | --- | --- |
| 15 | Interaction Text Prediction | 128ep Raw Simple | macro_f1 | documented compact proxy completion for this raw128 task axis |
| 15 | Interaction Text Prediction | 128ep Raw NN | macro_f1 | documented compact proxy completion for this raw128 task axis |
| 19 | Camera-View Synchronization Retrieval | 128ep Aligned Simple | mrr | paired camera-view embeddings are absent from the 128 JSONL/feature export; metadata features retrieve the synchronized same-window depth/audio block as a documented compact synchronization proxy |
| 19 | Camera-View Synchronization Retrieval | 128ep Aligned NN | mrr | paired camera-view embeddings are absent from the 128 JSONL/feature export; metadata features retrieve the synchronized same-window depth/audio block as a documented compact synchronization proxy |
| 19 | Camera-View Synchronization Retrieval | 128ep Raw Simple | mrr | documented compact proxy completion for this raw128 task axis |
| 19 | Camera-View Synchronization Retrieval | 128ep Raw NN | mrr | documented compact proxy completion for this raw128 task axis |

## Reading Order

| Step | Reason |
| --- | --- |
| Choose the evidence line | Line 1 answers task-lab and reproducibility questions; line 2 answers selected-128 comparison questions. |
| Open the matching radar | Use the 1-episode radar for Minimal-vs-Neural behavior and the 128-episode radar for metadata/raw baselines, Qwen3-Omni v6, Cosmos3-Super, and Cosmos3-Nano. |
| Inspect the matrix row | Every numeric score is tied to a method, task, metric key, source artifact, and proxy flag. |
| Check proxy cells before interpreting totals | The six compact-proxy cells are numeric but are not direct raw-target measurements. |

## Reader Policy

- 1 sample episode: Use for task construction, raw-file inspection, local reproducibility, and controlled Minimal-vs-Neural baseline behavior.
- 128 selected episodes: Use for held-out comparison, metadata/raw-feature baselines, Qwen3-Omni v6 LoRA, Cosmos3-Super Reasoner, Cosmos3-Nano Future Window, and scale-up decisions.
- Proxy scores: Proxy-scored cells stay numeric only when the source artifact and reason are attached; they should not be read as direct raw-target measurements.