Add private PAWBench benchmark input preview

README.md

@@ -0,0 +1,133 @@
+---
+pretty_name: PAWBench
+size_categories:
+  - n<1K
+tags:
+  - benchmark
+  - video-generation
+  - physical-reasoning
+---
+
+# PAWBench
+
+### A benchmark for distributional physical realism in generated videos
+
+> [!IMPORTANT]
+> **Private preview.** This repository contains the current materialized
+> PAWBench input package for review and integration. It is not yet a public
+> benchmark release or leaderboard submission endpoint.
+
+PAWBench asks whether a video model reproduces the *range and frequency* of
+outcomes that a physical scene can produce. It evaluates repeated rollouts from
+the same source image and action, rather than treating one plausible-looking
+video as sufficient evidence of physical realism.
+
+## At a glance
+
+| Component | Contents |
+| --- | --- |
+| Scenes | 50 physical scenes |
+| Calibration track | 25 scenes with reference outcome distributions |
+| Coverage track | 25 scenes with supported outcome labels |
+| Source images | 50 first-frame images |
+| Prompt material | One base prompt per scene and the `gt_guided` prompt bank |
+| Evaluator | [PAWBench code](https://github.com/Andrew0613/PAWBench) |
+
+## Repository layout
+
+```text
+.
+├── manifest.json
+├── scenes.jsonl
+├── source_images/
+│   └── <scene source image>.png|jpg
+└── prompts/
+    └── gt_guided/
+        ├── prompt_bank.jsonl
+        ├── README.md
+        └── review.md
+```
+
+`manifest.json` declares the package schema and points consumers to the scene
+table. `scenes.jsonl` is the benchmark contract: each line is one scene and
+paths are relative to this repository root.
+
+## Scene format
+
+Every scene row contains the model input needed to generate a rollout and the
+outcome contract needed to evaluate it.
+
+| Field | Meaning |
+| --- | --- |
+| `scene_id` | Stable scene identifier |
+| `split` | `calibration` or `coverage` |
+| `source_image_path` | Local first-frame image path |
+| `action` | Physical action to be performed |
+| `base_prompt` | Baseline image-to-video prompt |
+| `outcome_labels` | Canonical labels for observable outcomes |
+| `reference_distribution` | Calibration-only target distribution; `null` for Coverage |
+
+Illustrative Calibration record:
+
+```json
+{
+  "scene_id": "A-01",
+  "split": "calibration",
+  "source_image_path": "source_images/A01.png",
+  "action": "Flick the coin once.",
+  "outcome_labels": ["heads", "tails"],
+  "reference_distribution": {"heads": 0.5, "tails": 0.5}
+}
+```
+
+## Using the data
+
+Clone or download this dataset locally, generate repeated videos for each
+scene, then pass local video paths to the PAWBench evaluator:
+
+```python
+from pathlib import Path
+
+from pawbench import evaluate
+
+result = evaluate(
+    benchmark_path=Path("/path/to/PAWBench"),
+    videos=[
+        {
+            "sample_id": "my-model::A-01::r000",
+            "scene_id": "A-01",
+            "repeat_index": 0,
+            "video_path": "/results/A-01/r000.mp4",
+        }
+    ],
+    model_or_lane="my-model",
+    vlm={
+        "base_url": "https://your-vlm-provider.example/v1",
+        "model": "your-vlm-model",
+        "api_key_env": "YOUR_VLM_API_KEY",
+    },
+)
+```
+
+The evaluator derives the full 50-scene × 50-rollout grid from the scene
+table. Missing or malformed items remain visible as blockers and do not shrink
+the benchmark denominator.
+
+## Scope
+
+This dataset repository contains benchmark inputs only:
+
+- scene definitions, source images, and official prompt material are included;
+- PAWEval code and its rubric implementation live in the GitHub repository;
+- generated model videos, provider responses, and experiment result bundles are
+  intentionally excluded.
+
+The private preview preserves the currently materialized 50-scene package. Its
+final public roster and evaluator-rubric alignment are being completed before a
+public benchmark release.
+
+## Citation
+
+The paper and canonical citation will be added with the public release. For
+private preview work, record the exact Hugging Face revision and GitHub commit
+used for an evaluation.

manifest.json

@@ -0,0 +1,45 @@
+{
+  "schema_version": "pawbench.benchmark_inputs/v1",
+  "dataset_id": "PAWBench",
+  "benchmark_revision": 2,
+  "scene_count": 50,
+  "splits": {
+    "calibration": 25,
+    "coverage": 25
+  },
+  "scene_table": "scenes.jsonl",
+  "source_image_root": "source_images",
+  "prompt_banks": {
+    "gt_guided": {
+      "path": "prompts/gt_guided/prompt_bank.jsonl",
+      "scene_count": 50,
+      "outcome_prompt_count": 230,
+      "sha256": "47d7dd4adaca8d73f16a810c5c29785eb8930b8d295a4e0379cf1f9b93bd6bb7",
+      "owner_accepted_at": "2026-08-14"
+    }
+  },
+  "upstream": {
+    "repository": "Andrew613/PAWBench-Results",
+    "revision": "5fdea8a1e7a1e6ccf69bf5af9cf7947aefc58190",
+    "package_root": "benchmark/PAWBench_V2"
+  },
+  "authority": [
+    {
+      "path": "docs/artifacts/PAWBench/issue_233_v2_source_policy_authority_20260714/source_policy_authority.json",
+      "sha256": "6299512847c7c3d68d3762619dbc549ce41efea526d0b74ca7a95b0d2006b822",
+      "accepted_subsets": [
+        "calibration25_reference_distributions",
+        "coverage25_scene_inputs",
+        "coverage25_support_labels"
+      ]
+    },
+    {
+      "path": "docs/artifacts/PAWBench/issue_233_v2_source_policy_authority_20260714/calibration25_scene_inputs_authority.json",
+      "sha256": "6434001f43824b70ee6f80f8f9e1775ddced143fdf6c39bb26610975c21c1fb0",
+      "accepted_subsets": [
+        "calibration25_scene_inputs"
+      ]
+    }
+  ],
+  "claim_boundary": "Benchmark input materialization only; no generation, evaluation, metric, scientific, manuscript, or publication acceptance."
+}

prompts/gt_guided/README.md

@@ -0,0 +1,38 @@
+# GT-guided prompt bank
+
+This directory contains the owner-accepted semantic prompt bank for the
+GT-guided PAWBench experiment. Acceptance freezes prompt wording and allocator
+semantics; generation, evaluation, scientific acceptance, and paper use remain
+separate decisions.
+
+`prompt_bank.jsonl` contains one row per scene, in the same order as
+`../../scenes.jsonl`. Each row has only the stable `scene_id` join key and an
+ordered `outcome_prompts` array. The array's `outcome_label` values must exactly
+match that scene's `outcome_labels`, in the same order.
+
+Each array entry contains the complete prompt for one accepted outcome. It
+replaces the generic final-state sentence in the benchmark `base_prompt`:
+
+```text
+By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.
+```
+
+with a concrete result followed by the readability requirement:
+
+```text
+{Concrete outcome sentence}. By the final moment, all final positions, spatial relations, and object identities are clearly readable.
+```
+
+No outcome-control suffix is appended. `scenes.jsonl` remains the only owner of
+the split, order, source image, action, base prompt, outcome support, and
+reference distribution; consumers join on `scene_id` instead of copying those
+fields into this bank.
+
+The bank is the semantic vocabulary, not the K=50 execution roster. The
+Calibration allocator assigns outcome counts from the accepted reference
+distribution and the Coverage allocator balances counts across the accepted
+support. A zero-quota Calibration outcome remains in the vocabulary while
+receiving no execution row.
+
+The seven semantic boundary questions and their accepted wording are retained
+in `review.md`.

prompts/gt_guided/review.md

@@ -0,0 +1,899 @@
+# GT-guided prompt bank — accepted owner review
+
+Status: prompt wording and the seven semantic boundaries below were owner-accepted on 2026-08-14. This acceptance authorizes deterministic GT-guided roster compilation; generation, evaluation, scientific acceptance, and paper use remain separate.
+
+The machine-readable bank contains 50 scene records in the same order as `scenes.jsonl`. Each record has an ordered `outcome_prompts` array whose labels exactly match that scene's `outcome_labels`; across all scenes it contains 230 full prompts.
+
+For each outcome, the generic final-state sentence in the benchmark `base_prompt` is replaced by a concrete result sentence followed by the final-state readability requirement. No extra outcome-control suffix is appended.
+
+Review the exact final-state replacements below. Approving a replacement approves the corresponding full prompt.
+
+## Accepted semantic boundaries
+
+1. Confirm that every left/right outcome uses the viewer or camera coordinate frame where the rubric depends on visible image coordinates.
+2. Confirm qualitative wording for `near_miss` versus `clear_miss` in BC-01 and BC-05.
+3. Confirm the visible hand-width / multiple-row distinction for BM-01 small versus large unravel.
+4. Confirm that visible curling or reattachment selects BM-04 curl/reattach rather than generic partial peel.
+5. Confirm the qualitative top-versus-side localized-collapse distinction in BS-02.
+6. Confirm the qualitative shift-versus-scatter distinction in BS-03.
+7. Confirm BC-02 and BC-04 multi-contact trajectories are intentionally not disambiguated in the prompt bank beyond the current PAWEval rubric.
+
+## A-01 — calibration
+
+![A-01](../../source_images/A01.png)
+
+**Action:** Flick the coin once.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Flick the coin once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Reference distribution:** `heads`=1/2, `tails`=1/2
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `heads` | The coin comes to rest heads-side up. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `tails` | The coin comes to rest tails-side up. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## A-02 — calibration
+
+![A-02](../../source_images/A02.png)
+
+**Action:** Throw the die once onto the table.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Throw the die once onto the table. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Reference distribution:** `face_1`=1/6, `face_2`=1/6, `face_3`=1/6, `face_4`=1/6, `face_5`=1/6, `face_6`=1/6
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `face_1` | The die comes to rest with exactly one pip visible on its top face. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `face_2` | The die comes to rest with exactly two pips visible on its top face. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `face_3` | The die comes to rest with exactly three pips visible on its top face. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `face_4` | The die comes to rest with exactly four pips visible on its top face. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `face_5` | The die comes to rest with exactly five pips visible on its top face. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `face_6` | The die comes to rest with exactly six pips visible on its top face. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## A-03 — calibration
+
+![A-03](../../source_images/A03.png)
+
+**Action:** Spin the wheel once under the fixed pointer.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Spin the wheel once under the fixed pointer. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Reference distribution:** `red`=1/2, `blue`=1/2
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `red` | The wheel finishes with the fixed pointer clearly positioned over a red sector. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `blue` | The wheel finishes with the fixed pointer clearly positioned over a blue sector. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## A-04 — calibration
+
+![A-04](../../source_images/A04.png)
+
+**Action:** Spin the wheel once under the fixed pointer.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Spin the wheel once under the fixed pointer. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Reference distribution:** `red`=1/4, `blue`=3/4
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `red` | The wheel finishes with the fixed pointer clearly positioned over the red sector. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `blue` | The wheel finishes with the fixed pointer clearly positioned over the blue sector. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## A-05 — calibration
+
+![A-05](../../source_images/A05.png)
+
+**Action:** Release the ball once.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Release the ball once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Reference distribution:** `left`=1/2, `right`=1/2
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `left` | The released ball clears the central peg and commits to the half of the board on the viewer's left. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `right` | The released ball clears the central peg and commits to the half of the board on the viewer's right. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## A-06 — calibration
+
+![A-06](../../source_images/A06.png)
+
+**Action:** Release the ball once.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Release the ball once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Reference distribution:** `left`=1/4, `middle`=1/2, `right`=1/4
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `left` | The released ball comes to rest in the bottom bin on the viewer's left. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `middle` | The released ball comes to rest in the centered bottom bin. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `right` | The released ball comes to rest in the bottom bin on the viewer's right. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## A-07 — calibration
+
+![A-07](../../source_images/A07.png)
+
+**Action:** Release the ball once.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Release the ball once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Reference distribution:** `left_path`=1/2, `right_path`=1/2
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `left_path` | The released ball passes the central split and travels into the downstream branch on the viewer's left. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `right_path` | The released ball passes the central split and travels into the downstream branch on the viewer's right. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## A-08 — calibration
+
+![A-08](../../source_images/A08.png)
+
+**Action:** Release the ball once.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Release the ball once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Reference distribution:** `left`=1/2, `right`=1/2
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `left` | The released ball leaves the apex and commits to the side of the triangular ramp on the viewer's left. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `right` | The released ball leaves the apex and commits to the side of the triangular ramp on the viewer's right. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## A-09 — calibration
+
+![A-09](../../source_images/A09.png)
+
+**Action:** Move the hand upward once and let the pencil fall.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Move the hand upward once and let the pencil fall. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Reference distribution:** `falls_left`=1/2, `falls_right`=1/2
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `falls_left` | After release, the pencil's eraser end falls and comes to rest on the viewer's left of the vertical line through the sharpened tip's initial contact point. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `falls_right` | After release, the pencil's eraser end falls and comes to rest on the viewer's right of the vertical line through the sharpened tip's initial contact point. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## A-10 — calibration
+
+![A-10](../../source_images/A10.png)
+
+**Action:** Flick both coins once simultaneously.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Flick both coins once simultaneously. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Reference distribution:** `0_heads`=1/4, `1_head`=1/2, `2_heads`=1/4
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `0_heads` | Both coins come to rest tails-side up. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `1_head` | One coin comes to rest heads-side up and the other coin comes to rest tails-side up. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `2_heads` | Both coins come to rest heads-side up. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## A-11 — calibration
+
+![A-11](../../source_images/A11.png)
+
+**Action:** Release the ball once.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Release the ball once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Reference distribution:** `left`=0/1, `right`=1/1
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `left` | The released ball clears the offset peg and settles in the left tray. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `right` | The released ball clears the offset peg and settles in the right tray. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## A-12 — calibration
+
+![A-12](../../source_images/A12.png)
+
+**Action:** Let one small ball enter the inlet and pass through the split once.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Let one small ball enter the inlet and pass through the split once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Reference distribution:** `left_path`=1/2, `right_path`=1/2
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `left_path` | The released ball passes the central split and travels into the downstream branch on the viewer's left. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `right_path` | The released ball passes the central split and travels into the downstream branch on the viewer's right. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## A-13 — calibration
+
+![A-13](../../source_images/A13.png)
+
+**Action:** The man picks up exactly one ball once.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. The man picks up exactly one ball once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Reference distribution:** `red`=1/2, `blue`=1/2
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `red` | The single ball picked up by the man is red. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `blue` | The single ball picked up by the man is blue. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## A-14 — calibration
+
+![A-14](../../source_images/A14.png)
+
+**Action:** The man picks up exactly one ball once.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. The man picks up exactly one ball once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Reference distribution:** `red`=1/2, `blue`=1/2
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `red` | The single ball picked up by the man is red. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `blue` | The single ball picked up by the man is blue. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## A-15 — calibration
+
+![A-15](../../source_images/A15.png)
+
+**Action:** Release the ball once.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Release the ball once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Reference distribution:** `left`=1/2, `right`=1/2
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `left` | The released ball passes the triangular split and settles in the left bottom tray. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `right` | The released ball passes the triangular split and settles in the right bottom tray. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## A16 — calibration
+
+![A16](../../source_images/A16.png)
+
+**Action:** Release the ball once.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Release the ball once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Reference distribution:** `red_group`=1/4, `blue_group`=1/4, `green_group`=1/4, `yellow_group`=1/4
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `red_group` | The released ball settles inside one of the red pockets of the roulette wheel. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `blue_group` | The released ball settles inside one of the blue pockets of the roulette wheel. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `green_group` | The released ball settles inside one of the green pockets of the roulette wheel. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `yellow_group` | The released ball settles inside one of the yellow pockets of the roulette wheel. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## A17 — calibration
+
+![A17](../../source_images/A17.png)
+
+**Action:** Release the ball once.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Release the ball once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Reference distribution:** `left_basin`=1/2, `right_basin`=1/2
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `left_basin` | The released ball leaves the ridge and settles in the left basin. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `right_basin` | The released ball leaves the ridge and settles in the right basin. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## A18 — calibration
+
+![A18](../../source_images/A18.png)
+
+**Action:** Spin the spinning top once.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Spin the spinning top once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Reference distribution:** `red_sector`=1/4, `blue_sector`=1/4, `green_sector`=1/4, `yellow_sector`=1/4
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `red_sector` | The spinning top finishes with its center clearly inside the red sector. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `blue_sector` | The spinning top finishes with its center clearly inside the blue sector. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `green_sector` | The spinning top finishes with its center clearly inside the green sector. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `yellow_sector` | The spinning top finishes with its center clearly inside the yellow sector. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## A19 — calibration
+
+![A19](../../source_images/A19.png)
+
+**Action:** Turn the side crank briefly and open the bottom gate once so exactly one ball exits.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Turn the side crank briefly and open the bottom gate once so exactly one ball exits. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Reference distribution:** `red`=1/4, `green`=1/4, `blue`=1/4, `yellow`=1/4
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `red` | Exactly one ball exits the bottom gate, and that exiting ball is red. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `green` | Exactly one ball exits the bottom gate, and that exiting ball is green. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `blue` | Exactly one ball exits the bottom gate, and that exiting ball is blue. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `yellow` | Exactly one ball exits the bottom gate, and that exiting ball is yellow. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## A20 — calibration
+
+![A20](../../source_images/A20.png)
+
+**Action:** Release the ball once.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Release the ball once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Reference distribution:** `left`=1/4, `middle`=1/2, `right`=1/4
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `left` | The released ball comes to rest in the bottom bin on the viewer's left. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `middle` | The released ball comes to rest in the centered bottom bin. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `right` | The released ball comes to rest in the bottom bin on the viewer's right. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## A21 — calibration
+
+![A21](../../source_images/A21.png)
+
+**Action:** Roll both dice once.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Roll both dice once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Reference distribution:** `sum_2`=1/36, `sum_3`=2/36, `sum_4`=3/36, `sum_5`=4/36, `sum_6`=5/36, `sum_7`=6/36, `sum_8`=5/36, `sum_9`=4/36, `sum_10`=3/36, `sum_11`=2/36, `sum_12`=1/36
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `sum_2` | Both dice come to rest with their visible top-face values adding up to 2. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `sum_3` | Both dice come to rest with their visible top-face values adding up to 3. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `sum_4` | Both dice come to rest with their visible top-face values adding up to 4. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `sum_5` | Both dice come to rest with their visible top-face values adding up to 5. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `sum_6` | Both dice come to rest with their visible top-face values adding up to 6. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `sum_7` | Both dice come to rest with their visible top-face values adding up to 7. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `sum_8` | Both dice come to rest with their visible top-face values adding up to 8. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `sum_9` | Both dice come to rest with their visible top-face values adding up to 9. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `sum_10` | Both dice come to rest with their visible top-face values adding up to 10. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `sum_11` | Both dice come to rest with their visible top-face values adding up to 11. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `sum_12` | Both dice come to rest with their visible top-face values adding up to 12. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## A22 — calibration
+
+![A22](../../source_images/A22.png)
+
+**Action:** Release the ball once.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Release the ball once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Reference distribution:** `bin_1`=1/16, `bin_2`=4/16, `bin_3`=6/16, `bin_4`=4/16, `bin_5`=1/16
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `bin_1` | The released ball settles in bin 1, the leftmost of the five bottom bins. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `bin_2` | The released ball settles in bin 2, the second bottom bin from the left. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `bin_3` | The released ball settles in bin 3, the middle of the five bottom bins. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `bin_4` | The released ball settles in bin 4, the second bottom bin from the right. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `bin_5` | The released ball settles in bin 5, the rightmost of the five bottom bins. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## A23 — calibration
+
+![A23](../../source_images/A23.png)
+
+**Action:** Flick the coin once.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Flick the coin once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Reference distribution:** `heads`=1/2, `tails`=1/2
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `heads` | The coin comes to rest heads-side up. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `tails` | The coin comes to rest tails-side up. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## A24 — calibration
+
+![A24](../../source_images/A24.png)
+
+**Action:** Turn the side crank briefly and open the bottom gate once so exactly one ball exits.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Turn the side crank briefly and open the bottom gate once so exactly one ball exits. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Reference distribution:** `red`=1/4, `green`=1/4, `blue`=1/4, `yellow`=1/4
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `red` | Exactly one ball exits the bottom gate, and that exiting ball is red. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `green` | Exactly one ball exits the bottom gate, and that exiting ball is green. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `blue` | Exactly one ball exits the bottom gate, and that exiting ball is blue. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `yellow` | Exactly one ball exits the bottom gate, and that exiting ball is yellow. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## A25 — calibration
+
+![A25](../../source_images/A25.png)
+
+**Action:** Move the hand upward once and let the pencil fall.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Move the hand upward once and let the pencil fall. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Reference distribution:** `falls_left`=1/1, `falls_right`=0/1
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `falls_left` | After release, the pencil falls with its eraser end settling to the viewer's left of the image-vertical line through the sharpened tip's initial table-contact point. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `falls_right` | After release, the pencil falls with its eraser end settling to the viewer's right of the image-vertical line through the sharpened tip's initial table-contact point. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## BA-01-02 — coverage
+
+![BA-01-02](../../source_images/BA-01-02.png)
+
+**Action:** Gently touch the cat once.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Gently touch the cat once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Coverage support:** `cat_accepts_touch_or_leans_in`, `cat_remains_still_near_hand`, `cat_withdraws_but_stays_in_area`, `cat_leaves_interaction_area`, `cat_defensive_posture_without_contact`, `cat_active_defensive_contact`
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `cat_accepts_touch_or_leans_in` | After the gentle touch, the cat visibly relaxes into the contact, leans toward the hand, or rubs against it without withdrawing, leaving, showing a defensive posture, or making defensive contact. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `cat_remains_still_near_hand` | After the gentle touch, the cat stays beside the hand in the same position without accepting or leaning into the touch, withdrawing, leaving, showing a defensive posture, or making defensive contact. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `cat_withdraws_but_stays_in_area` | After the gentle touch, the cat recoils or pulls its head or body away from the hand while remaining in the same sofa interaction area, without defensive posture or defensive contact. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `cat_leaves_interaction_area` | After the gentle touch, the cat jumps or runs out of the touchable sofa interaction area without first making active defensive contact, so the hand can no longer continue the interaction. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `cat_defensive_posture_without_contact` | After the gentle touch, the cat remains in the sofa interaction area and arches its back, raises its fur, flattens its ears, opens its mouth in threat, or crouches defensively without touching the hand. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `cat_active_defensive_contact` | After the gentle touch, the cat swats, scratches, bites, or lunges into the hand with clear physical contact. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## BA-01-S1 — coverage
+
+![BA-01-S1](../../source_images/BA-01-S1.png)
+
+**Action:** Gently touch the cat once.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Gently touch the cat once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Coverage support:** `cat_accepts_touch_or_leans_in`, `cat_remains_still_near_hand`, `cat_withdraws_but_stays_in_area`, `cat_leaves_interaction_area`, `cat_defensive_posture_without_contact`, `cat_active_defensive_contact`
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `cat_accepts_touch_or_leans_in` | After the gentle touch, the cat visibly relaxes into the contact, leans toward the hand, or rubs against it without withdrawing, leaving, showing a defensive posture, or making defensive contact. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `cat_remains_still_near_hand` | After the gentle touch, the cat stays beside the hand in the same position without accepting or leaning into the touch, withdrawing, leaving, showing a defensive posture, or making defensive contact. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `cat_withdraws_but_stays_in_area` | After the gentle touch, the cat recoils or pulls its head or body away from the hand while remaining in the same sofa interaction area, without defensive posture or defensive contact. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `cat_leaves_interaction_area` | After the gentle touch, the cat jumps or runs out of the touchable sofa interaction area without first making active defensive contact, so the hand can no longer continue the interaction. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `cat_defensive_posture_without_contact` | After the gentle touch, the cat remains in the sofa interaction area and arches its back, raises its fur, flattens its ears, opens its mouth in threat, or crouches defensively without touching the hand. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `cat_active_defensive_contact` | After the gentle touch, the cat swats, scratches, bites, or lunges into the hand with clear physical contact. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## BA-01 — coverage
+
+![BA-01](../../source_images/BA-01.png)
+
+**Action:** Gently touch the cat once.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Gently touch the cat once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Coverage support:** `cat_accepts_touch_or_leans_in`, `cat_remains_still_near_hand`, `cat_withdraws_but_stays_in_area`, `cat_leaves_interaction_area`, `cat_defensive_posture_without_contact`, `cat_active_defensive_contact`
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `cat_accepts_touch_or_leans_in` | After the gentle touch, the cat visibly relaxes into the contact, leans toward the hand, or rubs against it without withdrawing, leaving, showing a defensive posture, or making defensive contact. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `cat_remains_still_near_hand` | After the gentle touch, the cat stays beside the hand in the same position without accepting or leaning into the touch, withdrawing, leaving, showing a defensive posture, or making defensive contact. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `cat_withdraws_but_stays_in_area` | After the gentle touch, the cat recoils or pulls its head or body away from the hand while remaining in the same sofa interaction area, without defensive posture or defensive contact. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `cat_leaves_interaction_area` | After the gentle touch, the cat jumps or runs out of the touchable sofa interaction area without first making active defensive contact, so the hand can no longer continue the interaction. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `cat_defensive_posture_without_contact` | After the gentle touch, the cat remains in the sofa interaction area and arches its back, raises its fur, flattens its ears, opens its mouth in threat, or crouches defensively without touching the hand. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `cat_active_defensive_contact` | After the gentle touch, the cat swats, scratches, bites, or lunges into the hand with clear physical contact. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## BA-02 — coverage
+
+![BA-02](../../source_images/BA-02.png)
+
+**Action:** Let the stranger approach the dog.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Let the stranger approach the dog. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Coverage support:** `dog_approaches_person`, `dog_holds_position_facing_person`, `dog_turns_away_without_retreating`, `dog_retreats_or_avoids_person`, `dog_leaves_interaction_area`, `dog_shows_defensive_posture`, `dog_active_defensive_advance_or_contact`
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `dog_approaches_person` | As the person approaches, the dog walks toward them and ends visibly closer to them without defensive posture or active defensive advance/contact. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `dog_holds_position_facing_person` | As the person approaches, the dog stays near its starting position and keeps its head and body facing them, with no approach, turn-away, retreat, leaving, defensive posture, or active defensive advance/contact. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `dog_turns_away_without_retreating` | As the person approaches, the dog turns its head or body away from them while staying near its starting position, without increasing distance, defensive posture, or active defensive advance/contact. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `dog_retreats_or_avoids_person` | As the person approaches, the dog backs away, sidesteps, or arcs away so it visibly increases distance from the person while remaining within the path-side interaction area, without defensive posture or active defensive advance/contact. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `dog_leaves_interaction_area` | As the person approaches, the dog runs out of the path-side interaction area without first making an active defensive advance/contact, so the person-dog interaction can no longer continue. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `dog_shows_defensive_posture` | As the person approaches, the dog stays near its starting position and bares its teeth, opens its mouth in threat, stiffens or lowers its body, raises its hackles, or pins its ears without advancing or making contact. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `dog_active_defensive_advance_or_contact` | As the person approaches, the dog lunges, charges, snaps, bites, or jumps toward them in a clear active defensive advance or contact. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## BA-03 — coverage
+
+![BA-03](../../source_images/BA-03.png)
+
+**Action:** Let the dog make one reach attempt toward the toy.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Let the dog make one reach attempt toward the toy. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Coverage support:** `goes_around_left`, `goes_around_right`, `jumps_over_middle`, `crawls_or_ducks_under_middle`, `does_not_attempt`
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `goes_around_left` | The dog passes around the camera-left end of the low obstacle toward the toy. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `goes_around_right` | The dog passes around the camera-right end of the low obstacle toward the toy. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `jumps_over_middle` | The dog jumps across the middle of the low obstacle with its body and legs visibly clearing the bar from above. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `crawls_or_ducks_under_middle` | The dog crosses through the middle by lowering its body and passing visibly under the bar rather than around either end. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `does_not_attempt` | The dog has a visible opportunity to reach the toy but does not commit to any path around, over, or under the obstacle. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## BA-06 — coverage
+
+![BA-06](../../source_images/BA-06.png)
+
+**Action:** Let the cat reach or step toward the tabletop objects once from the current poised posture.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Let the cat reach or step toward the tabletop objects once from the current poised posture. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Coverage support:** `cat_does_not_enter_tabletop_area`, `cat_enters_tabletop_area_without_object_contact`, `cat_touches_object_without_displacing`, `cat_displaces_non_cup_object`, `cat_displaces_cup_without_spill`, `cat_spills_or_tips_cup`
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `cat_does_not_enter_tabletop_area` | The cat remains off the tabletop after reaching or stepping toward it, with no paw or body placed on the tabletop surface. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `cat_enters_tabletop_area_without_object_contact` | The cat newly places a paw or part of its body on the tabletop surface without touching the cup or any other tabletop object. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `cat_touches_object_without_displacing` | The cat visibly touches the cup or another tabletop object, but the touched object does not shift, rotate, fall, or tip. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `cat_displaces_non_cup_object` | The cat visibly shifts, knocks over, or knocks down a tabletop object other than the cup while the cup does not move, tip, or spill. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `cat_displaces_cup_without_spill` | The cat visibly shifts or rotates the cup while it remains upright and no liquid or contents spill. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `cat_spills_or_tips_cup` | The cat visibly tips or knocks down the cup, or causes liquid or contents to spill from it. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## BA-07 — coverage
+
+![BA-07](../../source_images/BA-07.png)
+
+**Action:** Gently pull the leash once while the dog is eating.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Gently pull the leash once while the dog is eating. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Coverage support:** `dog_continues_eating_at_bowl`, `dog_pauses_or_lifts_head_near_bowl`, `dog_resists_taut_leash_stays_near_bowl`, `dog_steps_back_from_bowl`, `bowl_shifts_without_food_spill`, `food_spills_or_bowl_tips`
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `dog_continues_eating_at_bowl` | The dog keeps its head at the bowl and continues eating through the visible leash pull without stepping back, moving the bowl, or visibly bracing against a taut leash. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `dog_pauses_or_lifts_head_near_bowl` | The dog stops eating or lifts its head after the leash pull but remains near the bowl without visibly bracing against a taut leash, while the bowl remains upright and in place with no spill. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `dog_resists_taut_leash_stays_near_bowl` | The leash becomes visibly taut while the dog braces or leans against the pull and stays near the bowl without stepping back, while the bowl remains upright and in place with no spill. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `dog_steps_back_from_bowl` | The dog visibly steps back and increases its distance from the bowl after the leash pull while the bowl stays upright and in place. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `bowl_shifts_without_food_spill` | The bowl visibly slides or rotates during the dog and leash interaction while remaining upright with no food spilling out. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `food_spills_or_bowl_tips` | Food visibly spills out of the bowl or the bowl visibly tips or falls during the dog and leash interaction. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## BA-08 — coverage
+
+![BA-08](../../source_images/BA-08.png)
+
+**Action:** An off-screen hand enters the frame once and offers food to the bird.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. An off-screen hand enters the frame once and offers food to the bird. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Coverage support:** `bird_takes_food_from_hand`, `bird_flies_off_or_leaves_table`, `bird_moves_away_but_stays_on_table`, `bird_displaces_food_without_taking`, `bird_pecks_food_without_taking`, `bird_stays_near_without_food_contact`
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `bird_takes_food_from_hand` | The food visibly leaves the offering hand and is held, carried, or controlled by the bird. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `bird_flies_off_or_leaves_table` | The bird flies off or otherwise leaves the tabletop without first taking the offered food. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `bird_moves_away_but_stays_on_table` | The bird visibly increases its distance from the offered food and hand while remaining on the tabletop and without taking the food. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `bird_displaces_food_without_taking` | The bird visibly knocks, drops, or moves the offered food from its position without holding or carrying it, while the bird remains near the offer on the tabletop and does not move away or leave the table. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `bird_pecks_food_without_taking` | The bird visibly pecks or touches the offered food with its beak while the food remains in the hand or offer position and is not taken, while the bird remains near the offer on the tabletop and does not move away or leave the table. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `bird_stays_near_without_food_contact` | The bird remains near the offered food and hand without touching the food, moving away, or leaving the tabletop. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## BC-01-I2 — coverage
+
+![BC-01-I2](../../source_images/BC-01-I2.png)
+
+**Action:** Toss the ball once from the visible hand pose toward the cup.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Toss the ball once from the visible hand pose toward the cup. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Coverage support:** `clean_in_cup`, `contact_then_in_cup`, `cup_contact_then_out`, `near_miss_no_cup_contact`, `clear_miss`
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `clean_in_cup` | The tossed ball passes cleanly through the cup mouth without touching the cup or rim and remains inside the cup. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `contact_then_in_cup` | The tossed ball visibly contacts the rim or another part of the cup setup, then enters through the cup mouth and remains inside the cup. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `cup_contact_then_out` | The tossed ball visibly contacts the cup or rim, then bounces away and comes to rest outside the cup. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `near_miss_no_cup_contact` | The tossed ball passes close beside the cup mouth without touching any part of the cup and comes to rest outside the cup. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `clear_miss` | The tossed ball passes plainly far from the cup without touching it and comes to rest outside the cup. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## BC-01 — coverage
+
+![BC-01](../../source_images/BC-01.png)
+
+**Action:** Toss the ball once from the visible hand pose toward the cup.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Toss the ball once from the visible hand pose toward the cup. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Coverage support:** `clean_in_cup`, `contact_then_in_cup`, `cup_contact_then_out`, `near_miss_no_cup_contact`, `clear_miss`
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `clean_in_cup` | The tossed ball passes cleanly through the cup mouth without touching the cup or rim and remains inside the cup. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `contact_then_in_cup` | The tossed ball visibly contacts the rim or another part of the cup setup, then enters through the cup mouth and remains inside the cup. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `cup_contact_then_out` | The tossed ball visibly contacts the cup or rim, then bounces away and comes to rest outside the cup. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `near_miss_no_cup_contact` | The tossed ball passes close beside the cup mouth without touching any part of the cup and comes to rest outside the cup. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `clear_miss` | The tossed ball passes plainly far from the cup without touching it and comes to rest outside the cup. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## BC-02 — coverage
+
+![BC-02](../../source_images/BC-02.jpg)
+
+**Action:** Shoot the basketball once from the visible shooting pose toward the hoop.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Shoot the basketball once from the visible shooting pose toward the hoop. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Coverage support:** `swish`, `backboard_then_in`, `rim_then_in`, `rim_miss_no_score`, `backboard_miss_no_score`, `left_miss_no_contact`, `right_miss_no_contact`, `short_airball_no_contact`, `long_airball_no_contact`
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `swish` | The basketball passes cleanly down through the hoop without touching the rim or backboard and scores. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `backboard_then_in` | The basketball visibly strikes the backboard first, then passes down through the hoop and scores. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `rim_then_in` | The basketball visibly contacts the rim before passing down through the hoop and scoring. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `rim_miss_no_score` | The basketball visibly contacts the rim and does not score. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `backboard_miss_no_score` | The basketball visibly contacts the backboard and does not score. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `left_miss_no_contact` | The basketball passes completely to the camera-visible left of the hoop without touching the rim or backboard and does not score. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `right_miss_no_contact` | The basketball passes completely to the camera-visible right of the hoop without touching the rim or backboard and does not score. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `short_airball_no_contact` | The basketball falls short in front of the hoop without reaching or touching the rim or backboard and does not score. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `long_airball_no_contact` | The basketball passes beyond or over the hoop region without target contact and does not score. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## BC-03-I1 — coverage
+
+![BC-03-I1](../../source_images/BC-03-I1.png)
+
+**Action:** Roll the bowling ball once down the lane.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Roll the bowling ball once down the lane. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Coverage support:** `all_7_pins_down`, `six_pins_down`, `five_pins_down`, `four_pins_down`, `three_pins_down`, `two_pins_down`, `one_pin_down`, `zero_pins_down_after_contact`, `no_pin_contact`
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `all_7_pins_down` | The bowling ball visibly contacts the seven-pin rack and all seven pins are clearly knocked down. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `six_pins_down` | The bowling ball visibly contacts the seven-pin rack, exactly six of the seven pins are clearly knocked down, and exactly one pin remains standing. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `five_pins_down` | The bowling ball visibly contacts the seven-pin rack, exactly five of the seven pins are clearly knocked down, and exactly two pins remain standing. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `four_pins_down` | The bowling ball visibly contacts the seven-pin rack, exactly four of the seven pins are clearly knocked down, and exactly three pins remain standing. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `three_pins_down` | The bowling ball visibly contacts the seven-pin rack, exactly three of the seven pins are clearly knocked down, and exactly four pins remain standing. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `two_pins_down` | The bowling ball visibly contacts the seven-pin rack, exactly two of the seven pins are clearly knocked down, and exactly five pins remain standing. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `one_pin_down` | The bowling ball visibly contacts the seven-pin rack, exactly one of the seven pins is clearly knocked down, and exactly six pins remain standing. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `zero_pins_down_after_contact` | The bowling ball visibly contacts the seven-pin rack or an individual pin, but none of the seven pins is knocked down and all seven pins remain standing. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `no_pin_contact` | The bowling ball visibly misses the seven-pin rack without contacting any pin, and all seven pins remain standing. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## BC-03 — coverage
+
+![BC-03](../../source_images/BC-03.png)
+
+**Action:** Roll the bowling ball once down the lane.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Roll the bowling ball once down the lane. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Coverage support:** `all_7_pins_down`, `six_pins_down`, `five_pins_down`, `four_pins_down`, `three_pins_down`, `two_pins_down`, `one_pin_down`, `zero_pins_down_after_contact`, `no_pin_contact`
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `all_7_pins_down` | The bowling ball visibly contacts the seven-pin rack and all seven pins are clearly knocked down. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `six_pins_down` | The bowling ball visibly contacts the seven-pin rack, exactly six of the seven pins are clearly knocked down, and exactly one pin remains standing. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `five_pins_down` | The bowling ball visibly contacts the seven-pin rack, exactly five of the seven pins are clearly knocked down, and exactly two pins remain standing. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `four_pins_down` | The bowling ball visibly contacts the seven-pin rack, exactly four of the seven pins are clearly knocked down, and exactly three pins remain standing. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `three_pins_down` | The bowling ball visibly contacts the seven-pin rack, exactly three of the seven pins are clearly knocked down, and exactly four pins remain standing. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `two_pins_down` | The bowling ball visibly contacts the seven-pin rack, exactly two of the seven pins are clearly knocked down, and exactly five pins remain standing. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `one_pin_down` | The bowling ball visibly contacts the seven-pin rack, exactly one of the seven pins is clearly knocked down, and exactly six pins remain standing. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `zero_pins_down_after_contact` | The bowling ball visibly contacts the seven-pin rack or an individual pin, but none of the seven pins is knocked down and all seven pins remain standing. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `no_pin_contact` | The bowling ball visibly misses the seven-pin rack without contacting any pin, and all seven pins remain standing. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## BC-04 — coverage
+
+![BC-04](../../source_images/BC-04.jpg)
+
+**Action:** Kick the ball once toward the goal.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Kick the ball once toward the goal. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Coverage support:** `clean_goal`, `keeper_deflected_goal`, `frame_deflected_goal`, `goalkeeper_prevents_goal`, `post_out`, `crossbar_out`, `miss_left`, `miss_right`, `miss_high`, `shot_stops_short`
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `clean_goal` | The ball visibly crosses the goal plane into the net without contacting the goalkeeper, another defender, or any part of the goal frame. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `keeper_deflected_goal` | The goalkeeper visibly contacts and deflects the ball, after which the ball crosses the goal plane into the net. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `frame_deflected_goal` | The ball visibly contacts a goalpost or the crossbar and then crosses the goal plane into the net. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `goalkeeper_prevents_goal` | The goalkeeper visibly saves, catches, blocks, or deflects the ball away so that it remains outside the goal and never crosses the goal plane. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `post_out` | The ball visibly hits the left or right goalpost and rebounds or deflects away without crossing the goal plane. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `crossbar_out` | The ball visibly hits the crossbar and rebounds or deflects away without crossing the goal plane. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `miss_left` | The ball reaches the goal region but passes completely outside the screen-left side of the goal frame without contacting the goalkeeper or the goal frame. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `miss_right` | The ball reaches the goal region but passes completely outside the screen-right side of the goal frame without contacting the goalkeeper or the goal frame. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `miss_high` | The ball reaches the goal region but passes completely above the crossbar without contacting the goalkeeper or the goal frame. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `shot_stops_short` | The kicked ball loses motion and comes to rest on the field before reaching the goal mouth, without contacting the goalkeeper or the goal frame. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## BC-05 — coverage
+
+![BC-05](../../source_images/BC-05.jpg)
+
+**Action:** Toss the ring once from the visible hand pose toward the peg.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Toss the ring once from the visible hand pose toward the peg. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Coverage support:** `settles_around_peg`, `rests_on_peg_not_scored`, `contacts_peg_but_off`, `near_miss_no_peg_contact`, `clear_miss`
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `settles_around_peg` | The ring finishes settled around the wooden peg with the peg visibly passing through the ring's center opening. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `rests_on_peg_not_scored` | The ring finishes supported by, hanging on, or leaning against the peg without the peg passing through the ring's center opening. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `contacts_peg_but_off` | The ring visibly strikes the peg, then finishes off the peg and is no longer touching or supported by it. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `near_miss_no_peg_contact` | The ring passes visibly close to the peg without touching it, then finishes on the table beside the peg. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `clear_miss` | The ring travels clearly away from the peg without touching it, then finishes well separated from the peg on the table. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## BC-06 — coverage
+
+![BC-06](../../source_images/BC-06.jpg)
+
+**Action:** Throw the dart once from the visible hand pose toward the board.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Throw the dart once from the visible hand pose toward the board. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Coverage support:** `bullseye`, `outer_bull`, `inner_single_hit`, `triple_ring_hit`, `outer_single_hit`, `double_ring_hit`, `board_bounce_out`, `off_board_miss`
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `bullseye` | The dart finishes embedded in the small innermost center bullseye of the dartboard. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `outer_bull` | The dart finishes embedded in the outer bull ring immediately surrounding the innermost center bullseye. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `inner_single_hit` | The dart finishes embedded in an inner single-score wedge between the outer bull and the narrow triple ring. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `triple_ring_hit` | The dart finishes embedded in the narrow triple-scoring ring midway across the dartboard wedges. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `outer_single_hit` | The dart finishes embedded in an outer single-score wedge between the narrow triple ring and the narrow double ring. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `double_ring_hit` | The dart finishes embedded in the narrow double-scoring ring near the outer edge of the dartboard. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `board_bounce_out` | The dart visibly strikes the dartboard, rebounds away, and finishes without remaining embedded anywhere on the board. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `off_board_miss` | The dart completely misses the dartboard without touching it and finishes in the visible off-board area. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## BC-09 — coverage
+
+![BC-09](../../source_images/BC-09.png)
+
+**Action:** Release the ball once into the board.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Release the ball once into the board. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Coverage support:** `left_bin`, `middle_bin`, `right_bin`, `stuck_on_board`
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `left_bin` | The metal ball finishes settled inside the camera-left bottom collection bin. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `middle_bin` | The metal ball finishes settled inside the center bottom collection bin. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `right_bin` | The metal ball finishes settled inside the camera-right bottom collection bin. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `stuck_on_board` | The metal ball stops and remains visibly settled on the playing surface before reaching any of the three bottom collection bins. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## BM-01 — coverage
+
+![BM-01](../../source_images/BM-01.png)
+
+**Action:** Pull the loose yarn end once.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Pull the loose yarn end once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Coverage support:** `thread_pulls_free_cleanly`, `small_unravel`, `large_unravel_or_ladder`, `thread_snaps`, `thread_stays_attached_no_visible_change`
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `thread_pulls_free_cleanly` | The loose yarn end pulls completely free from the knitted fabric while the pull site shows no new released loops, laddering, widening gap, or hole. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `small_unravel` | Pulling the loose yarn releases only a short local loop beside the pull site, with the released strand staying close to the fabric and the surrounding knitted rows remaining intact. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `large_unravel_or_ladder` | Pulling the loose yarn creates an extended run, ladder, widening gap, or broad release across multiple knitted rows, or pulls out a long loop or strand spanning at least a hand-width from the fabric. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `thread_snaps` | The pulled yarn visibly snaps into two separated ends while the knitted fabric and the broken yarn ends remain clearly visible. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `thread_stays_attached_no_visible_change` | The loose yarn is visibly pulled but stays attached, and the pull site shows no new released loop, ladder, widening gap, hole, or snapped end. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## BM-02 — coverage
+
+![BM-02](../../source_images/BM-02.png)
+
+**Action:** Bend the chocolate bar once until it responds.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Bend the chocolate bar once until it responds. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Coverage support:** `no_visible_break`, `single_clean_break`, `partial_crack_not_separated`, `multiple_piece_break`, `crumbs_or_small_fragment_break`
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `no_visible_break` | After the bend, the chocolate bar remains one connected piece with no new crack line, separation, crumbs, flakes, or small fragments. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `single_clean_break` | The bend cleanly separates the chocolate bar into exactly two large pieces along one break line, with no additional pieces, crumbs, flakes, or small fragments. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `partial_crack_not_separated` | The bend creates a visible new crack line in the chocolate, but the bar remains materially connected as one piece with no separated section or fragments. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `multiple_piece_break` | The bend breaks the chocolate bar into three or more separated large pieces, with no crumbs, flakes, or small fragments. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `crumbs_or_small_fragment_break` | The bend visibly breaks brittle chocolate from the bar as newly created crumbs, flakes, or small fragments that remain visible beside the larger chocolate pieces. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## BM-03 — coverage
+
+![BM-03](../../source_images/BM-03.png)
+
+**Action:** Flip the egg once with the spatula.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Flip the egg once with the spatula. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Coverage support:** `lands_yolk_up_intact`, `lands_yolk_down_intact`, `folds_without_tearing`, `white_tears_yolk_intact`, `yolk_ruptures`, `spatula_contact_no_flip`
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `lands_yolk_up_intact` | The egg completes the flip and lands flat and unfolded in the pan with the intact yolk visibly facing upward and the egg white unbroken. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `lands_yolk_down_intact` | The egg completes the flip and lands flat and unfolded with the yolk side against the pan, so the yolk is no longer visible while the egg remains intact. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `folds_without_tearing` | During the flip, the egg folds over onto itself and settles visibly folded while both the egg white and yolk remain unbroken. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `white_tears_yolk_intact` | The egg white visibly tears or separates while the yolk remains intact and does not leak. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `yolk_ruptures` | During the flip, the yolk membrane visibly ruptures and liquid yolk leaks and spreads onto the egg white or pan. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `spatula_contact_no_flip` | The spatula visibly contacts the egg, but the attempted lift or turn does not complete and the egg never flips over. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## BM-04 — coverage
+
+![BM-04](../../source_images/BM-04.png)
+
+**Action:** Peel the tape from the cardboard once.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Peel the tape from the cardboard once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Coverage support:** `clean_full_peel`, `partial_peel_tape_still_attached`, `tape_tears`, `cardboard_surface_tears`, `tape_curls_or_reattaches_without_full_removal`
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `clean_full_peel` | The tape peels completely off the cardboard as one intact strip, leaving no tape attached and no lifted cardboard fibers or torn surface layer. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `partial_peel_tape_still_attached` | The tape peels up along part of its length but the remaining section stays visibly bonded to the cardboard when the peeling motion stops, and the lifted portion remains extended and does not curl, fold back, or reattach. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `tape_tears` | The tape strip visibly tears, splits, or breaks during the peel before clean full removal. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `cardboard_surface_tears` | The top cardboard surface layer visibly tears and lifts with the tape, leaving an exposed rough fibrous patch on the box. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `tape_curls_or_reattaches_without_full_removal` | The peeled tape curls, folds back, or sticks again before full removal. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## BS-01-I1 — coverage
+
+![BS-01-I1](../../source_images/BS-01-I1.png)
+
+**Action:** Flip the bottle once.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Flip the bottle once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Coverage support:** `upright_on_base`, `balanced_on_cap`, `side_laying`, `leaning_against_visible_support`
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `upright_on_base` | After the flip, the bottle comes to rest vertically on its base without support from any other object. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `balanced_on_cap` | After the flip, the bottle comes to rest upside down, balanced vertically on its cap without support from any other object. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `side_laying` | After the flip, the bottle comes to rest horizontally on its side on the tabletop without leaning against another object. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `leaning_against_visible_support` | After the flip, the bottle comes to rest tilted against a clearly visible support that bears its weight. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## BS-01 — coverage
+
+![BS-01](../../source_images/BS-01.png)
+
+**Action:** Flip the bottle once.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Flip the bottle once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Coverage support:** `upright_on_base`, `balanced_on_cap`, `side_laying`, `leaning_against_visible_support`
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `upright_on_base` | After the flip, the bottle comes to rest vertically on its base without support from any other object. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `balanced_on_cap` | After the flip, the bottle comes to rest upside down, balanced vertically on its cap without support from any other object. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `side_laying` | After the flip, the bottle comes to rest horizontally on its side on the tabletop without leaning against another object. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `leaning_against_visible_support` | After the flip, the bottle comes to rest tilted against a clearly visible support that bears its weight. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## BS-02 — coverage
+
+![BS-02](../../source_images/BS-02.png)
+
+**Action:** Pull the red-marked block once from the tower.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Pull the red-marked block once from the tower. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Coverage support:** `tower_remains_standing_after_removal`, `tower_remains_standing_but_displaced`, `top_section_collapse`, `side_section_collapse`, `full_collapse`
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `tower_remains_standing_after_removal` | After the red-marked block is pulled completely out, the remaining tower stays upright and essentially unchanged in its original position and alignment. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `tower_remains_standing_but_displaced` | After the red-marked block is pulled completely out, the remaining tower stays standing as one continuous stack but is visibly shifted or leaning from its original position or alignment. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `top_section_collapse` | After the red-marked block is pulled completely out, the upper section of the tower loses its stacked structure and falls while most of the lower tower remains standing. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `side_section_collapse` | After the red-marked block is pulled completely out, blocks from one side of the tower break away and fall while most of the tower remains standing. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `full_collapse` | After the red-marked block is pulled completely out, the tower loses its overall vertical structure and most of its blocks fall. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+
+## BS-03 — coverage
+
+![BS-03](../../source_images/BS-03.png)
+
+**Action:** Pull the tablecloth out once.
+
+**Prompt prefix:** Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Pull the tablecloth out once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image.
+
+**Coverage support:** `objects_remain_in_place`, `objects_shift_but_upright`, `objects_scatter_on_table`, `one_object_falls_on_table`, `multiple_objects_fall_on_table`, `one_object_leaves_table`, `multiple_objects_leave_table`
+
+| Outcome label | Final-state replacement |
+| --- | --- |
+| `objects_remain_in_place` | After the tablecloth is pulled clear, all tabletop objects remain in their original positions and orientations. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `objects_shift_but_upright` | After the tablecloth is pulled clear, the tabletop objects visibly shift but stay close together, with no object tipping over or leaving the table. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `objects_scatter_on_table` | After the tablecloth is pulled clear, the tabletop objects spread apart across the tabletop while remaining on the table, with no object tipping over. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `one_object_falls_on_table` | After the tablecloth is pulled clear, exactly one initially upright tabletop object tips over and remains on the tabletop, while no object leaves the table. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `multiple_objects_fall_on_table` | After the tablecloth is pulled clear, at least two initially upright tabletop objects tip over and remain on the tabletop, while no object leaves the table. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `one_object_leaves_table` | After the tablecloth is pulled clear, exactly one tabletop object ends off the tabletop while all other tabletop objects remain on the table. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |
+| `multiple_objects_leave_table` | After the tablecloth is pulled clear, at least two tabletop objects end off the tabletop. By the final moment, all final positions, spatial relations, and object identities are clearly readable. |

scenes.jsonl

@@ -0,0 +1,50 @@
+{"scene_id":"A-01","split":"calibration","split_order":0,"action":"Flick the coin once.","source_image_path":"source_images/A01.png","source_image_bytes":8188278,"source_image_sha256":"d90a04cf6dfce9e3c0fe8ac7360024f03efa44d019995dcc48eb017854dd89eb","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Flick the coin once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["heads","tails"],"reference_distribution":{"heads":0.5,"tails":0.5},"reference_distribution_exact":[{"denominator":2,"label":"heads","numerator":1},{"denominator":2,"label":"tails","numerator":1}],"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/A01.png"}
+{"scene_id":"A-02","split":"calibration","split_order":1,"action":"Throw the die once onto the table.","source_image_path":"source_images/A02.png","source_image_bytes":8018738,"source_image_sha256":"8770d0be9e8b909cad415eb93d05083834b1a876554a38bb2074585922bc9e8e","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Throw the die once onto the table. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["face_1","face_2","face_3","face_4","face_5","face_6"],"reference_distribution":{"face_1":0.16666666666666666,"face_2":0.16666666666666666,"face_3":0.16666666666666666,"face_4":0.16666666666666666,"face_5":0.16666666666666666,"face_6":0.16666666666666666},"reference_distribution_exact":[{"denominator":6,"label":"face_1","numerator":1},{"denominator":6,"label":"face_2","numerator":1},{"denominator":6,"label":"face_3","numerator":1},{"denominator":6,"label":"face_4","numerator":1},{"denominator":6,"label":"face_5","numerator":1},{"denominator":6,"label":"face_6","numerator":1}],"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/A02.png"}
+{"scene_id":"A-03","split":"calibration","split_order":2,"action":"Spin the wheel once under the fixed pointer.","source_image_path":"source_images/A03.png","source_image_bytes":8762767,"source_image_sha256":"fcb2e48564161f6827f46f7d9be8c0803fe14971abbaa9e58159ec2137c79a14","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Spin the wheel once under the fixed pointer. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["red","blue"],"reference_distribution":{"blue":0.5,"red":0.5},"reference_distribution_exact":[{"denominator":2,"label":"red","numerator":1},{"denominator":2,"label":"blue","numerator":1}],"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/A03.png"}
+{"scene_id":"A-04","split":"calibration","split_order":3,"action":"Spin the wheel once under the fixed pointer.","source_image_path":"source_images/A04.png","source_image_bytes":7880885,"source_image_sha256":"a158a705a4fa7b80ce780c35bb2d6ad11dee0ff565e02ec5da771aa572669bec","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Spin the wheel once under the fixed pointer. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["red","blue"],"reference_distribution":{"blue":0.75,"red":0.25},"reference_distribution_exact":[{"denominator":4,"label":"red","numerator":1},{"denominator":4,"label":"blue","numerator":3}],"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/A04.png"}
+{"scene_id":"A-05","split":"calibration","split_order":4,"action":"Release the ball once.","source_image_path":"source_images/A05.png","source_image_bytes":1904971,"source_image_sha256":"151c15377a14893cc1613467f6d86f9551f373568ebf5dbe29973b1b2432bd50","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Release the ball once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["left","right"],"reference_distribution":{"left":0.5,"right":0.5},"reference_distribution_exact":[{"denominator":2,"label":"left","numerator":1},{"denominator":2,"label":"right","numerator":1}],"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/A05.png"}
+{"scene_id":"A-06","split":"calibration","split_order":5,"action":"Release the ball once.","source_image_path":"source_images/A06.png","source_image_bytes":1999782,"source_image_sha256":"ebc27acae803b3c91b98d0a7b811c615308b50204e1dae2ac065e4099cd9b57b","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Release the ball once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["left","middle","right"],"reference_distribution":{"left":0.25,"middle":0.5,"right":0.25},"reference_distribution_exact":[{"denominator":4,"label":"left","numerator":1},{"denominator":2,"label":"middle","numerator":1},{"denominator":4,"label":"right","numerator":1}],"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/A06.png"}
+{"scene_id":"A-07","split":"calibration","split_order":6,"action":"Release the ball once.","source_image_path":"source_images/A07.png","source_image_bytes":7133726,"source_image_sha256":"9813c8a5aabbac9bacd49aaef17b90f4184f5bf03369910a7fd4bcfcf642d3a1","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Release the ball once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["left_path","right_path"],"reference_distribution":{"left_path":0.5,"right_path":0.5},"reference_distribution_exact":[{"denominator":2,"label":"left_path","numerator":1},{"denominator":2,"label":"right_path","numerator":1}],"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/A07.png"}
+{"scene_id":"A-08","split":"calibration","split_order":7,"action":"Release the ball once.","source_image_path":"source_images/A08.png","source_image_bytes":5956287,"source_image_sha256":"be3734ff1a4835235afff9cfbb7bc099545363189800b64204c9750fb99106f2","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Release the ball once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["left","right"],"reference_distribution":{"left":0.5,"right":0.5},"reference_distribution_exact":[{"denominator":2,"label":"left","numerator":1},{"denominator":2,"label":"right","numerator":1}],"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/A08.png"}
+{"scene_id":"A-09","split":"calibration","split_order":8,"action":"Move the hand upward once and let the pencil fall.","source_image_path":"source_images/A09.png","source_image_bytes":7145656,"source_image_sha256":"d4538997b6624277669b04958270edbc2b06a19cd47904aba7132a955bd386f0","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Move the hand upward once and let the pencil fall. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["falls_left","falls_right"],"reference_distribution":{"falls_left":0.5,"falls_right":0.5},"reference_distribution_exact":[{"denominator":2,"label":"falls_left","numerator":1},{"denominator":2,"label":"falls_right","numerator":1}],"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/A09.png"}
+{"scene_id":"A-10","split":"calibration","split_order":9,"action":"Flick both coins once simultaneously.","source_image_path":"source_images/A10.png","source_image_bytes":6661630,"source_image_sha256":"8fdc51b2c1043e9f4596a0b4bc16602cac1e9b08c4fd14b62a09111685d1ba85","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Flick both coins once simultaneously. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["0_heads","1_head","2_heads"],"reference_distribution":{"0_heads":0.25,"1_head":0.5,"2_heads":0.25},"reference_distribution_exact":[{"denominator":4,"label":"0_heads","numerator":1},{"denominator":2,"label":"1_head","numerator":1},{"denominator":4,"label":"2_heads","numerator":1}],"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/A10.png"}
+{"scene_id":"A-11","split":"calibration","split_order":10,"action":"Release the ball once.","source_image_path":"source_images/A11.png","source_image_bytes":5969075,"source_image_sha256":"6140065b06cfa568b27599824536aaa1d9e9c24eff3e8ba7d00a95682753817a","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Release the ball once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["left","right"],"reference_distribution":{"left":0,"right":1},"reference_distribution_exact":[{"denominator":1,"label":"left","numerator":0},{"denominator":1,"label":"right","numerator":1}],"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/A11.png"}
+{"scene_id":"A-12","split":"calibration","split_order":11,"action":"Let one small ball enter the inlet and pass through the split once.","source_image_path":"source_images/A12.png","source_image_bytes":6552414,"source_image_sha256":"b011b03e9c4e9ddc8ff6f49f9735ba92d7b6fb8aa81304dc3592a33b1130bf85","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Let one small ball enter the inlet and pass through the split once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["left_path","right_path"],"reference_distribution":{"left_path":0.5,"right_path":0.5},"reference_distribution_exact":[{"denominator":2,"label":"left_path","numerator":1},{"denominator":2,"label":"right_path","numerator":1}],"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/A12.png"}
+{"scene_id":"A-13","split":"calibration","split_order":12,"action":"The man picks up exactly one ball once.","source_image_path":"source_images/A13.png","source_image_bytes":6834824,"source_image_sha256":"314058baeced2ccc11cafeac1551b0de34e7ab9a907c234e5d6e6bb73eaee626","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. The man picks up exactly one ball once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["red","blue"],"reference_distribution":{"blue":0.5,"red":0.5},"reference_distribution_exact":[{"denominator":2,"label":"red","numerator":1},{"denominator":2,"label":"blue","numerator":1}],"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/A13.png"}
+{"scene_id":"A-14","split":"calibration","split_order":13,"action":"The man picks up exactly one ball once.","source_image_path":"source_images/A14.png","source_image_bytes":7722269,"source_image_sha256":"7640f0126c6e769ec085c5834e09a949d0a6b774476da1d6ca113b8771436916","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. The man picks up exactly one ball once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["red","blue"],"reference_distribution":{"blue":0.5,"red":0.5},"reference_distribution_exact":[{"denominator":2,"label":"red","numerator":1},{"denominator":2,"label":"blue","numerator":1}],"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/A14.png"}
+{"scene_id":"A-15","split":"calibration","split_order":14,"action":"Release the ball once.","source_image_path":"source_images/A15.png","source_image_bytes":1534095,"source_image_sha256":"baa2cf979da88ffe12b531b036a7c3d665327a514739acb1bbf32ed8905a7635","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Release the ball once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["left","right"],"reference_distribution":{"left":0.5,"right":0.5},"reference_distribution_exact":[{"denominator":2,"label":"left","numerator":1},{"denominator":2,"label":"right","numerator":1}],"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/A15.png"}
+{"scene_id":"A16","split":"calibration","split_order":15,"action":"Release the ball once.","source_image_path":"source_images/A16.png","source_image_bytes":2172379,"source_image_sha256":"ac50713c61441b1928f52d1fb54d4abc1cf2a71c69b9d3acb7a9adbc8f5e4c4f","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Release the ball once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["red_group","blue_group","green_group","yellow_group"],"reference_distribution":{"blue_group":0.25,"green_group":0.25,"red_group":0.25,"yellow_group":0.25},"reference_distribution_exact":[{"denominator":4,"label":"red_group","numerator":1},{"denominator":4,"label":"blue_group","numerator":1},{"denominator":4,"label":"green_group","numerator":1},{"denominator":4,"label":"yellow_group","numerator":1}],"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/A16.png"}
+{"scene_id":"A17","split":"calibration","split_order":16,"action":"Release the ball once.","source_image_path":"source_images/A17.png","source_image_bytes":1616881,"source_image_sha256":"18214f7f976b77d6bfa1d5e2c30ffce8fc515293d852f03ff5f87ae3f4f77457","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Release the ball once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["left_basin","right_basin"],"reference_distribution":{"left_basin":0.5,"right_basin":0.5},"reference_distribution_exact":[{"denominator":2,"label":"left_basin","numerator":1},{"denominator":2,"label":"right_basin","numerator":1}],"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/A17.png"}
+{"scene_id":"A18","split":"calibration","split_order":17,"action":"Spin the spinning top once.","source_image_path":"source_images/A18.png","source_image_bytes":1801846,"source_image_sha256":"82283f086f3bdbac314061cf423d7ef78c7c705231b07ccb21a2380b65a3f224","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Spin the spinning top once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["red_sector","blue_sector","green_sector","yellow_sector"],"reference_distribution":{"blue_sector":0.25,"green_sector":0.25,"red_sector":0.25,"yellow_sector":0.25},"reference_distribution_exact":[{"denominator":4,"label":"red_sector","numerator":1},{"denominator":4,"label":"blue_sector","numerator":1},{"denominator":4,"label":"green_sector","numerator":1},{"denominator":4,"label":"yellow_sector","numerator":1}],"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/A18.png"}
+{"scene_id":"A19","split":"calibration","split_order":18,"action":"Turn the side crank briefly and open the bottom gate once so exactly one ball exits.","source_image_path":"source_images/A19.png","source_image_bytes":2095456,"source_image_sha256":"c05317235c7ac6037caaff1427a4ae6268cd1726503c9db005b51334e4230c99","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Turn the side crank briefly and open the bottom gate once so exactly one ball exits. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["red","green","blue","yellow"],"reference_distribution":{"blue":0.25,"green":0.25,"red":0.25,"yellow":0.25},"reference_distribution_exact":[{"denominator":4,"label":"red","numerator":1},{"denominator":4,"label":"green","numerator":1},{"denominator":4,"label":"blue","numerator":1},{"denominator":4,"label":"yellow","numerator":1}],"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/A19.png"}
+{"scene_id":"A20","split":"calibration","split_order":19,"action":"Release the ball once.","source_image_path":"source_images/A20.png","source_image_bytes":1853281,"source_image_sha256":"25fc692aaf296a307eb016ebf5cb038afca7b09379ca55c4034a91b56041255a","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Release the ball once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["left","middle","right"],"reference_distribution":{"left":0.25,"middle":0.5,"right":0.25},"reference_distribution_exact":[{"denominator":4,"label":"left","numerator":1},{"denominator":2,"label":"middle","numerator":1},{"denominator":4,"label":"right","numerator":1}],"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/A20.png"}
+{"scene_id":"A21","split":"calibration","split_order":20,"action":"Roll both dice once.","source_image_path":"source_images/A21.png","source_image_bytes":1426889,"source_image_sha256":"c5ee1ec7869785a60dee0f8f654e5be16ecabc2ba4c0cd32622e436f0a653912","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Roll both dice once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["sum_2","sum_3","sum_4","sum_5","sum_6","sum_7","sum_8","sum_9","sum_10","sum_11","sum_12"],"reference_distribution":{"sum_10":0.08333333333333333,"sum_11":0.05555555555555555,"sum_12":0.027777777777777776,"sum_2":0.027777777777777776,"sum_3":0.05555555555555555,"sum_4":0.08333333333333333,"sum_5":0.1111111111111111,"sum_6":0.1388888888888889,"sum_7":0.16666666666666666,"sum_8":0.1388888888888889,"sum_9":0.1111111111111111},"reference_distribution_exact":[{"denominator":36,"label":"sum_2","numerator":1},{"denominator":36,"label":"sum_3","numerator":2},{"denominator":36,"label":"sum_4","numerator":3},{"denominator":36,"label":"sum_5","numerator":4},{"denominator":36,"label":"sum_6","numerator":5},{"denominator":36,"label":"sum_7","numerator":6},{"denominator":36,"label":"sum_8","numerator":5},{"denominator":36,"label":"sum_9","numerator":4},{"denominator":36,"label":"sum_10","numerator":3},{"denominator":36,"label":"sum_11","numerator":2},{"denominator":36,"label":"sum_12","numerator":1}],"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/A21.png"}
+{"scene_id":"A22","split":"calibration","split_order":21,"action":"Release the ball once.","source_image_path":"source_images/A22.png","source_image_bytes":1230741,"source_image_sha256":"067c1ca7d97a191c3c784297a8c79870d3e57f375a5d85386c8943035818c331","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Release the ball once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["bin_1","bin_2","bin_3","bin_4","bin_5"],"reference_distribution":{"bin_1":0.0625,"bin_2":0.25,"bin_3":0.375,"bin_4":0.25,"bin_5":0.0625},"reference_distribution_exact":[{"denominator":16,"label":"bin_1","numerator":1},{"denominator":16,"label":"bin_2","numerator":4},{"denominator":16,"label":"bin_3","numerator":6},{"denominator":16,"label":"bin_4","numerator":4},{"denominator":16,"label":"bin_5","numerator":1}],"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/A22.png"}
+{"scene_id":"A23","split":"calibration","split_order":22,"action":"Flick the coin once.","source_image_path":"source_images/A23.png","source_image_bytes":1506451,"source_image_sha256":"8e035c4dc8ccf4dd3bf5e6200ff9dcd0780670f6b01d960897a1358e5dbd295b","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Flick the coin once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["heads","tails"],"reference_distribution":{"heads":0.5,"tails":0.5},"reference_distribution_exact":[{"denominator":2,"label":"heads","numerator":1},{"denominator":2,"label":"tails","numerator":1}],"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/A23.png"}
+{"scene_id":"A24","split":"calibration","split_order":23,"action":"Turn the side crank briefly and open the bottom gate once so exactly one ball exits.","source_image_path":"source_images/A24.png","source_image_bytes":2052344,"source_image_sha256":"2ad0148a440e5a084f1935f42da1856c1f73a0580bddec42e46cb16f19c2972e","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Turn the side crank briefly and open the bottom gate once so exactly one ball exits. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["red","green","blue","yellow"],"reference_distribution":{"blue":0.25,"green":0.25,"red":0.25,"yellow":0.25},"reference_distribution_exact":[{"denominator":4,"label":"red","numerator":1},{"denominator":4,"label":"green","numerator":1},{"denominator":4,"label":"blue","numerator":1},{"denominator":4,"label":"yellow","numerator":1}],"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/A24.png"}
+{"scene_id":"A25","split":"calibration","split_order":24,"action":"Move the hand upward once and let the pencil fall.","source_image_path":"source_images/A25.png","source_image_bytes":1283840,"source_image_sha256":"c9ea6c92ae96812370e7e046bdb5f649d86e7bcb984da4b85a8ecbdbc6f8f603","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Move the hand upward once and let the pencil fall. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["falls_left","falls_right"],"reference_distribution":{"falls_left":1,"falls_right":0},"reference_distribution_exact":[{"denominator":1,"label":"falls_left","numerator":1},{"denominator":1,"label":"falls_right","numerator":0}],"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/A25.png"}
+{"scene_id":"BA-01-02","split":"coverage","split_order":0,"action":"Gently touch the cat once.","source_image_path":"source_images/BA-01-02.png","source_image_bytes":8927834,"source_image_sha256":"3ff018f2e150747fb8b9d398ea22a7a390afbfaaf62d292b600476d9da7c6987","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Gently touch the cat once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["cat_accepts_touch_or_leans_in","cat_remains_still_near_hand","cat_withdraws_but_stays_in_area","cat_leaves_interaction_area","cat_defensive_posture_without_contact","cat_active_defensive_contact"],"reference_distribution":null,"reference_distribution_exact":null,"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/BA-01-02.png"}
+{"scene_id":"BA-01-S1","split":"coverage","split_order":1,"action":"Gently touch the cat once.","source_image_path":"source_images/BA-01-S1.png","source_image_bytes":8256551,"source_image_sha256":"ad5cbbde37b083432de986712537a81009840553f64f5201207c44b9b6ed9209","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Gently touch the cat once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["cat_accepts_touch_or_leans_in","cat_remains_still_near_hand","cat_withdraws_but_stays_in_area","cat_leaves_interaction_area","cat_defensive_posture_without_contact","cat_active_defensive_contact"],"reference_distribution":null,"reference_distribution_exact":null,"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/BA-01-S1.png"}
+{"scene_id":"BA-01","split":"coverage","split_order":2,"action":"Gently touch the cat once.","source_image_path":"source_images/BA-01.png","source_image_bytes":8463456,"source_image_sha256":"0f9c13707e321a6803e0af54b8d0a7cbee8d661771e25a02d5b2d81ba31d7348","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Gently touch the cat once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["cat_accepts_touch_or_leans_in","cat_remains_still_near_hand","cat_withdraws_but_stays_in_area","cat_leaves_interaction_area","cat_defensive_posture_without_contact","cat_active_defensive_contact"],"reference_distribution":null,"reference_distribution_exact":null,"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/BA-01.png"}
+{"scene_id":"BA-02","split":"coverage","split_order":3,"action":"Let the stranger approach the dog.","source_image_path":"source_images/BA-02.png","source_image_bytes":9453620,"source_image_sha256":"d0d77d10c487b54fa4146a0f131d138ce4edd54a6384f63193cf82bb4146ed31","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Let the stranger approach the dog. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["dog_approaches_person","dog_holds_position_facing_person","dog_turns_away_without_retreating","dog_retreats_or_avoids_person","dog_leaves_interaction_area","dog_shows_defensive_posture","dog_active_defensive_advance_or_contact"],"reference_distribution":null,"reference_distribution_exact":null,"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/BA-02.png"}
+{"scene_id":"BA-03","split":"coverage","split_order":4,"action":"Let the dog make one reach attempt toward the toy.","source_image_path":"source_images/BA-03.png","source_image_bytes":7851132,"source_image_sha256":"56c0c49da81fc7dfd677a1112ab8c4dd791c1a6b0ff1dedef5a29b93bebde9d2","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Let the dog make one reach attempt toward the toy. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["goes_around_left","goes_around_right","jumps_over_middle","crawls_or_ducks_under_middle","does_not_attempt"],"reference_distribution":null,"reference_distribution_exact":null,"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/BA-03.png"}
+{"scene_id":"BA-06","split":"coverage","split_order":5,"action":"Let the cat reach or step toward the tabletop objects once from the current poised posture.","source_image_path":"source_images/BA-06.png","source_image_bytes":8623523,"source_image_sha256":"42517eec224621640959012f716d08bdad112f768b3a9e59016ce224e4f7308c","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Let the cat reach or step toward the tabletop objects once from the current poised posture. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["cat_does_not_enter_tabletop_area","cat_enters_tabletop_area_without_object_contact","cat_touches_object_without_displacing","cat_displaces_non_cup_object","cat_displaces_cup_without_spill","cat_spills_or_tips_cup"],"reference_distribution":null,"reference_distribution_exact":null,"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/BA-06.png"}
+{"scene_id":"BA-07","split":"coverage","split_order":6,"action":"Gently pull the leash once while the dog is eating.","source_image_path":"source_images/BA-07.png","source_image_bytes":8518227,"source_image_sha256":"09ce62a9361b0325101b15233ecd39035b4499d002604367efb9508324409277","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Gently pull the leash once while the dog is eating. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["dog_continues_eating_at_bowl","dog_pauses_or_lifts_head_near_bowl","dog_resists_taut_leash_stays_near_bowl","dog_steps_back_from_bowl","bowl_shifts_without_food_spill","food_spills_or_bowl_tips"],"reference_distribution":null,"reference_distribution_exact":null,"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/BA-07.png"}
+{"scene_id":"BA-08","split":"coverage","split_order":7,"action":"An off-screen hand enters the frame once and offers food to the bird.","source_image_path":"source_images/BA-08.png","source_image_bytes":7264728,"source_image_sha256":"5f184fe443ff010c492ce091d59815ad5d82b071cf5f707d6671a58bfd068b69","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. An off-screen hand enters the frame once and offers food to the bird. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["bird_takes_food_from_hand","bird_flies_off_or_leaves_table","bird_moves_away_but_stays_on_table","bird_displaces_food_without_taking","bird_pecks_food_without_taking","bird_stays_near_without_food_contact"],"reference_distribution":null,"reference_distribution_exact":null,"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/BA-08.png"}
+{"scene_id":"BC-01-I2","split":"coverage","split_order":8,"action":"Toss the ball once from the visible hand pose toward the cup.","source_image_path":"source_images/BC-01-I2.png","source_image_bytes":5553517,"source_image_sha256":"a26c6f6d341a6f1edaa78fe2fdb8e7fd988bccc4e7eec5b1e5c6080a0e78e6db","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Toss the ball once from the visible hand pose toward the cup. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["clean_in_cup","contact_then_in_cup","cup_contact_then_out","near_miss_no_cup_contact","clear_miss"],"reference_distribution":null,"reference_distribution_exact":null,"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/BC-01-I2.png"}
+{"scene_id":"BC-01","split":"coverage","split_order":9,"action":"Toss the ball once from the visible hand pose toward the cup.","source_image_path":"source_images/BC-01.png","source_image_bytes":1614593,"source_image_sha256":"b9349e2df01bafcd1d4f98d9252dcded4880d62314d8e0a6be25f3262d3353c9","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Toss the ball once from the visible hand pose toward the cup. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["clean_in_cup","contact_then_in_cup","cup_contact_then_out","near_miss_no_cup_contact","clear_miss"],"reference_distribution":null,"reference_distribution_exact":null,"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/BC-01.png"}
+{"scene_id":"BC-02","split":"coverage","split_order":10,"action":"Shoot the basketball once from the visible shooting pose toward the hoop.","source_image_path":"source_images/BC-02.jpg","source_image_bytes":224466,"source_image_sha256":"20ee53d95dbb0f24286537da8e8814f11a1596ff6daedc545c04d1f523d696a2","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Shoot the basketball once from the visible shooting pose toward the hoop. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["swish","backboard_then_in","rim_then_in","rim_miss_no_score","backboard_miss_no_score","left_miss_no_contact","right_miss_no_contact","short_airball_no_contact","long_airball_no_contact"],"reference_distribution":null,"reference_distribution_exact":null,"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/BC-02.jpg"}
+{"scene_id":"BC-03-I1","split":"coverage","split_order":11,"action":"Roll the bowling ball once down the lane.","source_image_path":"source_images/BC-03-I1.png","source_image_bytes":9563395,"source_image_sha256":"2fcf9df4538b817587d6c08593d9564845985515940b07421d0d6c87d21cba34","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Roll the bowling ball once down the lane. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["all_7_pins_down","six_pins_down","five_pins_down","four_pins_down","three_pins_down","two_pins_down","one_pin_down","zero_pins_down_after_contact","no_pin_contact"],"reference_distribution":null,"reference_distribution_exact":null,"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/BC-03-I1.png"}
+{"scene_id":"BC-03","split":"coverage","split_order":12,"action":"Roll the bowling ball once down the lane.","source_image_path":"source_images/BC-03.png","source_image_bytes":1712335,"source_image_sha256":"b46f2eed5a44c52d5598ae9050263d41eb75ada92e26d4742f73689bffd8d06d","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Roll the bowling ball once down the lane. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["all_7_pins_down","six_pins_down","five_pins_down","four_pins_down","three_pins_down","two_pins_down","one_pin_down","zero_pins_down_after_contact","no_pin_contact"],"reference_distribution":null,"reference_distribution_exact":null,"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/BC-03.png"}
+{"scene_id":"BC-04","split":"coverage","split_order":13,"action":"Kick the ball once toward the goal.","source_image_path":"source_images/BC-04.jpg","source_image_bytes":258764,"source_image_sha256":"1848f2b2a827f5104a1ee8b7be5ac57cdda440ca2888c135aaa4f176ed47da1b","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Kick the ball once toward the goal. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["clean_goal","keeper_deflected_goal","frame_deflected_goal","goalkeeper_prevents_goal","post_out","crossbar_out","miss_left","miss_right","miss_high","shot_stops_short"],"reference_distribution":null,"reference_distribution_exact":null,"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/BC-04.jpg"}
+{"scene_id":"BC-05","split":"coverage","split_order":14,"action":"Toss the ring once from the visible hand pose toward the peg.","source_image_path":"source_images/BC-05.jpg","source_image_bytes":113706,"source_image_sha256":"56e13f961cb6822715a4989b16584b0dce6ed82443cde1e5fbcef08aa9b3243d","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Toss the ring once from the visible hand pose toward the peg. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["settles_around_peg","rests_on_peg_not_scored","contacts_peg_but_off","near_miss_no_peg_contact","clear_miss"],"reference_distribution":null,"reference_distribution_exact":null,"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/BC-05.jpg"}
+{"scene_id":"BC-06","split":"coverage","split_order":15,"action":"Throw the dart once from the visible hand pose toward the board.","source_image_path":"source_images/BC-06.jpg","source_image_bytes":107785,"source_image_sha256":"0e2a2c2f95f08048756f6fd012609f575030a04f072ec49b2598280a634352ad","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Throw the dart once from the visible hand pose toward the board. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["bullseye","outer_bull","inner_single_hit","triple_ring_hit","outer_single_hit","double_ring_hit","board_bounce_out","off_board_miss"],"reference_distribution":null,"reference_distribution_exact":null,"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/BC-06.jpg"}
+{"scene_id":"BC-09","split":"coverage","split_order":16,"action":"Release the ball once into the board.","source_image_path":"source_images/BC-09.png","source_image_bytes":7951788,"source_image_sha256":"6b2d8eb1170eaf3d348674e2eb8eaccff24c359623704ef2a5389d92206db3cf","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Release the ball once into the board. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["left_bin","middle_bin","right_bin","stuck_on_board"],"reference_distribution":null,"reference_distribution_exact":null,"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/BC-09.png"}
+{"scene_id":"BM-01","split":"coverage","split_order":17,"action":"Pull the loose yarn end once.","source_image_path":"source_images/BM-01.png","source_image_bytes":2014281,"source_image_sha256":"380d36ee92e7767e663f485a2e6e8161544cdf2a2bd63c0c19511dde4937845b","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Pull the loose yarn end once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["thread_pulls_free_cleanly","small_unravel","large_unravel_or_ladder","thread_snaps","thread_stays_attached_no_visible_change"],"reference_distribution":null,"reference_distribution_exact":null,"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/BM-01.png"}
+{"scene_id":"BM-02","split":"coverage","split_order":18,"action":"Bend the chocolate bar once until it responds.","source_image_path":"source_images/BM-02.png","source_image_bytes":1838633,"source_image_sha256":"3ca59f910af6f8236aa79d5efe5ca6f0ed300c60c60a90dda85575a1cf1442f0","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Bend the chocolate bar once until it responds. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["no_visible_break","single_clean_break","partial_crack_not_separated","multiple_piece_break","crumbs_or_small_fragment_break"],"reference_distribution":null,"reference_distribution_exact":null,"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/BM-02.png"}
+{"scene_id":"BM-03","split":"coverage","split_order":19,"action":"Flip the egg once with the spatula.","source_image_path":"source_images/BM-03.png","source_image_bytes":1959317,"source_image_sha256":"66ff8e244c9763df9fe4800fda01f2d73569f84f9398e6d1bf423250aa43199b","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Flip the egg once with the spatula. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["lands_yolk_up_intact","lands_yolk_down_intact","folds_without_tearing","white_tears_yolk_intact","yolk_ruptures","spatula_contact_no_flip"],"reference_distribution":null,"reference_distribution_exact":null,"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/BM-03.png"}
+{"scene_id":"BM-04","split":"coverage","split_order":20,"action":"Peel the tape from the cardboard once.","source_image_path":"source_images/BM-04.png","source_image_bytes":2114210,"source_image_sha256":"7f16cd4a1d7cdd0120032d6599f6305465a64af415cef2a88c9dbcc425245ccd","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Peel the tape from the cardboard once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["clean_full_peel","partial_peel_tape_still_attached","tape_tears","cardboard_surface_tears","tape_curls_or_reattaches_without_full_removal"],"reference_distribution":null,"reference_distribution_exact":null,"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/BM-04.png"}
+{"scene_id":"BS-01-I1","split":"coverage","split_order":21,"action":"Flip the bottle once.","source_image_path":"source_images/BS-01-I1.png","source_image_bytes":1995747,"source_image_sha256":"fa40e8da0a48610fc8f7dca1bbbc97de2b3adc0dd49e704ac3f52e64ccee3047","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Flip the bottle once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["upright_on_base","balanced_on_cap","side_laying","leaning_against_visible_support"],"reference_distribution":null,"reference_distribution_exact":null,"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/BS-01-I1.png"}
+{"scene_id":"BS-01","split":"coverage","split_order":22,"action":"Flip the bottle once.","source_image_path":"source_images/BS-01.png","source_image_bytes":8312319,"source_image_sha256":"b74225d20122d5107a94be1557cf76ee85f315b8e26b37433ceef4995e9b6b6d","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Flip the bottle once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["upright_on_base","balanced_on_cap","side_laying","leaning_against_visible_support"],"reference_distribution":null,"reference_distribution_exact":null,"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/BS-01.png"}
+{"scene_id":"BS-02","split":"coverage","split_order":23,"action":"Pull the red-marked block once from the tower.","source_image_path":"source_images/BS-02.png","source_image_bytes":1924076,"source_image_sha256":"1bea243bd6120e9030da4931d942e9a7e307b9db6e697ddc0491532ccdfdb598","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Pull the red-marked block once from the tower. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["tower_remains_standing_after_removal","tower_remains_standing_but_displaced","top_section_collapse","side_section_collapse","full_collapse"],"reference_distribution":null,"reference_distribution_exact":null,"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/BS-02.png"}
+{"scene_id":"BS-03","split":"coverage","split_order":24,"action":"Pull the tablecloth out once.","source_image_path":"source_images/BS-03.png","source_image_bytes":1786714,"source_image_sha256":"8e0193980c95b5d0607ac56193a382f9b911fae9fa998ba0822cbcc598ba690f","base_prompt":"Using the provided image as the first frame, create one continuous real-time video. Use a locked-off stable camera. Pull the tablecloth out once. Keep the full interaction area and all key objects visible throughout. Preserve the visual style, lighting, object identities, and layout of the input image. By the final moment, the scene is stable enough that the final positions, spatial relations, and object identities are clearly readable.","outcome_labels":["objects_remain_in_place","objects_shift_but_upright","objects_scatter_on_table","one_object_falls_on_table","multiple_objects_fall_on_table","one_object_leaves_table","multiple_objects_leave_table"],"reference_distribution":null,"reference_distribution_exact":null,"upstream_source_uri":"hf://Andrew613/PAWBench-Results/benchmark/PAWBench_V2/source_images/BS-03.png"}