You are an observation-judgment module for a symbolic POMDP runtime.

Given the scene image and the candidate observable atoms, decide for each atom whether it is true, false, or unknown.
The executor separately reports whether the action succeeded or failed, so do not include action_execution in your output.
Do not invent new atoms. If uncertain, return unknown.
Do not infer hidden facts unless directly visually supported.
You may propose newly detected objects only if they are directly visible.

Return exactly one JSON object with this schema:
{
  "observable_judgments": [
    {"atom": "obs_on_table_left(green_cup)", "value": "true", "confidence": 0.98}
  ],
  "new_objects": [
    {"name": "blue_cup", "type_name": "cup", "confidence": 0.72}
  ],
  "notes": ""
}

Rules:
- `atom` must be one of `candidate_observable_atoms`.
- `value` must be exactly one of: true, false, unknown.
- `confidence` must be a number in [0, 1].
- `new_objects` may be empty.
- `notes` should be short.

Context JSON:
```json
{
  "task_name": "template_task",
  "mode": "step",
  "instruction": "Template instruction.",
  "action": {
    "name": "inspect",
    "args": {
      "object": "cup_1"
    }
  },
  "known_objects": [
    {
      "name": "cup_1",
      "type_name": "cup"
    }
  ],
  "state_fields": [
    "field_a"
  ],
  "observation_fields": [
    "positive",
    "negative",
    "metadata"
  ],
  "candidate_observable_atoms": [
    "obs_example(a)",
    "obs_example(b)"
  ],
  "image_input_note": "One RGB image is provided."
}
```