Video-Conditioned Persona Key Filtering

You produce a compact, OBJECTIVE summary of an egocentric video for a
downstream user-deviation data system. The camera wearer is the user.

Describe only what is observable or reasonably inferable about the TASK. Do NOT introduce persona
attributes (no personality, identity, goals-beyond-the-task, demographics). Do NOT describe the
wearer's appearance. Return valid JSON only.

Summarize this egocentric video as JSON with these fields:
- task: short task name
- task_goal: the successful end state
- observed_actions: list of {{action_id, description, time_range}}
- objects_and_tools: list of strings
- expected_or_inferred_steps: the reference/expected step sequence
- decision_points: moments where the user must decide something
- state_judgments: perceptual judgments the user must make (e.g. "judge doneness")
- hazards_and_constraints: safety / correctness constraints
- possible_deviation_opportunities: list of {{opportunity_id (like "d01"), description, related_actions}}
  — concrete places where a user could plausibly deviate from the expected trajectory
- uncertainties: things not visible/known

{context_block}

Keep it factual. Do not invent persona attributes. Return ONLY the JSON object.

You are evaluating persona keys for a user-deviation data generation system.

The input is an egocentric video (the camera wearer is the user), given to you as an objective task
summary. Your task is NOT to infer the user's actual persona. Your task is to decide whether each
provided persona key could be a useful CONDITIONING DIMENSION for generating plausible deviations from
the expected task trajectory in THIS particular video.

A key may be useful even when its value is not visible. Select it only if different plausible VALUES of
the key could cause meaningfully different deviation tendencies through a CONCRETE opportunity in the
video.

Do not assign values. Do not generate a persona. Do not generate deviations, dialogue, or questions.
Do not infer sensitive attributes (health, disability, ethnicity, religion, nationality, age) from
appearance — such keys may be selected only as hypothetical conditioning dimensions, never as claims
about the real wearer. Evaluate EVERY provided key and return valid JSON only.

VIDEO SUMMARY
{video_summary_json}

PERSONA KEYS TO EVALUATE (evaluate every one; use the exact key id)
{key_batch_json}

For each key:
1. Apply the counterfactual test: if two otherwise similar users differed only in this key, could they
   plausibly deviate from the expected task trajectory in meaningfully different ways in THIS video?
2. Assign relevance_score 0-3:
   0 = no credible connection to user deviation here
   1 = weak/indirect; mainly narrative or dialogue, not deviation
   2 = plausible, video-specific connection to at least one deviation opportunity
   3 = strong connection to multiple or high-impact deviation opportunities
3. selected = (relevance_score >= 2).
4. Give a concise VIDEO-SPECIFIC causal reason.
5. related_opportunity_ids: link selected keys to opportunity ids from the summary.
6. reason_codes for SELECTED keys (>=1), from: STEP_OMISSION, ACTION_SELECTION, ACTION_ORDER,
   OBJECT_OR_TOOL_SELECTION, QUANTITY_OR_PARAMETER_CHOICE, TIMING_OR_PACING, STATE_JUDGMENT,
   ATTENTION_OR_MONITORING, ERROR_RECOVERY, HELP_SEEKING, RISK_OR_SAFETY, GOAL_PRIORITY,
   DEPENDENT_OR_OTHER_PERSON_NEEDS, HEALTH_OR_ACCESSIBILITY_ADAPTATION, TASK_PERSISTENCE_OR_ABANDONMENT.
7. rejection_code for NON-selected keys, one of: NO_VIDEO_OPPORTUNITY, DIALOGUE_ONLY, NARRATIVE_ONLY,
   REDUNDANT_WITH_MORE_DIRECT_KEY, TOO_INDIRECT, STEREOTYPE_DEPENDENT, CONTRADICTS_VIDEO,
   INSUFFICIENT_CAUSAL_MECHANISM.

Do not assign any persona value. Do not claim the actual wearer has the attribute.
Return one decision for EVERY supplied key. Return ONLY the JSON object.

You are a strict reviewer of persona-key selections for user-deviation generation.
You remove keys with weak causal links, resolve redundancy, and check that every reason is
video-specific and that sensitive keys are framed as hypothetical conditioning dimensions (never
appearance-based inference). You must NOT add values or generate deviations. Return valid JSON only.

VIDEO SUMMARY
{video_summary_json}

INITIALLY SELECTED KEYS (with scores and reasons)
{selected_json}

For each key decide keep=true/false:
- remove if the causal link to a concrete deviation opportunity is weak or not video-specific;
- remove if it is redundant with a more direct selected key (say which in note);
- remove if a sensitive key is justified by appearance/stereotype rather than a task mechanism;
- keep keys that introduce a DISTINCT deviation mechanism.
Give a one-line note per key. Return ONLY JSON: {{"reviews":[{{"key","keep","note"}}]}}.