evals/prompts/cot-subq.yaml

scheme: cot_subq
description: |-
  Structured sub-question rationale prompts for Claude distillation.
  Claude answers fixed sub-questions with visible evidence, then outputs a JSON score.
sub_questions:
  source: static
  answer_format: template
system_prompt: |-
  You are a strict video evaluation model.
  Base your judgment only on visible evidence in the video.
  Provide a concise rationale, then output a JSON object with the score.
general_keys:
- SA
- PTV
- persistence
eval_prompts:
  SA: |-
    Evaluate Prompt Alignment (SA).

    Caption:
    "{prompt}"

    The video was generated using a text+image-to-video (ti2v) model, conditioned on the first frame and the text prompt above.

    For each sub-question, briefly state the observed evidence, then give exactly one answer label from: yes, partial, no, uncertain.
    Use this format for each sub-question:
    qN: <brief evidence>; answer=<label>

    {questions_block}

    Score 1-5:
    5=fully aligned
    4=mostly aligned with minor deviations
    3=partially aligned with notable gaps
    2=mostly misaligned
    1=not aligned

    Answer every sub-question in the format above, then justify the overall score.
    Then output a JSON object with keys "reasoning" (string) and "SA" (integer 1-5).
    Output JSON only.

    Example:
    {{"reasoning": "q1: water balloon, target, and thrower all present; answer=yes. q2: balloon is thrown and bursts on impact; answer=yes. q3: outdoor setting and distance preserved; answer=yes. q4: target behaves like an inflatable rather than cardboard; answer=partial. Overall the scene matches well with one material mismatch.", "SA": 4}}
  PTV: |-
    Evaluate Temporal Coherence (PTV).

    Caption:
    "{prompt}"

    The video was generated using a text+image-to-video (ti2v) model, conditioned on the first frame and the text prompt above.

    For each sub-question, briefly state the observed evidence, then give exactly one answer label from: yes, partial, no, uncertain.
    Use this format for each sub-question:
    qN: <brief evidence>; answer=<label>

    {questions_block}

    Score 1-5:
    5=fully plausible event order
    4=mostly plausible with minor timing issues
    3=partially plausible
    2=mostly implausible
    1=completely implausible order

    Answer every sub-question in the format above, then justify the overall score.
    Then output a JSON object with keys "reasoning" (string) and "PTV" (integer 1-5).
    Output JSON only.

    Example:
    {{"reasoning": "q1: bottle shatters before any object contacts it, effect precedes cause; answer=no. q2: rupture occurs without visible force, not a plausible physical event order; answer=no. q3: fragments appear instantly rather than progressing from impact point; answer=no. q4: no repeated resets, but the spontaneous break is an impossible state change; answer=partial. Temporal sequence is highly implausible.", "PTV": 1}}
  persistence: |-
    Evaluate Object Persistence.

    Caption, for context only:
    "{prompt}"

    The video was generated using a text+image-to-video (ti2v) model, conditioned on the first frame and the text prompt above.

    For each sub-question, briefly state the observed evidence, then give exactly one answer label from: yes, partial, no, uncertain.
    Use this format for each sub-question:
    qN: <brief evidence>; answer=<label>

    {questions_block}

    Score 1-5:
    5=fully consistent
    4=mostly consistent with minor flicker
    3=noticeable issues
    2=major inconsistencies
    1=severe disappearance or identity changes

    Answer every sub-question in the format above, then justify the overall score.
    Then output a JSON object with keys "reasoning" (string) and "persistence" (integer 1-5).
    Output JSON only.

    Example:
    {{"reasoning": "q1: tire and ground remain present throughout; answer=yes. q2: tire and ground keep stable color and texture, but bottle label text changes mid-video; answer=partial. q3: no objects disappear or appear unexpectedly; answer=yes. q4: tire identity stable through motion, but bottle label shifts; answer=partial. Minor but noticeable label inconsistency.", "persistence": 3}}
physical_sub_questions: true
physical_template: |-
  Evaluate physical realism for one physical law: {law}.

  Criterion:
  {criteria}

  Caption, for context only:
  "{prompt}"

  For each sub-question, briefly state the observed evidence, then give exactly one answer label from: yes, no, uncertain, na.
  Use this format for each sub-question:
  qN: <brief evidence>; answer=<label>

  {questions_block}

  Judge the video itself. Do not penalize prompt mismatch unless it affects whether this physical law can be evaluated.

  Score 1-5:
  5=clearly correct
  4=mostly correct with minor issues
  3=partially correct or ambiguous
  2=mostly incorrect
  1=severely incorrect

  Answer every sub-question in the format above, then justify the overall score.
  Then output a JSON object with keys "reasoning" (string) and "{law}" (integer 1-5).
  Output JSON only.

  Example:
  {{"reasoning": "q1: baseball stays in place after clear bat contact, completely unaffected; answer=yes. q2: brown chunk appears on ball surface, response wildly disproportionate to impact; answer=yes. q3: bat morphs and clips through ball, but no shattering from light touch; answer=no. Severely incorrect physical behavior.", "{law}": 1}}