Edit Step4_dp_refine Template

Step4_dp_refine v4

Viewer

Test Case:

Template

Preview

History

System prompt (optional) User prompt (template)

You are synthesizing decision points for NSPE ethics case {{ case_id }}.

## TOP ALGORITHMIC CANDIDATES (Scored by Q&C Alignment)

These candidates were composed algorithmically from extracted entities and scored against the board's actual questions and conclusions:

{{ candidates_block }}

## BOARD'S QUESTIONS (with Toulmin Analysis)

These are the actual ethical questions with their Toulmin structure:

{{ questions_block }}

## BOARD'S CONCLUSIONS (with Resolution Patterns)

{{ conclusions_block }}

## TASK

Synthesize {{ target_count }} decision points that:

1. **Preserve entity grounding** - Keep URI references from candidates
2. **Align with Q&C** - Each point should address real board concerns
3. **Merge similar candidates** - Combine candidates addressing the same issue
4. **Include the full Toulmin structure** - all six categories, each respecting
   its definition:
{{ toulmin_field_spec }}
5. **Coherent question-option structure** - Options must directly answer the question

CRITICAL FORMATTING REQUIREMENT:

- Do NOT use em dash characters anywhere in your output. Use commas or periods instead.

CRITICAL CONTINGENCY REQUIREMENT:

- Do NOT construct a decision point whose premise is a contingency the case text does not
  discuss: a party's future or hypothetical refusal, rejection, non-compliance, or any
  other event the facts, discussion, and conclusion never reach. A decision point may
  explore an alternative OPEN TO THE ACTOR at a moment the case narrates; it may NOT posit
  an un-discussed downstream scenario (e.g. "if the attorney then declines to correct the
  record, should the engineer withdraw?") and resolve duties for it.

CRITICAL ROLE LABEL REQUIREMENT:

- The role_label must be a SHORT identifier for the decision-maker: "Engineer A", "Firm C",
  "Engineers A and B", "Client", etc. Do NOT append obligation names, case descriptions,
  or topic keywords to the role_label.
  - BAD: "Engineer A Construction Observation Engineer", "Engineer H Public Hearing Testimony Completeness ZZZ Truck Stop"
  - GOOD: "Engineer A", "Engineer H"
- The role_label must be the agent who faces the decision. Do not assign a decision to a party
  who is not making the choice (e.g., do not assign a disclosure decision to the "Client"
  when it is the engineer who must decide whether to disclose).

CRITICAL COHERENCE REQUIREMENT:

The decision_question, description, and options must form a coherent decision structure:

1. The "decision_question" must be framed as an actionable choice the named role faces:
   - Format: "Should [role] [action A] or [action B]?" or "Must [role] [choice]?"
   - The question must present the core tension between competing courses of action.
   - Keep it to 1-2 sentences. Do not embed long subordinate clauses.
   - BAD: "Whether the obligation arose at point X or point Y" (analytical, not actionable)
   - BAD: "The interaction between principle X and principle Y" (abstract, no agent choosing)
   - GOOD: "Should Engineer A disclose the conflict to the client before accepting the project, or rely on internal firewalls?"

2. Each option must be a DIRECT ANSWER to the decision_question. If you read the question
   then read each option, the option must be a plausible response the named role could choose.
   - BAD: Question asks "when did the obligation arise?" but options are "Submit Report" / "Limit Disclosure"
   - GOOD: Question asks "Should Doe submit full findings or limit disclosure?" and options are
     "Submit Full Report" / "Limit Disclosure to Correcting False Data" / "Seek Ethics Guidance First"

CRITICAL OPTION REQUIREMENTS:

1. Each option MUST have a short "label" (3-8 words, Title Case, action phrase starting with a verb)
   and a longer "description" (1-2 sentences elaborating the action).
   The label is a DISCRETE CHOICE that a decision-maker selects from a list.
   - Good labels: "Disclose Conflict to Client", "Recuse from Evaluation", "Report to State Agency"
   - Bad labels: "Option A", "Proactively disclose AI tool usage and identify AI-generated sections to client before submission"
   The label must be distinct enough to distinguish options at a glance.

2. Each decision point MUST have 2-3 options that represent GENUINELY DEFENSIBLE positions.
   Do NOT create straw-man alternatives. Each option should be an action a reasonable
   professional could plausibly choose given competing pressures (time, cost, client
   relationship, scope of duty, professional judgment).

   - BAD (straw-man negation):
     O1: "Conduct rigorous line-by-line review before sealing"
     O2: "Seal without any verification"
   - GOOD (genuine tension):
     O1: "Conduct full independent technical review of all AI outputs before sealing"
     O2: "Apply standard firm QA protocols to AI outputs at the same level as conventional CAD software"
     O3: "Engage a third-party reviewer with AI expertise for safety-critical elements while applying standard review to remaining outputs"

   The non-board-choice options must represent positions with real justifications
   (efficiency, precedent, scope limitation, competing obligations) -- not simply
   omitting or refusing to perform the ethical action.

## OUTPUT FORMAT (JSON)

```json
[
  {
    "focus_id": "DP1",
    "source_candidate_ids": ["DP1", "DP3"],
    "description": "Clear description",
    "decision_question": "The key ethical question",
    "role_label": "Engineer A",
    "role_uri": "URI from candidate",
    "obligation_label": "From candidate",
    "obligation_uri": "URI",
    "constraint_label": null,
    "constraint_uri": null,
    "provision_labels": ["II.1.c"],
    "provision_uris": ["URIs"],
    "involved_action_uris": ["action URIs"],
    "toulmin_claim": "The course of action argued for",
    "toulmin_data": "The case facts appealed to (facts only)",
    "toulmin_warrants": "The general rules licensing facts to claim",
    "toulmin_qualifier": "Modal strength or attached conditions ('' if unconditional)",
    "toulmin_rebuttals": "The conditions of exception, 'would not apply if ...'",
    "addresses_questions": ["Q0", "Q1"],
    "board_resolution": "How board resolved this",
    "qc_alignment_score": 0.85,
    "intensity_score": 0.7,
    "options": [
      {"option_id": "O1", "label": "Disclose AI Usage to Client Before Submission", "description": "Proactively disclose AI tool usage and identify AI-generated sections to client before submission", "action_uri": "URI", "is_board_choice": true},
      {"option_id": "O2", "label": "Treat AI as Internal Drafting Tool", "description": "Treat AI tool as internal drafting software equivalent to CAD, disclosing only upon direct client inquiry", "action_uri": "URI", "is_board_choice": false},
      {"option_id": "O3", "label": "Disclose in Project Documentation Only", "description": "Disclose AI usage in project documentation without separate client notification, following existing firm software disclosure policy", "action_uri": "URI", "is_board_choice": false}
    ]
  }
]
```

Produce exactly {{ target_count }} decision points capturing the key ethical issues. Do NOT produce more than the requested count.

Loading preview...

Version History

No version history yet. Save changes to create versions.

Saved

7166 chars

Reference

Template Variables

Requested decision-point count for this batch, e.g. '4-6' or '2-3'; computed per batch of 5 in _llm_refine

Toulmin (1958) six-category field specification shared by all decision-point prompts

Q0-indexed questions with URI and Toulmin annotations (DATA/WARRANTS/REBUTTAL) joined from question_emergence via qc_refs.key_aliases normalization

Numeric case id, interpolated into the task header

Per-candidate blocks: focus_id with Q&C alignment score, description, question, role/obligation labels with URIs, matched questions, options with chosen/alternative tags

C0-indexed conclusions with URI, determinative principles, and resolution narrative joined from resolution_patterns via qc_refs.key_aliases normalization

Output Schema

type array

items

JSON Wrapper

Appended to the prompt by the extractor:

IMPORTANT: Wrap your response as a JSON object with exactly two keys:
{"new_step4_dp_refine_classes": [...], "step4_dp_refine_individuals": [...]}
If there are no individuals to report, use an empty array for "step4_dp_refine_individuals".