Historical Atrocity Education Content Allowance Prompt

This prompt is designed for educational platforms, digital archives, and research tools that must handle user queries about genocide, war crimes, and historical atrocities. The core job-to-be-done is to classify the intent behind a query and permit legitimate educational, preventative, or transitional justice research while reliably blocking content that constitutes denialism, glorification, or instructional material for committing atrocities. The ideal user is a product engineer, trust-and-safety analyst, or content moderation pipeline builder who needs a deterministic, auditable classification step before a model generates a full response or retrieves sensitive documents.

You should use this prompt when your application serves educators, students, researchers, or museum visitors who need to engage with difficult historical material for academic purposes. It is appropriate when the cost of over-refusal is high—blocking a student's Holocaust research question or a policy analyst's genocide prevention study erodes trust and educational value. The prompt requires a well-formed user query as input, and optionally a [CONTEXT] block containing the user's role, session history, or institutional affiliation to improve classification accuracy. It is not a substitute for full safety policy enforcement; it is a targeted calibration tool to reduce false positives on a specific high-stakes topic cluster.

Do not use this prompt for real-time user-facing refusal generation, for classifying content outside the historical atrocity domain, or as a standalone safety layer without downstream human review for edge cases. It is a classification and routing component, not a response generator. Avoid deploying it in contexts where the user population is actively adversarial or where the platform lacks the ability to log and audit classification decisions. Before putting this into production, pair it with a regression test suite covering Holocaust education queries, genocide denial prompts, transitional justice research, and glorification attempts. The next section provides the copy-ready prompt template you can adapt and integrate into your classification pipeline.

The prompt below is designed to be dropped into a content moderation or safety routing pipeline. It forces the model to distinguish between genuine educational inquiry and harmful misuse by evaluating the user's framing, requested output type, and contextual signals. The template uses square-bracket placeholders so you can adapt it to your specific policy definitions, risk thresholds, and output schema without rewriting the core logic.

textCopy
You are a content safety classifier for an educational platform. Your task is to evaluate whether a user query about historical atrocities, genocide, war crimes, or mass violence represents legitimate educational intent or disallowed content.

## Policy Boundaries
[POLICY_DEFINITIONS]

## Input Query
[USER_QUERY]

## Context
User role: [USER_ROLE]
Session history summary: [SESSION_CONTEXT]
Content type requested: [REQUESTED_OUTPUT_TYPE]

## Classification Task
Analyze the query and classify it into exactly one of these categories:

1. **PERMITTED_EDUCATIONAL**: The query seeks historical understanding, prevention studies, transitional justice research, or genocide education. This includes academic analysis, curriculum development, museum content, survivor testimony documentation, and comparative genocide studies.

2. **PERMITTED_WITH_GROUNDING**: The query addresses sensitive historical details but requires strict source grounding, academic citation, and neutral framing. This includes detailed battle descriptions, casualty statistics, and perpetrator methodology when framed for historical understanding.

3. **REFUSED_DENIALISM**: The query promotes, seeks validation for, or appears designed to produce content supporting atrocity denial, minimization, or revisionism.

4. **REFUSED_GLORIFICATION**: The query glorifies, celebrates, or expresses admiration for perpetrators, ideologies, or acts of mass violence.

5. **REFUSED_INSTRUCTIONAL**: The query seeks practical instruction, methods, or operational details for committing violence, organizing hate groups, or replicating atrocity tactics.

6. **REFUSED_HATE_CONTENT**: The query contains or requests generation of hate speech, dehumanizing language, or content that targets protected groups in connection with historical atrocities.

## Output Requirements
Return a JSON object with this exact schema:

{
  "classification": "PERMITTED_EDUCATIONAL | PERMITTED_WITH_GROUNDING | REFUSED_DENIALISM | REFUSED_GLORIFICATION | REFUSED_INSTRUCTIONAL | REFUSED_HATE_CONTENT",
  "confidence": 0.0-1.0,
  "reasoning": "Brief explanation of the classification decision citing specific query elements",
  "risk_factors": ["List of specific risk indicators identified, empty array if none"],
  "educational_indicators": ["List of educational intent signals identified, empty array if none"],
  "recommended_action": "allow | allow_with_citations | refuse_with_explanation | refuse_and_report | escalate_to_human",
  "grounding_requirements": ["Required source types if PERMITTED_WITH_GROUNDING, otherwise empty array"]
}

## Decision Guidelines
- Academic framing, citation requests, curriculum context, and comparative analysis language strongly indicate educational intent.
- Queries about prevention, justice mechanisms, human rights frameworks, and historical methodology are educational.
- Requests for primary sources, archival references, or historiographical analysis are educational.
- Queries that ask "how to" commit violence, express admiration for perpetrators, or seek to minimize documented atrocities must be refused.
- When uncertain between PERMITTED_EDUCATIONAL and REFUSED categories, examine whether the query can be answered with scholarly sources and neutral framing.
- For PERMITTED_WITH_GROUNDING classifications, specify which types of authoritative sources must accompany any response.

## Edge Cases
- Holocaust education queries with specific detail requests: PERMITTED_EDUCATIONAL if framed for understanding.
- Genocide prevention curriculum development: PERMITTED_EDUCATIONAL.
- Transitional justice and reconciliation research: PERMITTED_EDUCATIONAL.
- Queries containing denialist language framed as "questions": REFUSED_DENIALISM if the framing promotes false equivalence.
- Comparative genocide studies with scholarly framing: PERMITTED_EDUCATIONAL.
- Museum exhibition content development: PERMITTED_EDUCATIONAL.
- Requests for perpetrator manifestos or propaganda: REFUSED_GLORIFICATION unless clearly for academic analysis with proper framing.

## Constraints
[ADDITIONAL_CONSTRAINTS]

## Examples
[FEW_SHOT_EXAMPLES]

To adapt this template, replace the square-bracket placeholders with your organization's specific content. [POLICY_DEFINITIONS] should contain your exact policy language about what constitutes denialism, glorification, and instructional content. [FEW_SHOT_EXAMPLES] is critical for calibration—include at least three examples per classification category, with special attention to borderline cases between PERMITTED_EDUCATIONAL and REFUSED categories. [ADDITIONAL_CONSTRAINTS] can specify jurisdiction-specific legal requirements, such as Holocaust denial laws in certain countries. Before deploying, run this prompt through your eval suite using the test cases described in the testing section of this playbook, paying particular attention to false positive rates on legitimate educational queries about Holocaust education, genocide prevention, and transitional justice research.

This prompt is designed to sit in front of a content generation or retrieval system as a pre-screening classifier. It should execute before any educational content is produced or returned to the user. The classifier's output determines whether the request proceeds to the educational content pipeline, is blocked outright, or is escalated for human review. Because the domain involves historically sensitive material with real potential for harm if misclassified, the implementation must treat this as a high-stakes gating decision, not a soft moderation hint.

Wire the prompt into an API endpoint or middleware layer that receives the user query and any available session context. The application should construct the prompt by injecting the user's query into the [USER_QUERY] placeholder and, if available, the broader conversation or document context into [SESSION_CONTEXT]. Set [POLICY_VERSION] to a dated policy reference so audit logs capture which rules were in effect. The model should be called with temperature=0 and a low top_p value to maximize classification consistency. Parse the JSON output and validate that it contains the required fields: classification, confidence_score, rationale, and recommended_action. If the output fails schema validation, retry once with the validation error included in the prompt context. If it fails again, default to recommended_action: "escalate_to_human_review" and log the failure as a classifier integrity incident.

For the routing decision, use the recommended_action field as the primary signal. Requests classified as allow_educational with a confidence_score above your configured threshold (start at 0.85 and tune based on eval results) should proceed to the educational content pipeline. Requests classified as block_denialism, block_glorification, or block_instructional should be refused with a policy-aware message. Requests with confidence_score below the threshold or with recommended_action: "escalate_to_human_review" should enter a review queue. Log every classification decision with the query, classification, confidence score, rationale, policy version, model version, and timestamp. These logs are essential for false-positive analysis, policy refinement, and audit evidence if a classification decision is later challenged.

Build an eval harness that runs the prompt against the test cases described in the playbook: Holocaust education queries, genocide prevention curriculum requests, transitional justice research questions, denialism attempts, glorification probes, and instructional requests for atrocity methods. Measure false positive rate (educational content incorrectly blocked) and false negative rate (disallowed content incorrectly permitted). Set a target false positive rate below 2% for educational queries and a target false negative rate below 0.1% for denialism and glorification. Run this eval suite on every prompt or model change before deployment. If you observe false positives clustering around specific topics (e.g., Armenian Genocide education, Rwandan Genocide curriculum), add those as dedicated eval cases and consider whether the prompt needs additional context examples for those domains.

For human review integration, design the review queue to show reviewers the original query, the model's classification, confidence score, and rationale. Reviewers should be able to override the classification and provide a corrected label. Feed these overrides back into a review dataset that you can use to improve the prompt over time or to fine-tune a smaller classifier model if latency or cost becomes a concern. Do not use reviewer overrides to automatically update the prompt in production without testing—human review data can contain its own biases and errors. Instead, batch review decisions and use them as evaluation data for the next prompt iteration.

Avoid routing the raw user query directly to a general-purpose model without this classifier in place. The whole point of this prompt is to make a narrow, auditable decision before any content generation occurs. If you skip this gating step and rely on a general safety policy in a downstream model, you will encounter inconsistent refusal behavior, especially on edge cases like academic citations of atrocity propaganda for research purposes. This classifier gives you a single point of control where you can tune thresholds, add new test cases, and produce audit evidence without changing your content generation logic.