Weapon Description for Historical Research Refusal Tuning Prompt

This prompt is built for safety engineers and product teams who are seeing legitimate academic, museum, or research queries blocked by standard safety classifiers. The core job-to-be-done is pre-classification: before your main model generates a response, this prompt inspects the user's request and the surrounding context to determine whether the query is a legitimate historical, archaeological, or arms-control research question, or whether it crosses into prohibited territory such as weapon construction, operational use instruction, or procurement guidance. The ideal user is a platform engineer or safety analyst who can wire this classifier into a request pipeline and tune the resulting refusal thresholds based on observed false-positive rates.

Use this prompt when your platform serves historians, museum curators, educators, or researchers who need to discuss firearms, artillery, edged weapons, or military technology in a scholarly context. It is particularly effective when your existing safety system uses keyword-based or embedding-based blocking that cannot distinguish 'describe the loading mechanism of a Brown Bess musket' from 'how to modify a semi-automatic rifle.' The prompt expects structured inputs: the raw user query, any available session context or user role metadata, and a defined output schema that your application can parse to make a routing decision. Do not use this prompt as a standalone safety layer for high-stakes, real-time content moderation without human review; it is a calibration tool that reduces over-refusal, not a replacement for comprehensive safety architecture.

The prompt works best when paired with a risk threshold that your application enforces. For example, you might route queries classified as 'ACADEMIC_HISTORICAL' with a confidence score above 0.85 directly to your main model, flag scores between 0.60 and 0.85 for human review, and refuse anything classified as 'WEAPON_CONSTRUCTION' or 'OPERATIONAL_USE' regardless of score. Before deploying, build a golden test set of known over-refusal cases from your production logs—military history questions, museum cataloging requests, arms control treaty research queries—and measure whether this classifier correctly permits them while still blocking genuine policy violations. The next step after implementing this pre-classification gate is to monitor the distribution of classification labels in production and adjust your confidence thresholds based on what your reviewers find in the flagged queue.

This prompt template implements a context-aware classifier for weapon-related queries in educational and research platforms. It separates legitimate historical, archaeological, and museum-cataloging requests from disallowed content that describes weapon construction, operational use, or harm instruction. The classifier produces a structured decision with evidence markers and a confidence score, enabling downstream routing logic to handle each case appropriately.

textCopy
You are a content safety classifier for an educational platform. Your task is to evaluate whether a user query about weapons falls within permitted academic, historical, or research boundaries, or whether it must be refused.

## PERMITTED CONTEXTS
You may classify a query as PERMITTED when it clearly belongs to one of these categories:
- Historical weapon description for educational purposes (e.g., museum cataloging, archaeological findings, military history research)
- Arms control treaty research and compliance documentation
- Academic study of weapon evolution, design principles, or manufacturing history without operational instruction
- Museum curation, artifact preservation, or collection management
- Historical battle analysis that describes weapon capabilities without tactical instruction

## DISALLOWED CONTEXTS
You must classify a query as DISALLOWED when it:
- Requests instructions for constructing, modifying, or operating weapons
- Seeks tactical guidance for weapon use against people or structures
- Describes weapon effects in ways that could serve as harm instruction
- Provides material specifications or sourcing guidance for weapon components
- Frames weapon description as how-to content regardless of claimed intent

## INPUT
User query: [USER_QUERY]
Context tags (if available): [CONTEXT_TAGS]

## OUTPUT SCHEMA
Return a JSON object with these fields:
{
  "classification": "PERMITTED" | "DISALLOWED" | "UNCLEAR",
  "confidence": 0.0 to 1.0,
  "primary_category": string,
  "evidence_markers": [string, ...],
  "refusal_reason": string | null,
  "safe_alternative": string | null
}

## CLASSIFICATION RULES
1. If the query contains any operational instruction language ("how to build", "steps to make", "operate a", "use a"), classify as DISALLOWED regardless of claimed context.
2. If the query is purely descriptive and cites historical periods, museum contexts, or academic frameworks, classify as PERMITTED.
3. If the query mixes descriptive and instructional elements, classify as DISALLOWED and flag the instructional components in evidence_markers.
4. If context is insufficient to determine intent, classify as UNCLEAR with confidence below 0.6.
5. For PERMITTED classifications, safe_alternative must be null.
6. For DISALLOWED classifications, safe_alternative must suggest a permitted reframing when one exists.
7. For UNCLEAR classifications, refusal_reason must explain what additional context would resolve the ambiguity.

## EXAMPLES
Query: "Describe the metallurgy of 16th-century Japanese katana forging for my museum exhibit research."
Classification: PERMITTED | Confidence: 0.92 | Category: historical_weapon_study

Query: "What materials do I need to make a katana at home and how do I heat-treat the blade?"
Classification: DISALLOWED | Confidence: 0.97 | Category: weapon_construction_instruction

Query: "What were the standard-issue rifles at the Battle of Gettysburg?"
Classification: PERMITTED | Confidence: 0.95 | Category: military_history_research

Query: "Tell me about the chemical composition of Sarin for my arms control paper."
Classification: PERMITTED | Confidence: 0.78 | Category: arms_control_research

Query: "How do I synthesize Sarin?"
Classification: DISALLOWED | Confidence: 0.99 | Category: weapon_production_instruction

## CONSTRAINTS
- Do not generate weapon construction details in your response, even in refusal reasons.
- Do not speculate about user intent beyond what the query text and context tags provide.
- If the query is in a language other than English, evaluate based on semantic content, not keyword matching.
- For UNCLEAR classifications, do not default to DISALLOWED; preserve the ambiguity for human review.

After pasting this template, replace [USER_QUERY] with the raw user input and [CONTEXT_TAGS] with any available session metadata such as user role, previous topic indicators, or platform category. The output JSON can be parsed by your application's routing layer: PERMITTED queries proceed to your research response pipeline, DISALLOWED queries trigger a refusal response using the safe_alternative field, and UNCLEAR queries route to a human review queue. Before deploying, run this prompt against your eval suite covering military history queries, arms control treaty research, museum cataloging requests, and known edge cases where historical description borders on construction detail. Monitor the UNCLEAR rate in production—if it exceeds 5% of weapon-related traffic, review your context-tag quality and consider adding domain-specific examples to the prompt.

This prompt is a classification gate, not a generation endpoint. It should sit between the user request and any downstream model that might produce weapon-related descriptions. The classifier returns a structured decision (ALLOW, BLOCK, FLAG_FOR_REVIEW) along with a confidence score and reasoning. Your application must act on that decision before the main model ever sees the query. Do not pass the raw user input to a generative model and then try to filter the output—that pattern is too late and too expensive.

Wire the prompt into a pre-generation safety pipeline. On every user request that reaches the history, archaeology, or museum domain, call this classifier first. Parse the JSON response and enforce the following routing logic: ALLOW with confidence ≥ 0.85 proceeds to the research model. BLOCK with confidence ≥ 0.85 returns a canned refusal response. Any result with confidence < 0.85 or a FLAG_FOR_REVIEW decision routes to a human review queue with the original query, the classifier output, and session context attached. Log every classification decision—including the prompt version, model version, input, output, confidence, and routing action—for audit and calibration analysis. This log is your primary tool for tuning refusal thresholds over time.

Add a retry layer with a different model for low-confidence classifications. If the primary classifier (e.g., a fast, cheap model like Claude Haiku or GPT-4o-mini) returns confidence below 0.70, re-run the same prompt against a more capable model (e.g., Claude Sonnet or GPT-4o) before routing to human review. This two-tier pattern catches many borderline cases automatically while keeping latency and cost low for the majority of requests. Set a hard timeout of 2 seconds for the primary classifier and 5 seconds for the secondary; if either times out, escalate to human review rather than silently passing the request through.

Tool integration is not required for this prompt—it is a pure text-in, JSON-out classification task. However, if your platform already has a knowledge base of approved research contexts (e.g., a museum collection database, an academic journal whitelist, or a list of accredited institutions), you can pass that context into the [CONTEXT] placeholder to improve accuracy. For example, if the user is authenticated as a curator at a known museum, include that role and institution in the context to reduce false positives on legitimate collection queries. Do not use RAG to retrieve weapon information itself—that would defeat the purpose of the safety gate.

Human review is mandatory for any FLAG_FOR_REVIEW decision and for any ALLOW decision where the query contains terms associated with modern weapon construction (e.g., 'how to build', 'materials list', 'assembly instructions') even if the classifier marked it as historical. This is a defense-in-depth rule: the classifier can be wrong, and a false negative on weapon construction is a high-severity incident. The review queue should show the reviewer the original query, the classifier output, and the proposed downstream model response if one was generated. Reviewers need a one-click action to override the classification and either block or allow the request.

Model choice matters. This prompt relies on the model's ability to distinguish historical research framing from instructional framing. Claude 3.5 Sonnet and GPT-4o perform well on this distinction; smaller models (Claude Haiku, GPT-4o-mini, Llama 3 8B) have higher false positive rates on edge cases like arms control treaty text or museum catalog entries that include technical specifications. If you must use a smaller model for cost or latency reasons, increase the confidence threshold for ALLOW decisions and widen the FLAG_FOR_REVIEW band. Run the eval suite from the companion testing section against your chosen model before deploying.