The best solution in this scenario is to implement image moderation.
Azure AI Content Safety provides dedicated image moderation capabilities that automatically detect, classify, and score harmful visual content (such as violence, hate speech, sexual content, and self-harm). It natively allows you to configure threshold blocklists based on four severity levels (Low, Medium, High), perfectly matching the requirements.
Incorrect:
[Not A]
Optical Character Recognition (OCR) only extracts text embedded within an image. If a user uploads a harmful photo that contains no text (e.g., a violent image), OCR and keyword scanning will completely fail to detect it.
[Not B]
The Prompt shields feature is designed to protect generative AI models from text-based attacks, such as jailbreaking or prompt injection. It does not analyze or classify the visual pixels of an uploaded photo for harmful content.
[Not C]
While Azure AI Content Safety uses blocklists, standard text blocklists match specific exact terms or regex patterns. They are used for text moderation, not for evaluating the severity levels of raw visual imagery.
Reference:
https://supportzebra.com/blog/what-is-ai-image-moderation

To identify whether model input, model output, or expanded tool usage is driving the cost spike, you should focus on token usage.
Granular tracking of token usage breaks down the costs directly:
Input vs. Output Tokens: Tracing token usage isolates whether the user prompt (input) has grown or if the model is generating longer, more expensive responses (output).
Tool Calling: Expanded tool usage consumes extra tokens each time the agent schema processes an execution loop, which will clearly reflect as inflated token counts.
Reference:
https://www.mindstudio.ai/blog/subtraction-principle-agent-harness-optimization

The correct answer is D. Set the output_content_format=ContentFormat.MARKDOWN value . Azure Document Intelligence Layout API can return extracted document content in Markdown format, preserving semantic structure such as headings, paragraphs, sections, tables, and other layout elements. Microsoft's Document Intelligence layout guidance shows the Python SDK pattern for analyzing a document with the prebuilt-layout model and setting output_content_format=ContentFormat.MARKDOWN in the begin_analyze_document call. The Markdown output is returned in the top-level content section of the analysis result.
This minimizes development effort because the service produces structure-preserving Markdown directly, rather than requiring custom post-processing to reconstruct sections and table formatting from raw OCR spans. Microsoft's Markdown output documentation states that specifying Markdown output produces semantically structured content that maintains paragraphs, headings, tables, and other document elements in their proper hierarchy.
Option A only changes validation behavior and does not generate Markdown. Option B requests figures, not structured Markdown. Option C uses an incorrect parameter name; the documented SDK setting is output_content_format, not content. Reference topics: Azure Document Intelligence Layout API, Markdown output, PDF analysis, table extraction, and Foundry Tools document processing.

Explanation:
Box 1: No
Changing the content filtering configuration to low severity will not resolve the fine-tuning job issues.
Box 2: Yes
The difference between a 12% defect rate in Run1 and a 4% defect rate in Run2 is entirely consistent with the different severity thresholds configured for the two evaluations.
Box 3: Yes
The protected material evaluation metric is entirely unaffected by changes to the content harm severity threshold.
Reference:
https://learn.microsoft.com/en-us/python/api/azure-ai-evaluation/azure.ai.evaluation