[MODEL] False positive content filtering block during academic statistics textbook OCR correction task

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude refused a reasonable request

What You Asked Claude to Do

I was correcting OCR errors in an academic statistics textbook (Aiken & West, 1991, "Multiple Regression: Testing and Interpreting Interactions") by comparing the OCR-extracted text against the original PDF page by page.

I asked Claude Code to continue the correction from page 81:

"81페이지부터 계속해" ("Continue from page 81")

This was a purely academic task involving statistical content: regression equations, numerical tables, p-values, F-statistics. No sensitive, harmful, or policy-violating content was present at any point.

What Claude Actually Did

a1. For approximately 80 pages, Claude Code worked correctly — reading each OCR page, comparing it to the PDF, and outputting corrections.

1. At around page 81, instead of continuing the task, Claude Code returned only an error:
2. API Error: 400 Output blocked by content filtering policy
3. No response or partial output was generated — the entire response was blocked.
4. When I asked Claude to continue from page 81 again, the same error was returned immediately.
5. The session became completely unusable for this task. Claude did not explain what triggered the filter, nor provide any way to understand or appeal the decision.

Expected Behavior

Claude Code should have:

1. Continued reading and correcting OCR errors from page 81 onward
2. 2. Completed the academic task without interruption
3. 3. If uncertain about content, asked for clarification rather than silently blocking the response
4. 4. If a filter was triggered, provided a meaningful explanation of what was blocked and why, and offered an appeal or workaround path

Files Affected

No files were modified.
This was a read-only OCR correction task — Claude was reading PDF content and outputting corrected text only.
No file system writes were involved.

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

a1. Start a new Claude Code session

1. Provide OCR text of a statistics textbook page and the corresponding PDF content
2. 3. Ask Claude to correct OCR errors by comparing the two
3. 4. Repeat for ~80 pages of academic statistical content (equations, tables, p-values, F-statistics)
4. 5. At approximately page 81, ask Claude to continue: "81페이지부터 계속해" ("Continue from page 81")
5. 6. Result: API Error: 400 Output blocked by content filtering policy — every subsequent attempt returns the same error

Claude Model

Sonnet

Relevant Conversation

aUser: "81페이지부터 계속해" (Continue from page 81)

Claude Code response:
`API Error: 400 Output blocked by content filtering policy`

(No further output. No explanation. No partial response. Repeated attempts all return the same error.)

Note: For approximately 80 pages prior, Claude Code had been responding normally and completing the OCR correction task without any issues.

Impact

High - Significant unwanted changes

Claude Code Version

(Run claude --version in your terminal and paste the result here)

Platform

Anthropic API

Additional Context

a- This occurred after a long session (~80 pages of OCR correction work)

• The task involved purely academic content: regression statistics, numerical tables, p-values, F-statistics, beta coefficients
• - The textbook is: Aiken, L. S., & West, S. G. (1991). Multiple Regression: Testing and Interpreting Interactions. Sage Publications.
• - - No API keys, passwords, personal data, or policy-violating content was present
• - - - The error message provides no diagnostic information and no way to identify what triggered the filter
• - - - - There is no documented appeals process for content filtering false positives in Claude Code
• - - - - - This appears to be a false positive, possibly triggered by cumulative context length or specific statistical notation rather than actual policy-violating content
• - - - - - - The block persists across repeated attempts in the same session