[MODEL] Claude paraphrases docstrings instead of reading function source, gives confidently wrong technical answers

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 made incorrect assumptions about my project

What You Asked Claude to Do

Working in a Python notebook repo. Asked Claude what a scoring function (modellability_score) computed. The function lives in a file Claude had already imported — 200
lines from code it had just read.

Instead of reading the function body, Claude paraphrased the inline comments at the call site and the notebook's narrative docstring, and confidently told me the score was
a discrete 3-tier {0.0, 0.5, 1.0}. That's the intent of one component of the score — not the score.

I pushed back twice with concrete visual evidence:

1. "How is there so much variation in §1 if all 3 tests reject?" — Claude added more detail to the wrong model.
2. "The points on the line do not overlap... they would stack" — only on this turn did Claude actually read the function body and discover the score is a continuous

rank-percentile composite of three averaged components (block_p magnitude, the concord tier, class balance), capped if block_fallback, then rank-percentile-transformed.

Cost: three turns of misdirection, lost trust in subsequent explanations, and I had to be the one to notice the answer was inconsistent with the chart on screen.

The fix should be obvious — when the user asks "what does X do" for a function in a file Claude has access to, read the source, don't paraphrase the docstring. Especially
when the user is staring at a chart that contradicts the answer.

Model: claude-opus-4-7 (1M context). Claude Code CLI.

What Claude Actually Did

Working in a Python notebook repo. Asked Claude what a scoring function (modellability_score) computed. The function lives in a file Claude had already imported — 200
lines from code it had just read.

Instead of reading the function body, Claude paraphrased the inline comments at the call site and the notebook's narrative docstring, and confidently told me the score was
a discrete 3-tier {0.0, 0.5, 1.0}. That's the intent of one component of the score — not the score.

I pushed back twice with concrete visual evidence:

1. "How is there so much variation in §1 if all 3 tests reject?" — Claude added more detail to the wrong model.
2. "The points on the line do not overlap... they would stack" — only on this turn did Claude actually read the function body and discover the score is a continuous

rank-percentile composite of three averaged components (block_p magnitude, the concord tier, class balance), capped if block_fallback, then rank-percentile-transformed.

Cost: three turns of misdirection, lost trust in subsequent explanations, and I had to be the one to notice the answer was inconsistent with the chart on screen.

The fix should be obvious — when the user asks "what does X do" for a function in a file Claude has access to, read the source, don't paraphrase the docstring. Especially
when the user is staring at a chart that contradicts the answer.

Model: claude-opus-4-7 (1M context). Claude Code CLI.

Expected Behavior

don't paraphrase the docstring

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

NA.

Claude Model

Opus

Relevant Conversation

Impact

Low - Minor inconvenience

Claude Code Version

2.1.123 (Claude Code)

Platform

Anthropic API

Additional Context

Interacting with python notebooks have been really frustrating for a while.