Model poor performance: fails to synthesize patterns from data it already read and summarized

Model Poor Performance

Claude Opus 4.6 was asked to read all extraction data files (17 cases) and write a comprehensive summary. It successfully read every file and produced a detailed summary document. However, when later asked to solve a display issue that depended on a pattern visible in that same data, the model failed to recall or synthesize the pattern — despite having already processed the information.

This is not a bug in Claude Code tooling. This is a model performance issue — the model cannot effectively use information it has already consumed.

What Happened

1. User asked Claude to read all 17 test case extraction files and summarize them
2. Claude read every file, wrote a detailed summary (~400 lines) covering all cases
3. The extraction data clearly showed: criterion-A cases displayed truncated steps, criterion-B cases displayed all steps
4. When the user later asked Claude to fix the step display, the model did not use the pattern from the data it had already read
5. Instead, the model spent significant tokens on multiple wrong theories:

• First proposed a blanket "truncate all cases" approach (would have broken 3 cases)
• Then proposed the reference program overrides a setting internally (wrong)
• Then guessed the output was from a different method (wrong)
• Then asked the user questions that the data already answered

1. The user had to explicitly say "all data in the extracted file" to redirect the model back to data it had already read
2. Only then did the model re-read the files and find the criterion-A vs criterion-B pattern

Impact

• Wasted tokens on reading, understanding, and summarizing data without extracting the actionable insight
• Multiple round-trips of wrong guesses before arriving at the answer
• User had to guide the model to information the model already possessed
• Total wasted context: the full summary generation pass + multiple failed fix attempts
• Real cost to the user with zero value returned for those tokens

Expected Behavior

When the model has already read and summarized data, it should be able to synthesize patterns from that data when solving related problems — not require the user to point it back to the same files. Pattern recognition across previously consumed data is fundamental to the value proposition of a large context model.

Environment

• Claude Opus 4.6 (1M context)
• Claude Code CLI (VS Code extension)
• Task: Statistical analysis module development matching reference program output