/results: this note describes how testing data is captured

/results is a directory that gets created automatically when you run the script - it likely doesn’t exist yet. The script creates it with this line:

results_dir = Path(__file__).parent / "results"
results_dir.mkdir(exist_ok=True)

Each time you run a test, it saves a .json file inside it, like:

claude-api/results/20260302_142301_test1_short_html_no_limit.json

JSON is just a data format; it’s the raw output from each API call so you have a record of exactly what came back.

Should /results be committed? These are likely more helpful published for contributors to review, but reconsider as you test.

First Iteration Limits

Claude is interpreting what it received, not a raw dump. That’s why the character counts vary slightly between runs - Claude is estimating, not measuring precisely. The reliable, objective data in our results is actually the token counts in the usage block, which come directly from the API response metadata and don’t pass through Claude’s interpretation at all.

We’re relying on Claude to tell us what it received, rather than inspecting the raw fetch result directly. The script does try to capture the raw content length from the web_fetch_tool_result block, but that part isn’t working reliably - fetch_content_length_chars is showing 0 in the results.

Second Iteration Differences

Claude’s role is minimal. The prompt is just "Fetch this URL: {url}" with max_tokens=128 - we only need Claude to trigger the fetch, not describe anything. This also makes it much cheaper to run.

All analysis is Python. The analyze_content() function measures directly from the raw string - exact character counts, CSS indicator detection in the first 2000 chars, where the first heading appears, whether the content ends cleanly or mid-word. No estimation, no interpretation.

Output files - _raw_results.json and _raw_summary.md so they don’t collide with the files from web_fetch_test.py.

The raw content extraction depends on the web_fetch_tool_result block having a specific structure. If the API response format differs from what the script expects, raw_content_chars show None- which is itself a useful finding to document.