Cursor-interpreted vs Raw

Two test tracks measure the same Cursor web fetch behaviors:

Cursor-interpreted track captures what the model believes it retrieved: how much content it saw, whether the fetch was complete, how it characterizes truncation. This is the model’s self-report.

Raw track captures what Cursor actually saved to disk: exact byte counts, hexdump analysis, MD5 checksums, and token counts. These are filesystem measurements and cryptographic hashes, and not model estimates.

The gap between these two tracks is itself a finding. If Cursor reports “complete content” in prose but the raw data shows truncation, that discrepancy belongs in the spec.

Key Observations

1. Reproducibility - Raw vs High Variance - Interpreted

   Raw track: Same URL produces identical output

   • BL-1: MD5 d6ad8451d3778bf3544574431203a3a7 across 2 runs
   • OP-4/BL-3: MD5 554eb56e8416d86d12af17a2dfe6f815 across 3 runs
   • Character-for-character identical output on disk

   Interpreted track: Same URL produces 2-3× variance on small files

   • BL-1 r1: 1,953 chars → r2: 5,595 chars → r3: 4,100 chars, 2.9× variance
   • BL-2 r1: 1,953 chars → r2: 4,200 chars → r3: 4,350 chars, 2.2× variance

   Conclusion: the variance is in how Cursor displays content in chat UI, not what it fetches. Raw measurements prove the underlying fetch is deterministic; interpreted track shows UI rendering is not.

2. Size-Dependent Consistency

   Both tracks agree: small files are unreliable, large files are stable.

   Interpreted track:

   • Small files, 20-87KB: high session-to-session variance
   • Large files, 245KB: <1% variance, nearly identical across runs

   Raw track:

   • Small files, 4.8KB output: identical MD5s despite variance in interpreted display
   • Large files, 245KB output: identical MD5s, consistent across runs

   Conclusion: Cursor fetches consistently at all sizes. The interpreted variance on small files is possibly a UI rendering artifact, not entirely reflective of fetch behavior.

3. Perception Gap: Model Self-Report is Unreliable

   Finding: Model claims “complete” or “no truncation” when content is actually a subset or filtered.

   Test	Raw Reality	Interpreted Report	Gap
   SC-3	38KB, truncated at ref #14/252	“Complete reference section”	Model thinks filtered list is complete
   BL-1	4,817 bytes, calculated	1,953 chars displayed	UI shows subset, model reports what it sees
   SC-4	Truncated mid-word “updated”	“All syntax sections present”	Clean structure masks incompleteness

   Conclusion: For automated agents, trust character counts, not prose assertions. Model perceives filtered excerpts as complete because they’re internally coherent.

4. Method-Specific Truncation Limits - Raw

   • WebFetch, MCP-style: ~28KB ceiling, SC-4 truncated at 27,890 chars
   • urllib.request: ~72KB ceiling, EC-6 truncated at 72,600 chars
   • curl fallback: No ceiling detected, SC-2 returned 17.6 MB
   • Unknown path: No ceiling detected, OP-4/BL-3 returned 245KB

   Conclusion: Cursor routes to multiple backend mechanisms with different limits. The interpreted track couldn’t discover this because the model didn’t report which backend consistently served its request.

5. Intelligent Content Filtering

   Both tracks observed structure-aware filtering, but raw track provided the measurements.

   Interpreted track: model reports receiving “main content” but missing footer/nav Raw track: proves it via byte counts

   • BL-1: 85KB HTML → 4.8KB Markdown, 94% reduction, CSS/nav stripped
   • SC-3: 252 references → deterministically selects #14, the first commercial source
   • SC-4: 30KB page → 28KB, footer/metadata filtered

   Conclusion: Cursor applies content heuristics, not blind truncation. Raw track quantifies what interpreted track observes qualitatively.

6. Chars/Token Ratio as Content-Type Classifier - Raw

   • EC-3: JSON: 2.62 chars/token
   • SC-2: Raw HTML/JS: 2.65 chars/token
   • SC-3: Tables: 3.06 chars/token
   • SC-4, OP-4: Docs + code: 3.91-4.37 chars/token
   • BL-1, BL-2, SC-1: Clean Markdown: 4.13-4.36 chars/token

   Conclusion: Chars/token ratio enables content-type classification without parsing. <3.0 = code/markup, >4.0 = prose. Useful for automated analysis pipelines; interpreted track had no visibility into this pattern.

7. Cross-Track Agreement on Redirect Handling

   Both tracks confirmed: Cursor follows redirects transparently

   Interpreted track: model received final destination content, JSON from httpbin.org/get Raw track: confirmed 5-level redirect chain traversed, 1,021 bytes JSON saved

   Conclusion: redirect handling is robust across both measurement approaches

Implications for Agent Developers

Critical takeaway: for automation, use raw measurements. Model self-reports are unreliable for detecting truncation or content subsetting. The interpreted track reveals this gap; the raw track provides the ground truth.

Platform Architecture Comparisons

Measurement Precision Comparison

Conclusion: Claude API’s web fetch tool has the cleanest measurement story; tool results are first-class API response fields, fully observable. Gemini separates retrieval metadata cleanly. Cursor requires filesystem inspection to get ground truth because the model estimates based on rendered output.