Claude-interpreted vs Raw

The conclusions are similar: both confirm agentic aversion to CSS, boilerplate is heavy, Markdown is cleaner, and max_content_tokens cuts mid-content, but the two scripts produce meaningfully different data in three ways:

1. Measurement accuracy is completely different. The Claude-interpreted script estimated 13,700–14,200 chars for the short HTML page. The raw script measured 25,925 chars, nearly double. Claude was significantly underestimating. For the spec, only the raw numbers are citable.

2. The raw script found something the interpreted script missed. The default truncation limit finding, that the third test truncated at 20,696 chars even with no max_content_tokens set, only appears in the raw results. Claude didn’t flag this in its interpretation because it attributed the missing content to JavaScript rendering rather than a character limit. Both were actually happening simultaneously, but Claude only noticed one cause.

3. The raw script quantifies boilerplate precisely. “81% boilerplate before the first heading” and “97.5%” are exact, reproducible measurements. Claude’s equivalent estimates “~60-65%”, varied between runs of the same test. The raw numbers are what belong in the Agent-Friendly Documentation Spec, Known Platform Limits table.

The interpreted script is still useful as a record of what the agent perceives it received, which is arguably what matters most for agent developers and docs teams. A model that receives 25,925 chars but estimates 13,700 is working with a distorted picture of the content, and that gap itself is a finding worth noting.