Looks like the token count is accurate for Jev, more than 5x more input tokens.

Looks like the token count is accurate for Jev, more than 5x more input tokens. Because this approach is making a call for every line in the document and passing criteria with each line vs. a single prompt and response from the LLM. I am sure this could made more efficient, but **as-is it is a faster but not less expensive option.** (at least if you are using Deepseek V4 Flash on OpenRouter for the LLM)