[chaliy]

What changed

Adds the first real-model eval results as references in crates/markkit-eval/results/: markkit_qa-2026-10-28.{md,json} and markkit_extract-2026-10-28.{md,json}. Also updates knowledge/operations/eval.md (a QA results table, the extract re-run, and where the Doppler keys live) and knowledge/log.md. No code changes.

Why

The LLM eval (#56) had never run against real models. These results are the reference baseline to compare against when the HTML extraction changes.

Before / After

markkit_qa with Claude Sonnet 5.5, 23 questions per input:

input	correct	mean input tokens
markdown	23/23	464
main_content	22/23	377
raw html	23/23	945

markkit markdown is as accurate as raw HTML for about half the input tokens. The one miss is a known gap: main-content mode drops the forum question body. Every OpenRouter Muse Spark call returned HTTP 402 because the key's daily credit limit is too low, so those cases are N/A. The error text in the JSON has the key's dashboard ID redacted.

markkit_extract re-run: main_content passes 3/10 pages (it was 2/10; product_page passes again after the escaping fix). Mean scores are 87.4% for main_content and 73.6% for default.

Risk

• Low. Only results and docs change.

Checklist

• Tests added or updated (not applicable, results only)
• Backward compatibility considered

🤖 Generated with Claude Code