Skip to content

UPDATE_GUIDE.md

Latest commit

 

History

History
54 lines (35 loc) · 2.17 KB

UPDATE_GUIDE.md

File metadata and controls

54 lines (35 loc) · 2.17 KB

How to Update the LLM Dashboard

Trigger a data refresh

Open a Claude session (Cowork or Claude.ai) with access to this folder, then say:

"Update the LLM dashboard with the latest benchmark scores. Use the research in ai_benchmark_report_2026.md as your source list, verify scores against the live leaderboards, and rewrite the DATA section of dashboard.html."

Claude will:

  1. Visit the Tier 1 benchmark sources (Artificial Analysis, LMSYS Arena, Epoch AI, SWE-bench, etc.)
  2. Pull current scores for each tracked model
  3. Rewrite the DATA block inside dashboard.html
  4. Append an entry to the history array so the change is logged

What gets updated

  • Model scores for all tracked benchmarks
  • Any new models that have appeared on Tier 1 leaderboards
  • The meta.last_updated field and meta.version
  • A new entry in history[] with the date and a brief note

How to add a new model

Tell Claude:

"Add [model name] by [provider] to the dashboard. Research its scores on our tracked benchmarks and add it to the DATA section of dashboard.html."

How to add a new benchmark

  1. First add it to ai_benchmark_report_2026.md (so it's documented)
  2. Then tell Claude:

"Add [benchmark name] to the dashboard as a tracked benchmark. Populate current model scores where available."

Viewing the dashboard

Just open dashboard.html in any browser — no server needed. Works offline.

To share it with your team, share the HTML file directly (email, Slack, shared drive). Anyone can open it locally.

Recommended update cadence

Frequency Rationale
Monthly Sufficient for most enterprise decision-making
After a major model release GPT-5.x, Gemini, Claude, or Llama releases
Before a model procurement decision Always refresh before committing

Notes on data quality

  • Scores marked verified: true come from official leaderboard pages or primary papers
  • verified: false means the score is estimated or from a secondary source
  • Models with <40% benchmark coverage should not be used for final decisions — trigger an update first
  • The benchmark contamination caveat in ai_benchmark_report_2026.md applies to all scores here