Blog

• LLM evals · Jev · TypeSafe · LLM routing · AI evaluation2026-09-18 · 4 min

  Can Jev Route Between Cheap and Expensive LLMs?

  I tested Jev as a router between Haiku and Opus. Judging Haiku’s answer got 94% of Opus’s accuracy at 54% of the cost — but predicting failures before Haiku answered was much less effective.

  read →
• Harness · LLM · Combination2026-09-16 · 2 min

  Bigger Model or Better Harness?

  How much does the harness really matter? I tested several model + harness combinations on the same benchmark and found some surprisingly large differences in score, cost, and latency, including a cheap model going from worst to best depending on what it ran with.

  read →
• document QA · benchmark2026-08-30 · 5 min

  What Makes a Document QA Benchmark Hard

  Your agent passes your whole benchmark and still fails in production. 7 measured levers that make document QA question genuinely harder and more practical - each with an example you can copy and a diagram of why it bites.

  read →
• announcement2026-08-26 · 1 min

  The forum has its own database now — and this blog opens

  Same /threads URLs, new deployment and database underneath — and a blog opens alongside the forum.

  read →