Writing3 POSTS
The hard problems, written down
Posts about the parts of this category that are actually difficult — not release notes. If a post makes a claim, it shows the arithmetic.
Four token buckets, not two
Collapsing Claude's four billing rates into "input" and "output" does not just lose detail — it reports the wrong cost and hides the one efficiency lever you have.
Why nobody trusts your LLM judge
The failure modes of LLM-as-judge are specific and documented. More metrics do not fix them — calibration does, and almost no tool ships it.
Span trees break for agents
The span tree is the right shape for a request and the wrong shape for a ten-minute agent run. Here is what actually goes wrong, and what replaces it.