Short field notes from the Null Overflow bench — original stubs, not a scraped feed.

• 2026-09-10 · /lab/blog/calibration-loops-for-genaiCalibration loops for GenAI product sense

  Why human-in-the-loop magnitude checks beat raw leaderboard chasing when shipping retrieval and tool-use stacks.

• 2026-08-28 · /lab/blog/embedding-drift-on-crt-timeEmbedding drift on CRT time

  Vector indexes age quietly. A short field note on monitoring centroid shift without drowning in dashboards.

• 2026-08-04 · /lab/blog/eval-harnesses-that-fit-in-a-terminalEval harnesses that fit in a terminal

  Small golden sets, deterministic seeds, and phosphor-readable failure traces — how we keep evals runnable on a laptop.

• 2026-07-19 · /lab/blog/tool-calling-without-the-circusTool calling without the circus

  Schema discipline, idempotent side effects, and a refusal path that does not soft-hallucinate arguments.

• 2026-06-30 · /lab/blog/latency-budgets-as-product-copyLatency budgets as product copy

  If p95 is 4.2s, the UI should say so — treating timing as UX copy instead of an infra footnote.

← LabMAGNITUDE RUN