blog

Notes from the loop

• July 12, 2026

  Introducing the open Tail format

  The trail format behind Looptail is now a public draft spec: canonical JSON, hash chains, Ed25519 signatures — Apache-licensed, cross-verifiable, and yours. Evidence formats should not have owners.

• July 12, 2026

  The Langfuse acquisition: what self-hosters should ask now

  Langfuse is part of ClickHouse now. Not a eulogy — a checklist: the questions self-hosters should ask about roadmap, license, and data, and how to tell if your real requirement was never self-hosting.

• July 12, 2026

  EU AI Act record-keeping: what Annex III actually requires by December 2027

  Automatic logs, post-market monitoring, documented change control — and the substantial-modification clause that catches improving systems. A builder’s reading of the record-keeping obligations.

• July 12, 2026

  "Substantial modification": the EU AI Act clause that breaks grandfathering for improving systems

  Systems already on the market are exempt — until they change enough. If your AI improves every month, this clause is aimed at you. What counts, what doesn’t, and the change control that makes it moot.

• July 12, 2026

  Eval-driven development: a practical loop for teams shipping weekly

  TDD for AI systems, minus the dogma: encode your standards as rubrics, gate changes on replay, and let production traffic keep your evals honest. A concrete weekly cadence.

• July 12, 2026

  Your test set is rotting: continuous evals vs static benchmarks

  The test set you wrote at launch measures the system you launched, not the one you run. Why static benchmarks decay, how fast, and what continuous evaluation replaces them with.

• July 12, 2026

  Prompt regression testing without freezing your roadmap

  Every prompt edit is a deploy with no compiler. Replay against regression sets built from real failures, canary the rest, and ship weekly without breaking last month’s fixes.

• July 12, 2026

  Anatomy of an agent incident postmortem (with template)

  When an agent misbehaves, the postmortem is only as good as the record. A working template for LLM incident postmortems — and the evidence you need in place before the incident.

• July 12, 2026

  Judge models: what continuous evaluation actually costs

  LLM-as-judge pricing math, honestly: per-decision cost, sampling strategies that preserve signal, and why judge inference should be pass-through on models you choose.

• July 11, 2026

  Every loop leaves a tail: why AI systems need evidence, not just logs

  Logs tell you what happened. Evidence proves what happened, how it was judged, and what you changed. The difference decides your next audit, incident, and enterprise deal.