The public engineering journal

Dated field reports from the bench. What broke, what was changed, and what the change actually bought — written while it was still fresh.

LOG-00892026-06-20eval-harness · validation

Hardened the eval harness against fixture rot

Versioned every fixture and added a reproducible seed. Two cases that looked flaky were actually a real silent-failure bug in retrieval — fixed, then added a regression probe so it can never return quietly.

LOG-00882026-06-18delivery-os

Added a hard stage gate between Build and Validation

Builds were sliding toward release without an evidence pass. The new gate blocks on a missing eval report. More friction up front; far less rework downstream. Net cycle time went down, not up.

LOG-00872026-06-15spec-lint

spec-lint caught an untestable acceptance criterion

“Should feel fast” failed the lint. Rewriting it as a p95 latency budget surfaced a missing database index before a single line of code was generated.

LOG-00862026-06-13context-probe · eval

Mapped the long-context degradation curve

Ran the needle probe across window-fill levels. Recall holds until ~80% fill, then drops without the model signalling lower confidence. Documented the curve and added a fill-ratio guard.

LOG-00852026-06-09contract-tests

Pinned tool contracts after a silent schema drift

A model update reordered preferred arguments and broke integrations in staging. Pinned contract tests now run on every integration and alert on drift before release.