FIELD NOTESIG-01472026-06-19
Spec-first prompting cut rework across repeated build cycles
Writing an executable acceptance spec before any generation consistently reduced downstream correction passes versus ad-hoc prompting.
01 · SIGNAL DECK
SAMPLE FEEDrepresentative entriesTechnical intelligence, logged as it happens
A curated feed of experiments, findings, lessons, and incidents from active delivery. Each entry is a field observation — dated, classified, and traceable to a build.
EVALSIG-01462026-06-17
Long-context retrieval degrades silently past ~80% window fill
Models stopped citing mid-context evidence without signalling uncertainty. Caught by adversarial needle placement in the eval harness.
LESSONSIG-01452026-06-14
Agent loops need a hard validation gate, not a confidence threshold
Self-reported confidence was uncorrelated with correctness on refactor tasks. A deterministic test gate caught what scoring missed.
INCIDENTSIG-01442026-06-11
Tool-call schema drift broke integrations in staging
A model update changed argument-ordering preference. Pinned contract tests flagged it before it reached release.
FIELD NOTESIG-01432026-06-08
Smaller, gated increments beat large autonomous runs
Mean rework per change dropped when build increments were bounded to a single testable unit.