The Daily Edition is built on a 15-step pipeline that runs every morning at 06:00 CET. Each step is a standalone Python program orchestrated by a single runner. The pipeline takes about 90 minutes from first headline collection to published edition. It produces one of the most carefully verified pieces of editorial AI output we know how to produce.
The shape of the system
Three model tiers, assigned by cognitive type. Opus for editorial judgment, Sonnet for classification, Haiku for triage. Chosen because classification, creative editorial work, and triage are different cognitive tasks.
Strict separation between AI and code. AI for nuanced judgment; code for arithmetic, validation, and rule enforcement. We do not trust the model with weighted averages or ID lookups.
Verification at every stage. No AI output is accepted unverified. Twenty-nine deterministic checks run before the final AI critic sees the page.
The pipeline
15 steps: Newsletter ingestion, Event detection, Article scraping (141 feeds), Haiku pre-filter, Opus scoring (6 dimensions), Event boosting (multi-day cascade), Macro narrative matching, Article selection, Edition narration (5,000+ lines), Hook generation with leak-checking, Tier derivation, Final review, Website + JSON generation, SEO injection.
Cross-model scorer calibration
Scoring rubric cross-validated against Claude Opus 4.7, OpenAI GPT-5.4, and Google Gemini 3.1 Pro on 34 articles. Eight cross-validated examples plus one hypothetical anchor. Dimension values are medians across three models.
Verification surfaces
Coherence verification (Sonnet), description self-checks (Opus), hook leak detection, mechanism label validation, continuity delta validation, final review critic (Opus), 29 deterministic structural validators, and weekly feed health monitoring.
What we know we don't have
No human-graded reference set. No automated editorial-drift detection. English-only sources. Manual corrections processing.