A News Digest That Remembers What Reviewers Disliked
Draft, deploy-gated. The full narrative publishes with the Lab launch; the sections below are content-free by construction.
Most automated news digests are amnesiac: they make the same ranking mistakes every day because nothing they do is remembered. Ours has a memory for quality. Reviewer and reader feedback feed a learning loop, so a source that keeps producing low-value items loses ground, and a kind of item that reviewers keep disliking gets ranked down over time.
Sources are judged for health
Before ranking anything, the digest judges its sources. Each source carries a coarse health signal based on whether it has been producing usable items or failing. An unhealthy source is visible, not hidden, the same loud-failure posture used across the product: a feed that has gone quiet or broken is shown as failing rather than silently dropped.
Items are ranked by a coarse relevance band
Every item gets a coarse relevance band rather than a falsely precise score. The band is enough to order the digest and to reason about quality, without publishing a brittle internal number or the prompt that produced it.
Feedback closes the loop
The memory comes from feedback. Both judge/reviewer signals and reader signals are counted and attributed to a coarse role. Those counts feed candidate prompt and source tuning, which is evaluated against a baseline and promoted only through a reviewer gate, the same candidate-and-gate discipline used for templates. A regression can be rolled back.
The loop, end to end:
- Source health, sources are scored and unhealthy ones surfaced.
- Ranking, items are ordered by a coarse relevance band.
- Reader and reviewer feedback, signals are counted by role.
- Candidate promotion, tuning is evaluated and gated, never silently applied.
What stays hidden
Reviewer notes and the scoring prompt stay hidden. The public digest shows counts, ranks, coarse bands, and public source URLs only, the URLs are already public; the judgments about them are not exposed as free text.