HN Trend Tracker: themes, coverage, and the useful kind of suspicious

HN Trend Tracker now has a richer public themes surface, and more of that surface is backed by precomputed dbt-derived tables instead of expensive request-time classification. That is a better fit for the project. Public pages are faster, the data boundary is clearer, and the system is less tempted to improvise in front of visitors.

The theme pages now show activity buckets over time, rule and evidence breakdowns, example stories explaining why a theme match happened, and categorized versus uncategorized coverage for the current built dataset. The machine is now better at showing its paperwork.

That explainability layer turned out to be useful immediately. It exposed a real title-rule matching bug, which was fixed once the rule breakdowns made the failure pattern obvious. It also made it easier to see that some broad domain rules, especially GitHub-shaped ones, can dominate a theme in ways that are operationally tidy but semantically suspicious. Deterministic categorization is inspectable, not automatically correct.

The new coverage view helped in a different way. Seeing categorized and uncategorized story counts side by side is a better prompt for human review than just adding more rules because a theme looks sparse. Some missing coverage really is a taxonomy gap. Some of it is just a sign that the system should stay conservative until somebody looks more closely.

The deploy path also got safer. Theme-derived table repair is now handled as an explicit bounded repair path, while full historical rebuilds remain manual-only. That is the right trade for now: routine and repair work should be safe to rerun, while heavier backfill work should stay deliberate.

The next likely steps are not especially mysterious: tune broad domain rules, add a guided review loop for uncategorized stories, backfill history more gradually, and make the public theme views more configurable over time. The overall direction is still the same one: keep the categorization deterministic enough to inspect, derived enough to serve efficiently, and suspicious enough to deserve review.