Every event in Scoopkit gets classified into one of 10 categories and 57 subcategories, taxonomy version 1.0.0. Each classification also carries a confidence score, so a caller can tell a clean model release apart from a vague industry rumor without reading the article.
3 categories are fully free, 7 are paid or mixed, and the subcategories split 25 free to 32 paid. The classification runs once per event, right after the article gets merged into it.
Verified 2026-08-27 against the live /v1/taxonomy endpoint and a 100-event sample from the public API.
Facts
| Category | Tier | Subcategories | Example |
|---|---|---|---|
| Model Release & Capability | Free | 7 | New model release, benchmark result |
| Funding, M&A & Capital Events | Free | 7 | Venture round, acquisition |
| Research Publications & Technical Findings | Free | 6 | Preprint, dataset release |
| Product & API | Paid | 7 | Product launch, API deprecation |
| Safety & Policy | Paid | 8 | Safety incident, regulation proposed |
| Org & People | Mixed | 5 | Leadership change, new lab formed |
| Hardware & Infrastructure | Paid | 5 | Chip release, data center event |
| Agentic & Protocol | Mixed | 4 | Agent framework release, computer-use capability |
| Legal & IP | Paid | 5 | Copyright suit, IP or trade secret |
| General Industry | Mixed | 3 | Market analysis, enterprise adoption |
Why type a news event at all
A raw article gives you a headline and a body. It doesn't tell you whether the story is a model release, a funding round, or a lawsuit, and it doesn't tell you how sure anyone should be about that read. Filtering a feed by category means building that classifier yourself first.
Scoopkit does the classification once, at ingest time, so a query for MODEL_RELEASE events or SAFETY_AND_POLICY events just works. The category and subcategory live on the event record, not buried in the article text.
What type_confidence actually means
Every classification comes with a type_confidence score: 0.9, 0.6, or 0.3. These aren't a raw model probability. They're 3 fixed buckets, high, medium, and low, chosen because a model asked to output a bare 0 to 1 confidence number tends to give badly calibrated answers. Fixed buckets are honest about what the classifier can actually tell you.
In a live sample of 100 recent events, 66 landed in the high bucket, 30 in medium, and 4 in low. Most events get classified with real confidence. The low bucket exists for the genuinely ambiguous ones instead of forcing a confident-sounding number onto a guess.
How classification actually runs
Classification happens after clustering, not before. An article gets merged into an event first, then the event gets one classification call, not one call per article.
That call tries a free model first. Scoopkit runs a cascade of no-cost models through OpenRouter, and only falls back to a paid model, Claude Haiku 4.5, once the free options are exhausted. It's the same fallback model used for the clustering tie-break described in how Scoopkit deduplicates AI news headlines.
A real classified event
The Nvidia earnings event from the deduplication post landed in GENERAL_INDUSTRY.MARKET_ANALYSIS with type_confidence: 0.9. Nvidia reporting quarterly revenue is exactly the kind of clean, well-covered story that earns a high-confidence classification on the first pass.
Query that category directly with a free API key, or see the full field list at the API docs.
FAQ
How is a subcategory chosen?
The classifier reads the merged event's headline and summary and picks the closest match from the fixed list of 57 subcategories in taxonomy version 1.0.0. It's a closed set, not a free-text label, so 2 different events in the same subcategory are always comparable.
Which categories are free vs. paid?
Model Release, Funding & Deals, and Research & Papers are free. Product & API, Safety & Policy, Hardware & Infrastructure, and Legal & IP are paid. Org & People, Agentic & Protocol, and General Industry are mixed, with some subcategories free and some gated.
Can the classifier get it wrong?
Yes. That's what type_confidence is for. A low-confidence classification is a signal to double check the event yourself rather than a claim that the category is certain.