Scores & Classifications
Every capability attaches scores and classifications to its records so you can threshold the flow instead of reading it. The vocabulary is shared across the platform — this page is the glossary. Exact fields, ranges, and filter parameters for each endpoint live in the capability pages; what follows is what the signals mean and how to use them well.
Read this first: what a score is
Most Ultralayer scores are model-estimated judgments — an LLM assessing a record against a rubric at ingestion time. That design is deliberate: "how important is this?" and "how credible is this source?" are judgment calls, and a calibrated model applying one rubric consistently at wire speed does that job better than any fixed keyword rule. But it means you should hold the numbers correctly:
- They are estimates with variability, not measurements. Two near-identical items may score 0.58 and 0.64. Nothing meaningful separates them.
- Use them as thresholds and rankings, not as precise values. "Importance above 0.5" is a robust query; "sort by importance and trust the third decimal" is not.
- Semantics are consistent; distributions vary by product. Importance means "how much does this matter to markets" everywhere, but the score distribution over wire items differs from the one over developments. Calibrate per capability, per the guidance in each skill page.
The exception: cosine_similarity on search results is a mathematical distance between embeddings, not a model judgment. It is stable and precisely comparable — within one query's results.
The scores
| Score | Range | What it estimates |
|---|---|---|
| importance_score | 0–1 | How much the item matters to markets. The unfiltered flow is dominated by low-importance items — this is your primary noise gate. |
| trustworthiness_score | 0–1 | Credibility of the source — authority and attribution quality. Orthogonal to importance: a thin anonymous account can print important claims, and a top-tier wire prints routine ones. |
| utility_score | 0–1 | How informative a retrieved passage is — dense guidance language scores high, boilerplate scores low. Used in semantic search. |
| sentiment (entity-mention) | −1–1 | Tone toward one specific entity in one item. The precision tool for "negative news on X". See entities. |
| sentiment_score (passage) | −1–1 | Tone of a passage as a whole — not entity-specific. Same name family, different object; don't conflate the two. |
| impact_score | −1–1 | Signed exposure of one stakeholder to one development or situation: direction (winner/loser) and magnitude. See developments and stakeholder identification. |
| confidence_score | 0–1 | The model's own confidence in an assessment it produced. Low confidence is information: thin or ambiguous evidence. |
| surprise_score | 0–1 | How unexpected a development is relative to the prior narrative. Surprise, not importance — a scheduled CPI print can be maximally important and minimally surprising. |
| cosine_similarity | 0–1 | Embedding similarity between your query and a result. Mathematical, not model-judged. |
The classifications
Alongside the scores, records carry categorical labels. The ones that do the most work:
new · update · duplicate · correctionduplicate removes syndication copies; correction is rare but market-critical.confirmed · reported · rumor · speculation · forecasttrustworthiness_score, measures a different thing. See sources.news · analysis · opinion · press_release · data_releasepress_release + first-party sourcing = issuer originals.realtime · report_of_past · report_of_futurereport_of_future finds scheduled events before they happen.first_party · reporting · second_handprimary · significant · lowerPractical thresholding
Rules of thumb that hold up across the platform — starting points, not laws:
- Meaningful usually starts around 0.4–0.5. Below that, importance and utility scores mostly mark routine flow. A floor of
0.5on importance is a sensible default desk filter. - 0.7+ is a strong-signal band. High-importance wire scans, dense evidence passages, and trustworthy-source filters all live around here.
- Start loose, then tighten. Aggressive floors silently return empty results — a query with
min_importance_score: 0.9isn't wrong, it's just usually empty. Loosen until you see the distribution, then cut. - Combine orthogonal signals instead of maxing one. Moderate floors on importance and trustworthiness and a novelty filter beat a single extreme threshold on any one of them.
- For signed scores, filter both tails. Strong conviction in either direction (
≥ 0.3or≤ −0.3on sentiment or impact) is usually the interesting set; near-zero is noise. The two-sided filter syntax is documented per endpoint.
Each capability page gives concrete, tested threshold recipes in its use-case examples — those are the reference for real queries.