The principles behind every label

Deterministic where it counts

Scoring, thresholds, and label assignment are computed by fixed, versioned formulas — not left to a model's discretion. The same evidence always produces the same label.

Source independence, not source count

Corroboration is measured by how independent the sources are, not how many outlets ran the story. Syndicated or wire-sourced repeats of the same claim don't count as separate confirmation.

Duplicate resistance

Near-duplicate coverage of the same event is detected and grouped before it can inflate a signal's apparent strength.

Closed-set event classification

Every story is classified into one of a fixed set of event types. There is no open-ended, freeform category that could be quietly stretched to fit anything.

Confidence is explicit

Every label carries a confidence score — Low, Medium, or High — driven by source independence and the amount of corroborating data actually available.

Market confirmation, when available

Price and volume behavior is checked against the story to see whether the market is actually reacting. This confirmation step strengthens a label when it's available and is omitted, not faked, when it isn't.

Portfolio relevance

Signals are weighed against your actual holdings and exposure size, not treated as one-size-fits-all market alerts.

Conservative treatment of missing data

When a required input is missing or thin, the resulting confidence is capped low and the label is biased toward Watch or Healthy — insufficient evidence can never produce Avoid Adding or Opportunity Improving.

Bridge and wrapped-asset treatment

Wrapped or bridged representations of an asset are treated as a proxy for the underlying asset, not as an independent asset with its own signal history — and this treatment is disclosed on the relevant asset sections.

Conservative labels by design

The label set itself has no buy/sell language and no certainty claims built in — the six labels describe a state and a direction, not an instruction.

Outcome calibration

Past label accuracy is tracked internally and used to calibrate how the system labels future signals — this is a backend process during the beta, not yet a feature exposed in the product.

Shadow research is kept separate

Experimental scoring changes are tested against historical data internally before ever affecting a label you see — nothing untested reaches a live brief.