Ranking & the Semantic Layer
A source can produce thousands of findings. Severity tells you how bad each one would be if real — it doesn’t tell you which ones are genuine leads versus noise. That’s what ranking is for.
Severity is not importance. Similarity is not proof. These two rules govern everything on this page.
Importance: a second, independent score
Alongside severity, every finding gets an importance score from 0 to 1, built from meaning and context, not just detector type:
- Quality of the match — a real, readable record, or an OCR fragment?
- Detector confidence — how sure the detector was.
- Novelty — an outlier in its neighbourhood, or one of many near-identical matches?
- Context — how much genuine surrounding text supports the match.
- Severity — still a factor, but one of several, not the deciding one.
The result: a “Critical” finding repeated as boilerplate can rank below a “Medium” finding that’s genuinely unusual and well-supported. Ranking sorts by “worth my time,” not just “scary label.”
The reasons vocabulary
Every ranked finding comes with plain-language reasons explaining why it scored the way it did:
| Reason | What it means | What to do |
|---|---|---|
| Readable supporting context | The text around the match reads like a real document. | Treat normally. |
| Possible OCR fragment | The surrounding text looks garbled or scanned-and-broken. | Verify against the source first. |
| Unique evidence in this corpus | Nothing else in your data looks like this. | A stronger candidate for a lead. |
| Duplicate group | This exact match appears identically many times. | Likely a copied template — treat as one item. |
| Substantial surrounding context | There’s a lot of real text supporting the match. | Easier to judge without opening the source. |
| Cross-document recurrence | The same value shows up in several different documents, in different contexts. | A strong lead — it keeps appearing independently. |
| Common value, not discriminating | The value appears in so many assets it’s not distinctive (a shared inbox, a boilerplate field). | Deprioritise — not pointing at anything specific. |
| Known test value | Matches a documented test number or fixture (e.g. a payment-network test card). | Safe to ignore. |
| Repeated-digit pattern | Digits dominated by repetition (0000…, 1111…), typical of a placeholder. | Likely an artifact, not real data. |
| Severity noted separately | A reminder of severity, shown but not used to inflate importance. | Read severity and importance separately. |
Near-duplicate clusters
Findings whose surrounding content is essentially identical are grouped into a cluster across your entire corpus, even across sources. A cluster spanning only one source is usually a template repeated internally. A cluster spanning multiple sources means the same content is circulating between systems — which is itself worth a look, independent of what the finding is.
When ranking recalculates
Ranking isn’t a one-time judgement. It automatically recalibrates after scans finish landing, so importance scores reflect your current data, not a stale snapshot. If you want a fresh pass on demand — after a large bulk import, for example — you can trigger Reindex (re-read the content) and Recalculate (re-score importance) yourself from the Tune panel on the Fingerprints page.
Where this leads
Ranked findings feed everything downstream: the inquiries you watch, the leads proposed into a case, and what Autopilot chooses to act on. Ranking is also distinct from — and complementary to — the shared-value correlation described in Connections & Fingerprints.