Select a confidence tier to explore the design decisions
High confidence
75% and above
Borderline
60 – 74%
Low confidence
Below 60%
Example news item
Positive
NVDA beats Q2 earnings forecast by 18%
Reuters · 2h ago
Confidence
92%
Design decisions at this tier
Green bar — full width Perceptual signal that model output is reliable. Analysts can move faster with less cognitive load on verification.
No additional warning per item Adding a warning at 92% would be noise. The advisory banner at the top of the dashboard handles the persistent reminder to verify — it doesn't need repeating on every card.
Ranked higher in feed by default High-confidence signals surface first, respecting analyst time while still exposing lower-tier items below.
Analyst workflow at this tier
Scan headline→ Read sentiment pill→ See green bar→ Add to watchlist→ Spot-verify source
Example news item
Negative
Supply chain disruptions hit semiconductor stocks
FT · 4h ago
Confidence
65%
Design decisions at this tier
Amber bar — a visual pause Amber is universally understood as "proceed with caution." It doesn't stop the analyst but introduces intentional friction — a moment of hesitation before acting.
Surfaces in "low confidence" filter Items 60–74% appear in the borderline filter view, letting analysts batch their deeper review work rather than context-switching per item.
Hover tooltip: "mixed sentiment language detected" Surfacing the model's uncertainty reason helps analysts know where to focus their verification — not just that confidence is low, but why.
Analyst workflow at this tier
Scan headline→ See amber bar→ Read full article→ Check hover reason→ Manual classification
Example news item
Negative
Regional bank exposure to CRE raises concerns
Barron's · 6h ago
Confidence
54%
Design decisions at this tier
Red bar — but item stays visible A critical decision: low-confidence items are NOT hidden or suppressed. A 54%-confidence negative signal on a major sector is still analytically relevant. The red bar flags it without removing it.
"Add assessment" action available Analysts can manually reclassify low-confidence items. Corrections feed back into the model as labeled training data, closing the human-in-the-loop.
Batched in a dedicated review queue Low-confidence items queue for a scheduled review session rather than demanding immediate attention, protecting analyst focus during high-volume periods.
Analyst workflow at this tier
See red bar→ Skip or queue→ Scheduled review→ Add assessment→ Correction logged
Anti-patterns considered and rejected
Hiding low-confidence items entirely — analysts lose visibility into model blind spots
Showing a raw percentage without color encoding — too much mental math per article
Blocking action until analyst verifies — creates bottleneck, destroys efficiency gain