Confidence and provenance
How HermesAI shows you where every claim came from and how trustworthy the cluster is.
Every draft HermesAI produces carries two kinds of editorial signal:
- Provenance — which source supports which claim.
- Confidence — a weighted read on how solid the cluster is.
Together they let you review at speed without losing rigor.
Provenance, in plain terms
When HermesAI synthesises a story from multiple sources, it tracks which source supports each substantive claim. You see this attached to the draft in two ways:
Source list
Every outlet whose coverage went into the cluster — name, URL, tier, recency.
Claim-level attribution
Click any sentence to see which source supports it. If a sentence has no support, that itself is a signal — usually means the AI overreached and needs editing.
How confidence is weighted
The confidence read is a weighted signal, not certainty.
Prop
Type
The output is a normalised number, displayed as High / Medium / Low in most surfaces, with the underlying number visible when you need it.
What confidence is not
Confidence is not certainty
A high confidence read means HermesAI saw multiple credible sources agreeing. It does not mean the story is true beyond doubt.
A low confidence read means either few sources cover it or sources disagree. Either way, slow down — but don't auto-reject.
The editor still does the call. Confidence is a triage aid, not a publish-or-don't switch.
Conflict markers
When sources disagree on a substantive fact, HermesAI surfaces both versions instead of picking one quietly.
In the draft body
A small marker near the claim, with both versions on hover or click.
In the source list
Sources tagged as conflicting on this cluster, so you know who said what.
In the confidence read
Conflicts pull the confidence read down, which is exactly the editorial signal you want.
If the AI ever 'smoothed over' a conflict — picked one version without flagging — you'd never know. HermesAI is built so this doesn't happen.
Source tiers
Prop
Type
Tiering is set in the curated catalog. You don't change tiers per source — that's part of the catalog's quality control.
How to read these signals as an editor
Glance at confidence
High → solid ground. Medium → check the conflict markers. Low → slow down, look hard.
Scan the source list
How many sources, what tiers, how recent. If it's all tier 3 from the last hour, treat as preliminary.
Spot conflicts
If conflict markers exist, read both versions before deciding.
Click claim-level attribution
For any sentence that feels too confident, check what's supporting it. If there's no provenance, edit.