Credibility & Validity

Krataxis is built around graded confidence. An item is rarely simply "true" — it is reported, then corroborated, then confirmed. Two distinct measures express this, and it's worth keeping them apart: the validity score (a continuous number on one item) and the credibility tier (a discrete rung a cluster has climbed). Read together, they tell you how much weight a finding can bear.

The validity score

The validity score is a single number from 0.0 to 1.0 describing how trustworthy an individual intel item is. It's a blend of several independent factors:

Two consequences are worth internalising:

• Source reliability is earned, not assumed. A source analysts repeatedly confirm becomes more trusted; one repeatedly refuted decays. The platform's judgement of your sources sharpens the longer you use it.
• There is a ceiling without a human. Automation alone can take an item only so far — full confidence requires analyst verification. However well-corroborated, a machine-scored item never quite reaches certainty until a person signs off.

The credibility tiers

Where the validity score grades one item, the credibility tier grades a cluster — the group of items judged to describe the same real-world event. As corroborating sources accumulate, the cluster climbs a ladder:

The number of independent sources reporting the same thing is the strongest signal the system has — which is why clustering (merging duplicate reports of one event) is so central. Ten copies of one press release are not ten sources; the platform works hard to tell the difference.

Some sources arrive already corroborated. A structured news-event source like GDELT knows how many distinct outlets carried the same event, and the platform seeds that event's corroboration from the outlet count directly — so a globally-reported incident enters the picture already standing on many independent legs, rather than accumulating them one duplicate at a time. The same discipline applies: it is distinct outlets that count, not raw article volume, for exactly the press-release reason above. This is what lets the platform lean a credible, widely-reported event apart from a single, possibly motivated claim from the moment it arrives.

Deception flags

Corroboration can be manufactured. A coordinated campaign can push many "sources" repeating the same false claim, which would otherwise inflate a cluster's tier. The deception flag is the countermeasure: when a cluster shows signs of coordinated, manufactured consensus, it's flagged — and the flag acts as a hard ceiling.

A confirmed deception flag forces a cluster back to REPORT, regardless of how many sources back it. No amount of corroboration can promote an item the system believes is being gamed. Deception flags surface in the Deception Flags panel (the header badge carries the unread count), where you can review the cluster and record a verdict.

How to read this as an analyst

• Carry the tier with the claim. "15 strikes were reported" and "15 strikes are confirmed" are different epistemic states — never flatten them.
• Check both measures. A high validity score on one item is not the same as a high credibility tier on the cluster.
• Be suspicious of a REPORT sitting on heavy corroboration — it may be deception-flagged, not merely thin. Open it and see.
• Final authority is human. Automation grades the evidence; your confirmation is what moves a finding to CONFIRMED — and what the audit record stands on.

Where to next

• How Intelligence Flows — where scoring and tiering sit in the pipeline.
• Escalation & Predictions — what the platform does with graded intel once it has it.