An AI agent treats every fact it reads as equally reliable, because nothing in the data tells it otherwise. A price from an hour ago and a price from last quarter look identical on the page. A figure two systems agree on and a figure they contradict each other on carry the same apparent weight. So the agent acts with the same assurance on both, and when it is wrong, it is wrong with total confidence.
A data trust score is a measured, multi-component rating of how much confidence to place in a given fact, calculated from how fresh it is, how complete its source is, how well it agrees across systems, whether its entity relationships hold, and how it matches historical patterns. It travels with the data, so both AI and people know when to act and when to verify.
What a data trust score is
Start from the right standard. Confidence is not certainty, and the bar worth holding is to act with confidence, not to wait for a certainty that never arrives. The PolyPhaze Trust Score™ makes that standard operational by measuring five components and carrying the result with the fact:
Freshness: how current the fact is, with a signal that travels alongside it.
Source completeness: whether the value came from a full source or a partial one.
Cross-source consistency: whether the systems that touch the fact agree on it.
Referential integrity: whether the entity relationships behind the fact hold.
Historical pattern match: whether the value sits inside the range history would predict.
Together they produce one rating a reader can weigh, rather than a binary clean-or-dirty flag that hides how a fact actually stands.
What each component catches
Each of the five components exists to prevent a specific kind of wrong action:
Freshness catches the stale fact: a price or balance that changed after the data was last read, so an agent does not treat yesterday’s number as today’s.
Source completeness catches the partial record: a value pulled from a feed that was missing fields, so a thin source is not mistaken for a full one.
Cross-source consistency catches the disagreement: two systems reporting different values for the same fact, so a contested number is flagged rather than picked at random.
Referential integrity catches the broken link: an entity relationship that no longer holds, so a fact is not attached to the wrong customer or order.
Historical pattern match catches the anomaly: a value far outside the range history would predict, so a likely error surfaces before it drives an action.
Trust score versus data quality score
A data quality score is usually a static, pass-or-fail flag computed in a batch job: the record is clean or it is not. A trust score is a live, multi-component rating that travels with the fact and can be weighed against the specific action in front of it. The difference matters most at the moment of action.
Data quality score
Trust score
What it measures
One dimension, often completeness or validity
Five dimensions, combined into one rating
When it runs
In a batch, periodically
Continuously, as data changes
Form
A pass-or-fail flag
A weighted score that travels with the fact
How AI uses it
Filters bad records out
Gates each action against the risk it carries
How the score moves as data changes
A trust score is not stamped once and forgotten. It moves as the data underneath it moves:
Freshness decays with time, so a fact that was reliable this morning carries a lower score by the afternoon if its source has not refreshed.
A new conflicting record drops cross-source consistency the moment it lands, flagging a fact that was solid an hour earlier.
A resolved entity raises referential integrity, so confidence climbs as the foundation reconciles more of the estate.
A value that breaks the historical pattern pulls the score down immediately, before the anomaly can drive an action.
Because the score is live, an agent checking it before acting is reading the state of the data right now, not a rating left over from the last batch run.
Setting thresholds by action
The score is one number; the bar it has to clear depends on what the action costs if it is wrong. The threshold flexes with the risk:
A low-stakes action, like flagging a record for later review, can proceed on a moderate score.
A reversible action, like drafting a recommendation for a person to approve, can clear at a middling bar.
A high-stakes, hard-to-reverse action, like issuing a large credit or releasing an order, requires a high score across the components that matter to it.
This is how one scoring mechanism governs a careful action and a routine one without a separate rule for each. The score measures confidence; the threshold encodes consequence.
How an agent uses the score
It checks before it acts. A stale freshness signal stops a retail pricing agent from acting on a price that changed an hour ago. A low cross-source consistency score routes an insurance claims action to a person instead of paying on a contested figure. A high score across all five components clears the agent to proceed without a human in the loop. The score is the gate, and it turns confidence from a feeling into a number the action can be measured against.
The same score serves people, not only agents. An analyst looking at a figure scored low on source completeness knows to confirm it before putting it in front of the board, instead of discovering the gap in the room. One measure, read by humans and agents alike, applied before the action rather than explained after it.
The examples span industries. A thin source-completeness score holds a financial services exposure calculation until the missing positions arrive. A broken referential link stops a manufacturing inventory agent from decrementing stock against the wrong part. A low pattern-match score flags an insurance claim whose amount sits far outside the norm for its type. The components stay the same; the systems and the stakes change with the industry, and each has its own industry page.
Where the score comes from
The PolyPhaze Trust Score is computed in the Knowledge Fabric, continuously, as part of resolving and reconciling the estate. It travels with the fact wherever the fact goes, and it is model-agnostic, so the same confidence measure governs whichever model happens to be driving the agent. Confidence becomes a property of the data, available to every reader, rather than something each application has to reinvent on its own.
What happens without a score
Strip the score out and an agent treats every fact as equally good, because nothing tells it otherwise. It acts on the stale price and the fresh one with the same assurance, pays the contested figure as readily as the confirmed one, and decrements the wrong part as confidently as the right one. None of these trips an alarm at the moment, because the data was technically present and the action technically succeeded. The cost shows up later, in the gap between the wrong action and the discovery, and it compounds with every action the agent takes in between. A trust score closes that gap by making confidence visible before the action rather than reconstructing it after the loss.
The failure is quietest at high volume. A claims agent paying thousands of small claims, a pricing agent adjusting thousands of SKUs, a reconciliation agent touching thousands of records: at that scale, even a small fraction acting on untrustworthy data is a large absolute number, and without a score nothing separates the safe majority from the risky few.
How the score travels with the data
A score that lived in one system would help only inside that system. The point of computing it in the foundation is that it travels: through a query, across a join, into an agent’s working context, and on to the next agent in a chain. When one agent writes a result, the confidence it acted on is attached to that result, so the next reader inherits it rather than starting blind. Confidence stops being a local property of one application and becomes a property of the fact itself, available everywhere the fact goes.
A high score is not certainty
No, and the distinction is the whole design. A trust score does not promise a fact is correct; it measures how much evidence supports acting on it now. A high score means the fact is fresh, complete, consistent across systems, structurally sound, and in line with history, which is the strongest position the data can be in. It is a measure of confidence, not a guarantee, which is why the standard is to act with confidence rather than to wait for a certainty that real data never reaches. Holding that standard is what lets an agent move at all, instead of escalating everything or trusting everything.
Frequently asked questions
What is a data trust score?
A measured rating of how much confidence to place in a fact, built from freshness, source completeness, cross-source consistency, referential integrity, and historical pattern match. It travels with the data so AI and people know when to act.
How is it different from a data quality score?
A quality score is usually a static, pass-or-fail flag. A trust score is multi-component and travels with the fact, so an agent or analyst can weigh it against the specific action’s risk in the moment.
Can people see the score, or only agents?
Both. The same score an agent checks before executing is visible to an analyst weighing whether to confirm a figure, so humans and agents act against one consistent measure of confidence.
Does scoring every fact slow the system down?
No. The score is computed as data is resolved and reconciled, not recalculated at the moment of the query, so it is available instantly when an agent or analyst checks it.
Keep reading
See what trusted data looks like on your own systems with the PolyPhaze Knowledge Fabric, request a demo →