Skip to main content

Trust Score: Assessing Answer Quality

What is a Trust Score?​

Every Data Answer is independently scored on how well it is backed by your business context and your actual data. The trust score appears with the answer and tells you, at a glance, how much verified context stands behind it before you rely on it.

The score is one of three levels:

ScoreWhat it means
ConfidentThe elements the answer's correctness rests on — definitions, filters, joins, and business rules — were verified against your business context, data, and memory.
Needs AttentionThe answer is likely usable, but something could not be fully verified or deserves a closer look before you rely on it.
Review RequiredA verified problem was found — for example, the answer conflicts with verified context — or Mantra could not produce an answer at all.

A lower score doesn't always mean the answer is wrong. It means less of it could be verified; the report tells you exactly which parts so you can judge for yourself.

How Scoring Works​

After Mantra produces an answer, it independently scores the answer in the background. Scoring identifies the answer's load-bearing elements and assumptions, and verifies each one against your verified context, the actual values in your data, and any guidance you've given before. Every reason for the score cites the evidence it was checked against.

Because scoring runs after the answer, you'll briefly see Analyzing trust score while it completes; the score replaces it when the report is available.

Mantra composes answers only from data it actually discovered in your environment. Scoring independently re-verifies this: every table and column the answer references is checked against your catalog, with your access control applied; anything that can't be confirmed is flagged in the report and can lower the score.

Reading the Report​

Select the trust score to open the full report. At the top you'll find the score, a plain-language summary, and the main reasons behind it. Below that, the report gives a more granular view of the evidence, covering areas such as:

  • The business rules and definitions that applied to your question, and whether the answer honored them
  • The data used — which tables and columns the answer drew on, and whether the relationships between them are grounded in known joins or were inferred
  • The filter values in the query, checked against the values that actually exist in your data
  • Your own input — clarifications you answered and guidance you've given, and whether it was applied to this answer
  • How well the answer's explanation is supported by the query and its results
Live data verification

In environments where live verification is enabled, filter values are checked directly against your data so the report can tell you not just that a value was unverified, but that it was confirmed present or found to be absent.

When Mantra Can't Answer​

If Mantra declines to produce a SQL-backed answer because the available data and context can't support the question, the report explains the gaps in its attempts to answer: what data and context were found, and where the answer stalled: a concept with no backing data, an ambiguous term, a missing definition, etc.

This makes a non-answer actionable: often the fix is adding the missing context to the 360° Context Hub, rephrasing the question, or providing feedback directly. The report points at what would be most useful.

Improving Answers with Feedback​

The trust score is also a way for you to teach Mantra. From the report you can:

  • Agree or disagree with the score, with a note explaining why
  • Correct any finding, or supply what's missing when the report identifies a gap — for example, a definition or rule Mantra didn't have
  • Confirm a relationship Mantra inferred, so it's treated as reliable going forward

When you submit feedback, the answer is re-scored with your input taken into account.

Your guidance is remembered​

Corrections and confirmations become personal guidance: they shape how Mantra answers your future questions and scores your future answers, without you having to repeat them. Reports show your guidance alongside the evidence, including whether it was applied to the answer at hand.

Your guidance is personal — it applies to your answers, not your whole organization. When your guidance conflicts with a shared business rule, your words win for your answers, and the report marks the rule that is superseded. Shared business rules themselves are managed in the 360° Context Hub.