Why trustworthy numbers are hard

SQL generators write queries. They do not know your business, your metric definitions, or which table is the source of truth. That is the gap SEMANTIX is built to close.

Three ways a correct query still gives a wrong answer

Context is scattered. A metric definition lives in a wiki page, a spreadsheet, or one person’s head. Every question re-litigates which table, which filter, which grain. Two teams answering the same question get two different numbers, and both are defensible.

AI guesses without governance. Given a schema, a model will produce a plausible query. It will hallucinate a column, join on the wrong key, apply a filter nobody agreed to, and return confident nonsense with a chart on top.

Drift eats trust. When a number moves, the first question is always is this real? Days disappear into deciding between a genuine business change and a broken upstream pipeline. Once that happens twice, nobody trusts the dashboard anymore — including the correct one.

What grounding changes

Grounding means the answer is derived from context that a human curated and certified, rather than assembled on the fly from whatever schema was nearest.

Concretely, in SEMANTIX:

• An entity carries a description, its lineage back to the source, and a trust level a person assigned. The agent grounds in that, not in the bare table name.
• A certified metric has been reviewed and marked as cleared for decisions. An uncertified one is still visible, but it does not masquerade as a number someone signed off on.
• A business rule attached to a scope is an instruction the agent honours — for example, that cancelled orders are excluded, or that a fiscal year starts in April. Auto-captured findings arrive uncertified and have to be promoted by an owner, so nothing becomes policy by accident.
• Every trust or certification change is written to an audit log, so you can see who decided what and when.

What you get back

• Answers you can defend. “Revenue was 4.2M in Q2” traces back to a named metric, a rule set, and a query you can read.
• Fewer arguments. The definition is in one place, so it is argued once.
• Drift triage instead of drift panic. Scheduled runs and anomaly detection tell you which slice moved, so “is this real?” gets answered in minutes.

Note

Grounding is not a guarantee of correctness. A certified metric with a stale source, or a business rule nobody has revisited in a year, will still mislead you. The point is that the failure is visible and attributable rather than anonymous — which is what makes it fixable.

Continue with grounding in practice for how this works in the product, or context scopes for how it is organised.