Model Context Layer

Pointing a model at your database does not work

The obvious version of this idea is to give a language model credentials and a schema and let it query away. It demonstrates well and fails in production, and the reason is not the query language. Your DBAs write excellent SQL. The reason is that a schema records what happened without recording what it means, so the model has to infer the rest. Which table is authoritative. Which of four revenue columns anyone actually uses. What a join across these two quietly implies. It infers well most of the time. That is the failure mode, not the success case: a figure that is wrong one time in ten looks exactly like the nine that are right, so nobody catches it at the point of use. It surfaces later, after it has already reached a position, a valuation or an investment committee.

Firms that hold a great deal of data and a capable DBA team usually understand this instinctively, which is why enthusiasm for these tools rarely survives contact with the people accountable for the numbers.

Verified operations, not generated queries

Contour takes the other route. Your data is resolved into a governed semantic model first, so the meaning is written down rather than inferred, and the MCP layer exposes a constrained set of typed, tested operations over that model: defined metrics, entities and relationships, each with known semantics. The assistant composes those operations. It does not improvise joins across a raw warehouse.

This moves correctness out of the model's guess and into platform code that is versioned, reviewed and tested in CI. It is the difference between a number you can defend and a number you have to check.

Lineage and provenance on every figure

These are two different guarantees and a defensible number needs both. Lineage is how a figure was derived: the chain of operations applied, in order. Provenance is what it was derived from: the specific records those operations read, in which source system, as of when. Lineage without provenance tells you the method but not whether the inputs were the right ones; provenance without lineage tells you the inputs but not what was done to them.

Every result returns with both. An analyst can follow any number back to source without leaving the conversation, and a reviewer can reconstruct weeks later both how a figure was derived and which records it stood on. Answers that cannot be traced are answers that cannot be used.

What your DBAs keep

Access is read-only and enforced at the interface. Row and column entitlements defined in the model are respected, so an assistant acting on behalf of a user reaches exactly what that user could reach directly and nothing further. Every call is logged with the operations invoked and the records touched. Your production systems are never queried by a model, and your DBA team retains the control surface it already has.