Language models answer fluently whether or not they’re right. Handing them the raw data doesn’t fix that: burying complexity inside a black box doesn’t solve it, it hides it.
Intelligible puts a layer of components between your data and the AI. A component is a result you can read, like a table of feature effects or the distribution of one column. The AI answers from components instead of guessing from rows, so every answer has something underneath it you can check.
Data
CSV uploads and database connectors
Components
Computed, versioned, readable
AI systems
Summand chat, or any MCP client
Something found or fitted about your data, written down so a person or an agent can read it. Summand has eleven you can pick, five of them in beta. In Elbi a component comes from a derivation, a Python function you can open and change.
Connect your data
Upload a CSV, or on Enterprise connect Postgres, Snowflake or another database. Summand copies it into its own Iceberg store; Elbi pulls from your sources on your machine.
Compute components
Tools compute what's true about the data: distributions, feature effects, correlations, fitted models. They run when data lands, when a person or agent asks, and on any schedule you set.
Keep every result
Each component is stored with the dataset it came from and the time it was computed. A re-run adds a new version and leaves the old one where it was.
Answer from the results
When a model needs to know something about your data, it calls a tool and gets back a small structured result instead of raw rows. Summand's chat works this way, and any MCP client can call the same tools.
Reuse them
The next question, app or agent reads the stored component instead of working it out again.
A semantic layer holds rules people wrote by hand, like how revenue is defined. Nothing derives them from the data. The empirical layer holds what the data actually does: what was found automatically, and what a data science team fitted and wants an agent to be able to read.
You need both, and they don’t compete. Semantic layers are standardised under Apache Ossie, and Elbi imports and exports that format. Elbi is short for Empirical Layer by Intelligible.
Agents can add to it. In Elbi an agent proposes a derivation, it runs in a sandbox, and it serves only once it passes verification, or waits for a person to approve it if you turn on manual certification.
Every answer traces back to components you can open: an EBM's feature effects, a column's statistics, the data and model that produced them. That's what makes AI deployable where transparency is required.
Two teams read the same table two ways, and an AI system invents a third. Compute it once as a component, and people and agents all read the same thing.
Most analyses run once and disappear. Components persist and re-run when the data changes, so the next analysis starts from the last one.
Components reach any AI system over MCP and stay put when you switch providers. Summand's built-in assistant runs on Claude; Elbi works with whatever model your client brings.
The proof
Nine statistical traps. Elbi’s verification gates get all nine right; claude-sonnet-5 on its own returns a different verdict on 6.
See the benchmark →