AtomicDB organises everything it holds in one hierarchy: quadrant → knowledge area → model → concept → item. It is how data enters, how access is granted, and how a question knows where to go. Above it sits a second hierarchy — the AI tiers that reason over it.
Follow a single customer number from the top of the AtomicDB grid to the item every record points at. Hover a level below the diagram to see what it holds.
The broad sector a body of knowledge belongs to. An organisation sees its data at a glance and confines a question to the quadrants it ticks.
The scope and the security boundary. Access is granted per knowledge area, to groups — so a case, a client or a department can be sealed off by the engine itself.
One source, as it arrived: a spreadsheet, a table, a whole database, a mainframe file. Sources are never altered — a merged view is just another model.
What a column means. Concepts that share a name across models are bridged automatically — that is where joins come from.
A distinct value, stored once. Every record that mentions it holds an association to this one item — so a correction here is a correction everywhere.
Model, concept and item is the only way data enters the AtomicDB grid — for imports, repairs and migrations alike. So the structure an application declares is exactly the structure the engine holds.
The model and its concepts are created first: what this source is and what each column means.
Rows arrive and pass the verification gate. A value that breaks a rule is set aside for a person, never trimmed.
Every value is placed once and every record is linked to it. Only now is the import complete — and searchable.
Above the data sits a reasoning hierarchy. A language model may propose; the records decide. Rules are rows, an obligation is accepted by a person, and any question can be asked as of a date. Everything runs on the machine — nothing leaves it.
Plain-language questions become a walk of the grid. Every answer carries its evidence table and the exact text the model was given.
Paths, clusters, anomalies and the same thing under two names — each conclusion records the chain of relationships behind it.
Law and obligations held word for word as data. A verdict names the provision it applied; a violation is set aside for review.
A 2:46 narrated walk through Ask, ATLAS and AXIOM on the same question — and why the same question on the same date always returns the same evidence.