The Columna Thesis (v0.4)

A number owes you its assumptions.

To explain why, we have to start further back than anyone in this industry usually starts — not at the tool, not at the query language, but at what data is.

I. Data is something about something else

Every datum is a predication: a value said of something. The measure is what is said; the anchor — a set of coordinate dimensions — is what it is said of. A column is a measure over its anchor: the true atom of analytical meaning.

And a table? A table is a container. Form, not substance. Its keys and layouts hint at meaning — but hints are not declarations, and nothing checks them. (Others have circled this truth before: every key-value pair is a small predication; the object-database movement reached for substance over form. The instinct is old; the missing piece was the algebra.) This is why fifty years of table operations have been necessarily mechanical — you can only do mechanical things to a container, because a container has no aboutness to respect. Operations over columns are substantial: they must honor what the column says, and what it is said of.

Tables are how data is stored. Columns are what data says.

II. ColumnA: the algebra of substance

The name is the thesis: Columna stands for Column Algebra. The essence of data operation is columns and their operators — mappers and reducers — composing into expressions. Some expressions are grammatically well-formed; some are not. The algebra draws that line — and only that line. Meaning it deliberately leaves free.

What governs it is the nature of each column, captured in three anchors: the V-anchor (where its values live), the M-anchor (where they leak, and by what mechanism), the B-anchor (which directions of reduction are barred) — plus its family, which binds it to its kin under a common reducer. From these, legality follows. Whether you may sum this, roll that up, divide these two — not policy, not convention, not a style guide: theorems.

Relational algebra was the algebra of containers — select, project, join, closed over tables, indifferent to meaning. The algebra of the substance was never written. This is it.

III. The Manifold: the dataset without its costume

Strip a dataset of its arbitrary materialized form — the particular tables someone once chose, the joins those choices force, the layout accidents frozen into schemas — and what remains is the Manifold: the column set itself, with everything essential about each column and nothing incidental. All of the substance, none of the packaging.

This is why we believe the Manifold is the construct an AI agent should meet. Hand a model the tables and you hand it the distractions: containers to misread, joins to fan out, layouts to mistake for meaning. Hand it the Manifold and it faces only the essence — what each column says, of what, and how it may lawfully be treated. A mathematician’s manifold is defined without reference to any embedding space; so is ours. The tables were only ever one chart on it.

IV. The liberation: metadata independence

Metadata was never one thing. It is two kinds, fused: the load-bearing kind that governs what operations are legal and what value results, and the communicative kind that helps a person (or a model) understand. The three anchors capture the first kind completely — the operational governance layer — and in doing so set the second kind free. But because the load-bearing kind historically had no home, the industry forced names to carry it — and then built a governance regime to police the names, and called it the semantic layer. It was never semantic. It was correctness wearing meaning’s clothes: semantics conscripted into structural duty, and therefore frozen.

Language solved this before we were born: grammar decides well-formedness independent of meaning — “colorless green ideas sleep furiously” is perfect grammar and perfect nonsense — and that autonomy is what makes language both checkable and infinitely expressive. Data deserves the same constitution. Put correctness in the grammar layer — the algebra above, small, closed, machine-checkable. The moment it has a home, meaning is demobilized: names become pure communication, free to be colloquial, translated, plural, generated — because nothing operational rides on a string anymore.

Codd separated logical from physical and called it data independence. We claim the sequel: metadata independence — meaning separated from correctness so each can finally do its job. Fused, they corrupt each other; that fusion is the last fifty years. One inversion, two liberations.

V. The stakes: the analyst who never hesitates

For fifty years the two layers lived fused and scattered — some in schema structure, some in column names, some in documentation, and the decisive remainder in the analyst’s mind. The argument between human analysts was our only safety mechanism. That mechanism is gone. The AI analyst answers instantly, confidently, singularly. The debt came due the day the analyst stopped being human.

A language model is a semantic engine — magnificent at meaning, constitutionally incapable of guaranteeing correctness. Hand it today’s fused metadata and its greatest strength becomes its signature failure: it reads avg_order_value and confidently means something by it. Split the layers and the model is finally in its element — living entirely in semantics while every operation it proposes is adjudicated by a grammar it cannot touch. Hallucination, injection, drift: they can corrupt only the semantic layer. The grammar layer is the blast wall. The model proposes; the grammar disposes; the human decides among genuine alternatives. And the system’s answers come in four moods — serve, disclose, clarify, refuse — because “it depends,” said precisely, is often the most correct answer there is.

VI. What we learned about ourselves

Vigilance fails; structure holds. We made the silent-binding mistake ourselves — twice — after publishing the theory that names it. The third time, our own system caught us: asked for a metric that didn’t exist, it refused to guess. The test was wrong; the model was right. The defense must be architectural, because attention is demonstrably not enough. Even ours.

VII. The practice, and the proof

The discipline needs no product. Declare your anchors. Name your populations. Mark your materiality. Refuse, out loud, what your data cannot answer — any stack, starting today.

Columna is our proof that the discipline can be made executable — the reference implementation of the algebra, open, so no one has to take our word for anything.

If a number has ever lied to you by omission, this movement is yours.

Sign it. Practice it. Bring us the failure modes you’ve witnessed.