Foundations
Foundations
Most arguments about analytics are arguments about tools. This page is not.
Underneath the tools are questions nobody has settled: what an analytical thing actually is; what makes it reasonable to rely on an answer; and what changes when the thing doing the reasoning is not a person. They are older than our work and larger than it.
We work on them for a practical reason. Governance cannot be stronger than its account of the thing being governed — and a machine cannot enforce a distinction it cannot represent. Better answers about analytics eventually require better foundations underneath them.
This is where we currently think those problems are. It is a map of terrain, not a table of contents, and it can change.
Data · Certainty · Intelligence
Three enduring problems that analytical work has to contend with — not an exhaustive account of analytics, not three stages of anything, not three equally developed theories, and not three things we claim to have invented. They are uneven, and the unevenness is visible below rather than smoothed out.
Data
What is the analytical thing we are working with?
What it opens
- What gives analytical data its identity — what makes two numbers the same thing, or different things?
- Which transformations preserve that identity, and which quietly replace it with something else?
- How can an analytical object stay derivable and consistent as it moves through systems that know nothing about each other?
Where we are
This is the most developed of the three. Identity, existence and lawful derivation have a worked account with proofs behind them, and a query language exists that can only ask for meaning a declared model can defend. The container got its mathematics half a century ago, in the relational model; what the container was carrying did not.
Codd's relational model (1970) is the reference point here — cited as the thing that was settled, not as work datumwise claims or extends.
Ways in
- When is it data? a four-minute walk, from a single number on a plinth
- The Theory of Data in One Afternoon the whole idea in one sitting, no prior reading
- The Theory of Data the full account, with the proofs — readable here
- The Three Worlds of Analytics why business meaning, data, and material data need different governance — readable here
What is unsettled
Absence is only definable relative to a domain — you cannot say a value is missing without saying missing from what. Where that domain comes from, when nobody has declared one, is not settled.
Certainty
What gives us sufficient grounds to rely on what we know?
What it opens
- What is actually carrying the confidence — knowledge of how something works, a model of how someone chooses, or observed regularity?
- How far does a ground reach, and what happens when confidence earned on one is spent as though another had been established?
- What survives the passage from evidence to inference, to claim, to decision, to action?
Where we are
Substantial, and recent. The families of ground are distinguished and their limits described; the passage from governed evidence to a licensed claim has an account of its own. The word is used narrowly here — being sufficiently sure, on warranted grounds, for the reliance at issue — and not as a philosophical theory of knowledge.
Ways in
- The Ground for Certainty the current foundational publication — readable here
- The governed passage from evidence to a claim that is licensed to travel
- The Statistical Bridge the full account — readable here
What is unsettled
How far observed regularity warrants anything beyond the conditions it was observed under. Behavioural evidence is the ground most often spent past its reach, and the boundary is easier to state than to locate.
Intelligence
How do intelligent actors understand situations, form purposes, reason, choose, and act?
What it opens
- What changes when intelligence itself becomes part of the analytical machinery, rather than the thing operating it?
- How should a person work with an intelligent agent — and a machine with another machine?
- When an actor can explain a principle, reason about it, and act against it, what have we actually established about what it will do?
Where we are
The least settled of the three, and we would rather say so than fill the space. One thing is held firmly, and it is a boundary rather than a theory:
Interpretation is useful. Interpretation is not authority.
A person or a machine may interpret, propose, explain and clarify. Execution requires governed authorisation. That is a line we are willing to defend and to build to; it is not an account of how intelligence works.
Ways in
- Never let your agent touch the database the position, and what to put between them
What is unsettled
Almost all of it, and it is worth being specific rather than vague: what an agent’s proposal actually is; what makes a delegation legitimate; what one machine owes another when it hands over a result; and what grounds could warrant relying on an actor whose goals and self-regulation are not established. There is no datumwise theory of intelligence, and we are not writing one to complete a pattern.
How these relate
Not as a sequence, and not as three parts of one thing. They are three different questions, which is why they can be at three different stages of being answered. Nothing here says a result must pass through them in order, and nothing here says that together they cover the subject.
Where they do meet, they meet as different intellectual jurisdictions — different questions each account claims to govern. That is a statement about what a body of work answers for, not about what a running system is permitted to do; those are two senses of the same word and only the first is meant here.
Analytical Governance sits beside them, not inside them
Analytical Governance is the discipline governing the legitimacy of the analytical service — whether the purpose a person brought survives the passage into a request, a result, an answer, and a consequential use. It necessarily runs into foundational problems in all three territories, and it is filed under none of them: it is not a component of Data, an application of Certainty, or a policy about Intelligence. It has its own account, and its own doorway — reached through Architecture, the shelf for work that governs a passage rather than establishing what exists.
Work that is not ours
These are shared problems with long histories, and where we lean on work that is not ours we name it and link it. A link here is context, not endorsement: it does not make that work part of what datumwise holds, and it carries no datumwise authority.
Every deposited work, with its current version, date and record, is on the research corpus page. This page is the discussion; that page is the record.