Home › All guides

Enterprise Intelligence Architecture

The system a model runs inside, across four layers. Eight areas, each derived from peer-reviewed sources and each naming them. Free to read, in full, without an account.

Last reviewed: 2026-09-27 · Українською

How the four layers fit together

Read bottom-up. The order is a dependency chain, not a menu — each layer can only be as reliable as the one beneath it. That is why platforms that begin at layer 4, an agent pointed at a warehouse, tend to fail in ways that look like model problems and are not: the model is answering correctly from a substrate that cannot tell two customers apart. Each area below stands on its own and names the peer-reviewed work it derives from.

Where this comes from

Every one of the 24 lessons behind these eight areas cites the peer-reviewed work it derives from, and each area ends with its own source list. The reading corpus behind the track runs to 1834 papers, of which 1268 are mapped to specific areas. Where a statement rests on something the corpus does not hold — a standard, a specification, a project's own documentation, an article of law — the area names that primary source as well, in a list of its own.

Layer 1 — the substrate

Everything above inherits this layer's latency and its failure modes. The decisions here are physical: how many bytes a query reads, how many of them cross the network, and how evenly the work divides. A stage finishes at the speed of its slowest task, never its average one, so skew rather than raw volume is usually what breaks a pipeline. Get this wrong and no amount of modelling above it recovers the response time.

Layer 2 — meaning

Storage holds rows; it does not hold meaning. This layer decides what a word denotes — whether “customer” is an account, a billing entity or a person — and whether two records describing the same organisation are recognised as one. Without it every query author reinvents the definition privately, and two dashboards answer the same question differently with neither being wrong. Here a platform stops being a filesystem and becomes a shared vocabulary.

Layer 3 — control

A platform that cannot say who may read a field, or where a number came from, is unusable anywhere its answers carry consequences. This layer covers both halves: policy evaluated at query time rather than frozen into copies, and lineage that survives transformation so a figure can be traced back to its source. These are the properties auditors ask for first and the ones retrofitted last, at the highest cost.

Layer 4 — reasoning

This is where models finally act, and where they inherit every constraint settled below. An agent reasoning over an unresolved entity graph will confidently merge two companies into one; an agent without lineage cannot defend its answer when challenged. The areas here cover how a decision is represented so it can be reviewed afterwards, and how several agents reach agreement without a single point of failure.