What is not a cognitive unit

Teams reach for "cognitive unit" as a polite label for whatever involves a model this sprint. That collapses the settlement back into mechanism vocabulary. The definition is only as clear as its exclusions, and exclusions are how you prevent "any LLM call with JSON" from becoming your architecture.

Helpful context: What is a cognitive unit states the positive definition. Where the cognitive unit ends and the agent begins covers the agent boundary. This page is the fast negation list with enough depth to use in design review.

Common confusions

Not a prompt. Every cognitive unit contains a prompt the way every function contains statements. Containment is not synonymy. The prompt coaxes. The contract obligates.

Not a model. The model is substrate. The cognitive unit is the judgment edge callers invoke, measure, and substitute. Swapping GPT for Claude changes implementation. It does not rename the responsibility.

Not a chain. Three model calls in order are three cognitive units and orchestration code, not one primitive. "We only expose one API" does not merge three decisions.

Not a Skill. A Skill composes judgments to perform a known capability. Unity is the procedure, not one open decision. Meal planning as a whole is a Skill-shaped capability. Assessing Tuesday practicality is a cognitive unit inside it.

Not an agent. An agent owns a goal, loops, and may act on the world after gates. A cognitive unit answers one question and stops. Agency begins when something decides what work to undertake next.

Not a feature or chat session. Features bundle many judgments. Chat turns accumulate transcript without contractual edges. Bundling does not create decision locality.

Not deterministic code. Code that always returns the same output for the same input is not a cognitive unit, it is code. If your "cognitive unit" scores 100% on every suite case, you likely rented a model for arithmetic.

Not RAG as a label. Retrieval is a mechanism. If retrieval serves AssessMealPracticality, it lives in the operational profile or deterministic preprocessing, not as a name that hides which decision the caller is buying.

What this looks like in practice

The entire Weekly Meal Companion is not one cognitive unit. Outcome-level systems compose many edges. Retrieval of recipe candidates is not automatically a cognitive unit unless bounded as one judgment, because GenerateCandidates owns candidate production, not "everything about recipes."

Artifact	cognitive unit?	Why
AssessMealPracticality	Yes	One open decision, contract, suite
Meal Planning Agent	No	Owns goal and loop
Weekly meal planning feature	No	Bundles many judgments
Allergy validation function	No	Closed, deterministic
"The RAG pipeline"	No	Mechanism, not a decision edge
Three chained prompts in one export	No	Three cognitive units + orchestration

When someone points at a region of a file and calls it a cognitive unit, three questions settle the matter. What single judgment does it own? If the answer contains "and," split per one cognitive unit, one open decision. What does it return when grounds are thin? If the answer is "best effort," you have a prompt, not abstention or referral. Which suite describes its behaviour? If none, you have an experiment, not architecture. Missing answers mean the work lacks packaging, which is fine in discovery but not in production without a plan to close the gap.

RAG as a mechanism inside GenerateCandidates does not make "RAG" the cognitive unit. Callers buy candidate generation, not embedding search. A thin wrapper that sends text to search and returns snippets has not yet decided which judgment search serves. Tool wrappers become cognitive units only when bounded as one wrong-answerable decision with evaluation.

A cognitive unit is a judgment with a contract, not a prompt with a name.

The cognitive unit · Ch. 2

Why negative space matters

Mislabeling widens everything at once: costing tables fog, suites attach to the wrong responsibility, substitution breaks, and agents absorb judgments that remain inspectable edges. Negative definitions are how paying for reliability stays honest, because you cannot buy wrappers for a judgment you never named.

Pipelines marketed as agents suffer the same collapse. Pipelines are not agents when no goal ownership exists, and they are also not single cognitive units when they hide many decisions. A chat session accumulates transcript. It does not accumulate contracts. Messages are not cognitive units unless each turn owns a declared judgment with structured returns, which chat UIs rarely enforce. Middleware that "calls the LLM" for logging, summarisation, or routing is often several unnamed judgments plus infrastructure. Naming the judgments or admitting the middleware is experimental is the honest first step.

Product managers speak in features: "meal planning," "invoice auto-approval." Architects speak in edges: named cognitive units, human steps, deterministic gates. A feature roadmap that lists models instead of edges will reproduce prompt soup at scale. The negation list is a translation guide for those conversations. When someone says "we have a cognitive unit for that," the three reviewer questions surface maturity faster than buzzword bingo.

Recognising mislabels in practice

Sales demos collapse interpret, retrieve, reason, and act into one animated chain. Architecture unpacks the same flow into named cognitive units, deterministic gates, and agent pursuit. The demo serves selling well but fails as specification if copied verbatim into production. When a vendor says "our agent handles meal planning," the relevant question is which judgments are named, measured, and substitutable. If the answer is none, the product is a transcript generator, not Thoughtware.

The conversation that surfaces this most often goes: "Can we wrap the whole flow in one cognitive unit for simplicity?" The architectural response is: "Which single judgment does the household buy? Planning is an outcome. Inside it: interpret week, generate candidates, assess evenings, compose, critique, patch, judge variety. Each is a wrong-answerable decision. One wrapper means one bill, one suite, one owner for six failures." That response is negative space applied in real time. The goal is not to win meetings, it is to prevent uninspectable bundles from shipping because the vocabulary collapsed.

Artifact	Owns goal	Owns one judgment	Typical failure
cognitive unit	No	Yes	Wrong verdict on one question
Skill	No	No (composes many)	Procedure order or seam
Agent	Yes	No	Wrong pursuit or stop
Pipeline	No	No (often unnamed)	Unevaluated joins

Teams can adopt negative space before full library maturity: stop calling things cognitive units when they are prompts, split one compound, move one allergy check to code, add abstention to one return shape. The negation list is a checklist for incremental correction.

"We built an AI meal planner" is a product sentence, not a cognitive unit. The Weekly Meal Companion composes agents, libraries, memory, deterministic shell, evaluation, and human steps. Negative space prevents the product name from absorbing architectural precision. Mechanism labels proliferate: RAG, agentic, copilot, chain. Architecture labels are fewer and sharper: cognitive unit, Skill, agent, code. When a meeting uses mechanism vocabulary, translating to decision vocabulary before assigning owners is the habit that keeps architecture visible.

When vendors claim "agentic meal planning in a box," asking for a cognitive unit catalog, names, decisions, suites, purity levels, surfaces maturity faster than any deck. When candidates describe "building an LLM feature," asking which judgments they named and measured reveals the depth of their practice. Negative space questions are the fastest diagnostic available.

Precise negation keeps the cognitive unit primitive sharp enough to build libraries, costing, and evaluation on top. Sharp edges enable sharp ownership.

What to do next

Auditing one production "AI feature" and listing every model call, marking each as cognitive unit, agent step, or mislabeled mechanism, is the fastest way to apply these distinctions. Chains split until each cognitive unit owns one wrong-answerable decision. Closed checks move out of templates into deterministic code. Abstain/refer and suite pointers attach before the work earns the label.

See purity and the five levels and pipelines are not agents.

Read next: Purity and the five levels.