Concept
Cognitive units
How cognitive units are bounded, contracted, and distinguished from what they are not.
Cover for Cognitive units19 concepts in this section.
All sections
Illustration for What a name buysShared vocabulary pays rent
10 minWhat a name buys
Engineers resist naming meetings until they see what a precise name for a judgment actually buys, boundaries, tests, libraries, and someone to blame when the answer is wrong.
Illustration for One cognitive unit, one open decisionOne open call per cognitive unit
9 minOne cognitive unit, one open decision
Each cognitive unit owns exactly one wrong-answerable judgment. If you have two judgments, you have two cognitive units, or an undeclared mess hiding cost and ownership.
Illustration for Three rings of a cognitive unitIdentity, policy, environment
9 minThree rings of a cognitive unit
Identity, policy, and environment separate what makes a cognitive unit the same cognitive unit, what callers may override, and what the runtime supplies, so versioning arguments have a place to land.
Illustration for A contract, not a promptSpecify before you prompt
10 minA contract, not a prompt
Vendor prompt templates coax models. A cognitive unit is specified as a contract, typed inputs, outputs, authority, and evaluation, reviewable like any other interface.
Illustration for Grounds are part of the answerReasons the caller can check
9 minGrounds are part of the answer
A cognitive unit returns grounds, checkable claims about the case. Reasons belong in the contract, not in model traces or fluent prose alone.
Illustration for Async functions and the cognitive unitCall, await, enforce
9 minAsync functions and the cognitive unit
Cognitive units behave like async functions with governed side effects, call, await a structured judgment, enforce deterministic post-processing. The analogy speeds adoption and clarifies boundaries.
Illustration for The cognitive unit in the systemWhere open judgment lives in the stack
9 minThe cognitive unit in the system
In a Thoughtware system, cognitive units are where open judgments live, composed by agents, packaged in libraries, bounded by deterministic code, not ad hoc invoked from every feature.
Illustration for What is not a cognitive unitEdges of the definition
9 minWhat is not a cognitive unit
A cognitive unit is not a whole agent, RAG pipeline, tool wrapper, or chat turn. Negative space keeps the edge sharp, one named judgment with a contract.
Illustration for Purity and the five levelsHow much world touches the call
12 minPurity and the five levels
Purity is five engineering capabilities, not a moral virtue. Each level from sealed to acting trades predictability for reach. declare the trade on purpose.
Illustration for Six shapes of cognitive unitsSix common forms
11 minSix shapes of cognitive units
Cognitive units recur in six responsibility families. Callers need to know which kind of judgment they are buying, because suites, failure modes, and authority differ even when the underlying contract is the same.
Illustration for Cognitive capability should be encapsulatedHide cognition behind a contract
10 minCognitive capability should be encapsulated
Encapsulation hides how judgment is produced while exposing what callers need: the responsibility, the result contract, and the evaluation identity. That separation makes substitution, sharing, and measurement economical.
Illustration for Judgment should be explicitName who owns the call
9 minJudgment should be explicit
Every open judgment should be named, bounded, and located in a cognitive unit or declared human step. Explicit judgment is how teams test, substitute, and assign ownership when models participate in decisions.
Illustration for Judgment should have localityPut the owner where the work is
9 minJudgment should have locality
One primary home per open judgment. Accidental duplicates across cognitive units, UI, and middleware produce contradictory behaviour, untraceable failure, and evaluation that argues with itself.
Illustration for Human judgment remains part of the systemSome calls stay with people
11 minHuman judgment remains part of the system
Some decisions stay with people by design for consent, consequence, and contested values. Human steps are architectural nodes, not gaps waiting for the next model release.
Illustration for Evaluation is part of architectureJudgment needs a test plan
9 minEvaluation is part of architecture
Evaluation hooks belong in cognitive unit contracts from day one. Graders, exact checks, and promotion criteria turn judgment into dependable material instead of demo applause.
Illustration for Architecture outlives implementation vocabularyStructure survives renames
9 minArchitecture outlives implementation vocabulary
Models, frameworks, and vendor terms churn. Stable architecture, including cognitive unit maps, libraries, memory forms, and authority boundaries, persists as the asset teams migrate between substrates.
Illustration for Maturity moves settled work out of runtimeSettled work leaves the loop
9 minMaturity moves settled work out of runtime
As decisions close through evidence, mature systems migrate work to deterministic code, endorsed expertise, or library promotion. Cost and variance fall without abandoning measurement.
Illustration for A cognitive unit is one callAtomicity at the boundary
10 minA cognitive unit is one call
One cognitive unit invocation means one bounded judgment, one price edge, and one reproducible extent of work. Hidden multi-step chains inside a single name break the unit of account.
Illustration for Why you cannot read a cognitive unitBlack box by design
10 minWhy you cannot read a cognitive unit
Reading a template records what was asked, not what happens. Behaviour evidence from suites replaces source-reading as the dependable account of judgment quality.