One cognitive unit, one open decision

A team wraps "generate and critique the weekly plan" in one template because the model can do both in one call. Reviewers disagree about whether Tuesday's failure came from composition or from critique. Cost is one line item. Caching keys blur two decisions. When critique improves, composition evidence goes stale, and nobody notices until a household rejects the whole plan.

Cohesion applied to cognition rests on a simple claim: a cognitive unit that owns two decisions cannot be measured, priced, or reused as either one of them. The model's ability to perform several judgments in one pass is not permission to collapse several architectural edges into one name.

Helpful context: What is a cognitive unit defines the contract. Most decisions are several decisions explains why compounds arrive disguised as one feature. This page is the operational rule for cognitive unit boundaries.

Wrong-answerable judgment

The decision must be something competent reviewers can disagree about or get wrong. Closed arithmetic wrapped in a cognitive unit is not an open decision, it is rented reliability. A pipeline that classifies, compares, assesses tolerance, and drafts a note is several judgments glued together, even when a single model call produces all of them at once.

Compounds hide cost, hide failure, and force every caller to take the whole bundle. They make substitution impossible: you cannot swap the critique model without touching composition evidence. They make diagnosis theatrical: someone reads the entire template and guesses. The architectural response is to split until each name answers one question and stops.

Experienced architects recognise compounds before they ship. A cognitive unit name that contains "and" is suspect, because GenerateAndCritiquePlan is two owners sharing a mailbox. A template that asks two unrelated questions in one instruction block is two decisions disguised as efficiency. An evaluation suite that mixes calibration metrics with generation quality is blurring different failure modes into one score. One cache key covering unrelated inputs, practicality and variety sharing a closure measure, means neither decision caches cleanly. Reviewers arguing about which paragraph failed are experiencing a boundary that was never drawn.

None of these smells are fatal on day one of prototyping. They are fatal when the feature becomes infrastructure. A cognitive unit is one call for the same reason: above that boundary, you have traded predictability for adaptivity, and the trade is deliberate. Stronger models may collapse splits that existed for capacity, but they do not licence unnamed multi-decision compounds "because the model can." Classify by responsibility, not mechanism. Six shapes of cognitive units share one primitive but each still owns one decision. Shape tells you typical I/O patterns. It does not grant permission to merge judgments.

What this looks like in practice

In the Meal Companion, "generate and critique plan" splits into two named cognitive units:

cognitive unit	Open decision
ComposeWeek	Create a coherent provisional weekly plan under constraints
CritiquePlan	Identify material weaknesses in the complete plan

Different judgments, different failure modes, different suites. The agent orchestrates both, and neither absorbs the other's responsibility. When critique flags Thursday's meal as too repetitive, the repair passes through RecommendMealSubstitution without regenerating interpretation. When composition fails to use spinach before Wednesday, the team knows which suite to extend without rereading critique cases.

A recurring shape appears when separating compounds: closed checks on the way in, an open core, and closed checks on the way out. Deterministic code validates allergies and schema. The cognitive unit assesses credibility or fit. Deterministic code applies patches and calculates portions. That shape keeps arithmetic failures from masquerading as judgment failures.

The invoice parallel from the books carries the same structure. "Should this variance be approved automatically?" hides date arithmetic (closed), threshold comparison (closed), explanation credibility (open factual), and vendor tolerance policy (open preferential). One button, four decisions, three correct homes. Collapsing them is not simplification, it is uninspectable risk. Shipping that as one cognitive unit makes every failure a judgment failure, even when the date math was wrong.

Why the split matters

Need	Compound cognitive unit	Split cognitive units
Evaluation	One score for mixed jobs	Suites per judgment
Caching	Keys that do not match any one decision	Closure measured per class
Substitution	Replace nothing without rewriting the chain	Swap one judgment behind its contract
Diagnosis	Read every template and guess	Point at the failing decision
Pricing	One foggy bill	One call, one price per edge
Ownership	Shared blame	Named owner per edge

One cognitive unit, one open decision implies one owner for production behaviour. When Tuesday practicality fails, on-call knows which suite to check, which template version deployed, and which team owns the domain library, not which monolith function to grep. Compounds dissolve that clarity. The pager fires on "meal planning," five judgments share the blame, and the fix is "tune the prompt" because nobody can localise failure.

Cache keys must match decision boundaries. A compound that assesses practicality and variety in one call cannot cache either decision cleanly, because callers who need only practicality pay for variety recomputation and vice versa. Split edges make cache hit rates meaningful signals of closure.

Judgment should be explicit (WP Pr. 1) and one cognitive unit, one open decision are the same discipline at different zoom levels. Explicit naming at system level, single decision at edge level. Violating the latter makes the former theatrical, because names on compounds still hide several judgments.

When stronger models arrive

Strong models collapse tasks that existed because weaker models needed help. That is a gift, and a trap. The gift: fewer calls, lower latency, simpler demos. The trap: unnamed multi-decision compounds that return when the next model is weaker, the next regulator asks for evidence, or the next team needs to substitute one judgment.

The architectural rule does not forbid orchestration. An agent may call ComposeWeek and CritiquePlan in sequence, or parallel when safe. The meal agent's loop is one user-facing operation composed of many awaits. Composition is not compound identity, and each await remains a priced, measured, owned edge. Skills encode known composition patterns, agents choose among them, and neither construct replaces the rule that each cognitive unit inside the composition owns exactly one open decision.

If a stronger model genuinely retires a split, the retirement is explicit: deprecate a cognitive unit, merge suites with documented rationale, update costing tables. Explicit retirement also preserves the historical evaluation record so that future teams can understand why the split once existed and what evidence justified collapsing it. Silent merger is how compounds return. During prototyping, a compound template may be acceptable debt if tracked as debt: list the judgments inside, set a split milestone, attach no production SLA. Shipping the compound without a split ticket is how teams discover six months later that substitution is impossible. Document compound splits in the same architecture artifact as cognitive unit declarations, because future teams will not remember which demo prompt combined critique and compose.

AssessMealPracticality in a domain library is reusable because it owns one decision. GenerateAndCritiquePlan in a repo is not reusable, it is a fork magnet. Separation is the library unlock teams feel when they first publish a domain package. When splitting a compound, running old and new paths in shadow, compound versus split chain, and comparing outcomes and cost provides the evidence to split with confidence rather than on faith. Split before scale, not after trust breaks.

One cognitive unit, one open decision.

The cognitive unit · Ch. 2

What to do next

The operational habit is writing the wrong-answerable decision sentence for each named edge, rejecting closed arithmetic disguised as judgment by moving it to code, and splitting compounds until each name answers one question. Composition through agents or code preserves measurable edges rather than merging them.

What a name buys explains why identity matters. This principle explains what the identity must cover.

Read next: Three rings of a cognitive unit.