Six shapes of cognitive unit

A platform team publishes a catalog of "AI capabilities." One entry is labeled RAG. Another is tool-calling. A third is critic-in-the-loop. Product managers pick from the list as if the labels described interchangeable parts. Six months later, evaluation is a mess: a retrieval wrapper is graded on calibration, a judge is graded on recall, and nobody can explain why Tuesday's dinner plan failed because the responsibility was never named.

Mechanism labels describe how work might be done. Architecture needs labels for what cognitive responsibility is being purchased: interpretation, judgment, criticism, selection among strategies, and so on. Those responsibilities recur across products. Naming them consistently is how contracts, suites, and libraries stay legible.

This page is a catalog of six shapes that share one primitive contract (what is a cognitive unit) while differing in the open decision each edge owns. Every example maps to the Weekly Meal Companion, the household dinner planner used across this collection, so names stay stable from page to page.

Helpful context: One cognitive unit, one open decision applies inside every family. Purity and the five levels describes how much reach each cognitive unit may take. This catalog helps design interfaces and evaluation before models or retrieval are chosen.

Why mechanism names fail

"RAG cognitive unit" tells a caller that documents may be retrieved. It does not tell them whether the edge interprets a request, judges practical fit, or critiques a finished plan. Retrieval might appear inside several shapes. The shape tells you which decision you are invoking.

The same confusion appears when teams name edges after model features: reasoning, structured output, multi-step chains. Those are implementation choices inside identity or policy rings (three rings of a cognitive unit). The shape belongs in the decides field and in the name, because that is what downstream code, agents, and graders must depend on. When shapes blur, compounds appear: one template that classifies, compares, assesses tolerance, and drafts a note. Each verb is a different open decision with different failure modes. Splitting until each named cognitive units owns one wrong-answerable judgment is the design discipline that prevents evaluation from mixing metrics that belong to different families.

The six families

The books describe six recurring families. They are one primitive doing six jobs, not six incompatible artifact types. Every family still returns structured results, may abstain or refer, and binds to a suite when the judgment matters.

Interpretation answers what this material or request means in this context. The Meal Companion edge InterpretWeek turns the originating household sentence into a structured week representation: busy evenings, ingredient deadlines, repetition preferences. Interpretation concerns meaning rather than whether a plan is good. Misreads still require evaluation when downstream composition depends on the representation.

Recommendation answers what should be done given explicit criteria and a named gap. RecommendMealSubstitution proposes a local repair when CritiquePlan has named a weakness. The criteria and gap are inputs. The recommendation is bounded: one substitution, not a replan of the entire week unless authority expands.

Judgment answers what is the case here under household constraints. AssessMealPracticality decides whether a candidate meal fits one evening, given busy calendars and effort budgets. This is the shape most teams reach for first when they say "the model decides." The decision must stay explicit and local (judgment should be explicit).

Selection answers which of several declared options applies, within a ceiling. Strategy choice among expertise patterns is selection: pick the Busy Week Pattern when the applicability guard passes, otherwise return to full deliberation. Selection becomes dangerous when it swallows responsibility that belongs above it, such as which outcome the product owes or which authority boundary applies.

Epistemic work answers what is known, missing, or enough to proceed. AskTargetedQuestion forms the smallest useful question when information or authority is missing. Epistemic cognitive units are not a substitute for abstaining on the subject decision when grounds are thin. They acquire or surface knowledge state without silently guessing past a hard boundary.

Judge and criticism answer how good this artifact or outcome is against declared standards. CritiquePlan identifies material weaknesses in a complete plan. JudgeWeeklyVariety, JudgeWeekdayPracticality, JudgeHouseholdFit, and JudgeLocalRepairQuality live in a judge library because calibration matters. An unevaluated judge is an opinion, not a measurement.

1. 01InterpretationWhat the request means
2. 02JudgmentWhat is the case here
3. 03SelectionWhich strategy applies
4. 04CriticismHow good is this plan

Mechanism may vary. The family states the responsibility the caller buys.

Mapping the Meal Companion

The grounding reference lists named cognitive units for the Meal Companion. The table below assigns each to a shape, serving as a consistency check when adding edges to a new spec.

cognitive unit	Shape	Open decision (short)
InterpretWeek	Interpretation	Represent this week's planning request
GenerateCandidates	Recommendation / generation	Produce candidates for declared needs
AssessMealPracticality	Judgment	Does this candidate fit this evening
ComposeWeek	Judgment (agent-led)	Coherent plan under competing goals
CritiquePlan	Judge / criticism	Material weaknesses in the complete plan
RecommendMealSubstitution	Recommendation	Local repair for one named weakness
AskTargetedQuestion	Epistemic	Smallest useful question when blocked
JudgeWeeklyVariety	Judge / criticism	Meaningful variation across the week
JudgeWeekdayPracticality	Judge / criticism	Practical fit across the schedule
JudgeHouseholdFit	Judge / criticism	Alignment with intent and preferences
JudgeLocalRepairQuality	Judge / criticism	Did the patch fix the weakness

Agent-private cognitive units (InterpretWeek, ComposeWeek, CritiquePlan) still obey shape rules. Privacy affects library placement, not the responsibility family.

Suites follow the shape

Evaluation that ignores shape produces false confidence. A judge cognitive unit needs calibration stratified by case class: busy evenings, leftover preferences, allergy edge cases. An epistemic cognitive units is graded on whether the question was the smallest useful acquisition, not on whether the household liked the final plan. A selection step is graded on whether it picked the right strategy for the terrain (judgment terrain), not on fluency.

The distinction becomes visible when attaching a suite. Ask which failure would embarrass the product if it slipped. Interpretation failures misread the week. Judgment failures pick impractical meals. Judge failures accept weak plans. Shape tells you which labeled failures belong in the suite design.

These are one primitive doing six jobs rather than six kinds of thing.

The cognitive unit · Ch. 8

Common design mistakes

Teams often pick a shape from habit rather than from the open decision. A generator template that also scores its own output mixes recommendation or judgment with criticism. The compound hides cost, breaks caching, and makes promotion impossible because two decisions share one identity version.

Another mistake is treating selection as judgment. Choosing among three complete plans that all pass hard checks is composition strategy under agent leadership. Choosing which expertise pattern to apply when guards pass is selection with a documented ceiling. Collapsing the two produces agents that "feel smart" while ownership stays untraceable.

Epistemic edges are sometimes used to bypass human judgment that should remain. Asking the household a clarifying question is appropriate. Using a question template to smuggle medical interpretation past authority is not. Shape helps audit whether the edge respects authority is granted.

A correction loop using shapes

Walk a single correction loop using shapes explicitly. The household rejects Tuesday's meal after seeing the provisional plan. CritiquePlan (judge) names the weakness: active effort exceeds the busy-evening budget. RecommendMealSubstitution (recommendation) proposes a swap that preserves accepted meals. JudgeLocalRepairQuality (judge) asks whether the patch resolves the weakness without breaking variety or household fit. Deterministic code validates schema and allergy rules, then applies the patch transactionally.

Notice what did not happen: CritiquePlan did not also generate replacements. RecommendMealSubstitution did not replan the whole week. Each call has one shape, one suite family, and one price edge. The agent orchestrates the sequence. The shapes make the sequence reviewable.

If a spec lists "meal planning AI" without these names, there is no way to attach suites, libraries, or substitution rules. Shape is the first normalization step before cognitive capability should be encapsulated.

Authority limits by shape

Each shape carries typical authority limits that callers enforce in orchestration code. Interpretation may read context grants. Recommendation may propose bounded patches. Judgment may return verdicts but not execute purchases. Selection may pick among declared strategies when guards pass. Epistemic work may ask questions but not infer medical policy. Judges may score artifacts but not rewrite them without a separate generation or recommendation edge.

Violations often appear when a generator template also "fixes" plans silently. That is acting without declaration. The fix belongs in deterministic code or in a separate recommendation plus validation pair. Shape gives reviewers vocabulary to catch scope creep during pull request review.

Library designers document shape in the public contract even when mechanism varies. Two implementations of AssessMealPracticality may use different retrieval profiles. Both remain judgment-shaped edges with the same decision sentence and fit verdict schema.

What to do next

When adding or renaming a cognitive unit in a spec, four checks keep shapes clean: state the open decision in one sentence that competent reviewers could disagree about, assign a shape by responsibility rather than by RAG or model brand, split compounds until each name owns one decision and one primary home, and draft suite categories that match the shape's characteristic failures.

If you inherit a catalog labeled by mechanism, rewrite the public contract around shapes first. Mechanism belongs inside the encapsulated implementation, not in the name callers depend on.

Read next: Cognitive capability should be encapsulated.