Why you cannot read a cognitive unit

Read the body of a function and you know what it does. Code review rests on that claim for deterministic work. Open judgment breaks the habit without removing the need for trust.

A template for AssessMealPracticality reads plausibly. Busy evenings stay under twenty-five minutes of active effort. Reasons cite constraints. Nothing in the text guarantees Tuesday behaviour on labeled weeks. Move the same template to a smaller model and behaviour shifts. Leave the template unchanged while the provider updates weights under the same version label and behaviour shifts without a repository diff.

Reading records hope. Suites record behaviour. This intermediate note is the epistemic guard for architects: you cannot read a cognitive unit and know whether it performs. You can measure it on labeled cases bound to identity.

Helpful context: Evaluation is part of architecture requires suites in contracts. Reliability is not a confidence score explains operational trust. A cognitive unit is one call defines the boundary being measured.

The gap reading cannot close

Ask a concrete question: how often does AssessMealPracticality correctly flag an over-ambitious meal on a busy Tuesday?

Nobody answers from the template alone. Reviewers may agree the instructions sound reasonable. Reasonableness is not calibration. Households experience wrong answers anyway.

The gap between request and behaviour is the whole of evaluation engineering. No amount of careful template authoring eliminates that gap, because the same words processed by a different model, a different context length, or a different decoding configuration produce different behaviour on the same labeled case. The gap is a structural property of open judgment that makes measurement the only dependable account of quality. Deterministic code closes part of the gap with exact tests. cognitive units close the remainder with graders, labeled cases, stratified figures, and promotion rules tied to identity versions. Callers who need certainty must move work to code or to eval, not to longer prompts.

Fluency is not evidence. Confidence theatre fills UI with scores that measure model state, not fit for purpose. Reading is the architect's version of the same trap. Polished prose feels like due diligence.

1. Read the templateHope, not behaviour
2. Run the suiteMeasured behaviour
3. Bind to identityTemplate + model + date

Source-reading was never the account of behaviour. Suites are.

What review still does

Review remains valuable for clarity, hole design, abstain paths, and policy language. Review catches contradictions and missing authority limits before expensive measurement.

Review does not replace measurement. Treat review as authoring discipline. Treat suites as behaviour truth for promotion and incident response. When someone says "we read the prompt carefully," translate the claim into suite coverage and figures, or mark the capability experimental in the architecture record.

Identity binding

Figures attach to identity: template version, model pairing, evaluation date, and wrapper declarations when used. Substrate shifts under stable templates are identity events. Teams remeasure on schedule even when diffs are empty.

Evidence dies when the template changes when edits invalidate baselines without rebuilding suites. Binding prevents silent drift from masquerading as stability.

Old habit	Thoughtware replacement
Read the function	Read the suite record bound to identity
Diff the patch	Remeasure after identity or model change
Ship on review alone	Ship on measured behaviour for classes that matter
Trust demo weeks	Stratify labeled cases from terrain

What this looks like in practice

The Meal Companion team maintains labeled weeks: two busy evenings, spinach deadline, leftover preference, temporary pasta avoidance that must not become durable knowledge.

Before promoting AssessMealPracticality v2, they run suites stratified by case class. Busy-Tuesday accuracy must meet tolerance. Allergy-edge cases must not regress. Variety interactions must not worsen.

A vendor updates the default model. Template diff is empty. Scheduled remeasurement shows busy-Tuesday accuracy dropped. Promotion halts. Incident response targets identity and pairing, not household "pickiness." JudgeLocalRepairQuality receives similar discipline. Loop compression without judges is blind patching. Reading critique templates does not prove patches preserve accepted meals.

A template records a hope.

The cognitive unit · Ch. 9

Implications for libraries and substitution

Encapsulation enables substitution when figures travel with library entries. Reading a release note is insufficient. Consumers check published suite results for their case classes.

Substitution without remeasurement is faith-based deployment. Architecture teams treat unexplained behaviour shifts under stable templates as first-class incidents with identity versioning, not as random noise. Tuning by wrapping adds wrappers while preserving evidence chains. Wrappers change behaviour too. They receive measurement, not exemption because someone read the wrapper prompt once.

Intellectual humility as engineering discipline

This note sounds skeptical. The skepticism serves reliability. Products judged like people (products judged like people) fail in slow motion when teams trust prose over figures.

Accepting unreadability of static inspection pushes evaluation earlier in design, where Principle 6 wanted it. It pushes stratified cases, not aggregate demo scores. It pushes honest abstention when grounds are thin.

Common mistakes

Legal review substituting for suites. Compliance reads policies. Behaviour still drifts.

Prompt diff as release gate. Text unchanged therefore safe. Substrate changed.

Single demo path. Suites cover happy weeks while production lives elsewhere.

Anecdotal support as eval. Tickets are signals, not graders.

Incidents without readable sources

Production incidents on cognitive edges start from suite history and traces, not from rereading templates. Ask which identity version ran, which model pairing applied, which case class failed, and whether substrate changed without a diff. Those questions orient remediation.

When teams default to "fix the prompt," they often edit identity without rebuilding baselines. Behaviour shifts for callers who never saw a semantic version bump. Treat prompt edits that change behaviour as identity events with remeasurement, even when reviewers prefer to call them tweaks.

Reading remains useful for authoring discipline: hole design, abstain language, authority limits. Reading fails as release gate. Principle 6 and this note together push measurement earlier until teams internalize that templates are hopes and suites are facts.

Pairing with reliability work

Reliability investments attach to named edges. Double-sample high-consequence classes on AssessMealPracticality. Single-sample low-consequence screening. Those choices appear in cost models because atomicity makes edges visible. Paying for reliability depends on knowing what was bought.

When reliability work wraps unnamed compounds, spend rises without traceability. Incidents still end in prompt edits because no edge identity exists to remeasure. Atomicity is the prerequisite for principled reliability spending.

Habit change for code reviewers

Train reviewers to ask for suite diffs alongside template diffs. Ask which case classes moved. Ask whether identity semantic version bumped. Those questions feel unfamiliar for teams from deterministic codebases. They become natural after the first regression caught without a template change.

Reading templates remains appropriate for tone, abstain wording, and policy clarity. The habit change is not abandoning review but reframing what review accomplishes. Review catches intent errors, policy contradictions, and missing abstain paths before measurement begins. Measurement catches behaviour errors that intent alone cannot predict. Separate reading habits from release gates consciously so both can coexist without confusion.

What changes in the next track

This note closes the cognitive unit track in the curriculum. The Skills and agents track asks how named cognitive edges compose into procedures and goals. Decide, perform, and pursue describe responsibilities, not sizes. Measurement does not get left behind. Skills inherit suite expectations from the edges they call.

Readers arriving from this page carry three habits forward: name open judgments, bind suites to identity, and treat unreadable templates as normal rather than as emergencies. Measurement culture survives the track boundary when teams keep labeled cases alongside feature backlogs.

What to do next

Bind suites to identity: template version, model pairing, date. Remeasure after edits, wrapper changes, and scheduled substrate checks. Prefer tuning by wrapping when improving reliability without discarding evidence chains, and treat unexplained shifts under stable templates as identity events requiring figures.

This sets up the Evaluation track: evaluation as engineering. For template edits that invalidate figures, see evidence dies when the template changes.

Read next: Decide, perform, pursue.