Skip to content

Four Closures

The four closures are a way of thinking, not just a closeout schema. You use them to figure out why an AI output feels wrong when nothing has actually failed.

A wave is closed only when all four are explicit. Three out of four isn't closed.

The Four Dimensions

DimensionQuestion it answers
AuthorityWhere does truth for this surface live, and is it internally consistent?
SemanticAre the contract's required fields, failure modes, and complex cases all explicit?
ConsumerDo the systems / readers that consume this seam use it correctly?
Drift ResistanceCan the next implementer easily walk this design back?

Authority Closure

Where does truth for this surface live, and is it internally consistent?

Authority closure asks: after this change, is there exactly one canonical owner? Has any parallel truth surface (shadow caches, app-local convenience state, dual-read paths) been eliminated or explicitly admitted?

Closes whenFails when
Canonical owner is uniqueTwo surfaces both claim authority
No parallel truth remainsAn informal cache has become canonical
Admitted alterations are explicitSilent owner-cut reopens occurred

Common authority failure: The AI added a helper that effectively re-implements behavior owned by another package, creating shadow truth.

Semantic Closure

Are the contract's required fields, failure modes, and complex cases all explicit?

Semantic closure asks: would two different implementers of this contract produce equivalent behavior? Are MIME types, error reasons, retry policies, and edge case behaviors specified, or were they "left to the implementation"?

Closes whenFails when
Required fields specifiedRequired field omitted
Failure modes pinnedHappy-path-only
Complex cases not deferred as assumptions"Edge case" parked silently

Common semantic failure: Happy-path-only — the AI implemented what success looks like and left failure handling implicit.

Consumer Closure

Do the systems / readers that consume this seam use it correctly, and can shortcuts not silently become mainline?

Consumer closure asks: does the actual consumer (the app, the docs reader, the operator) experience this work as solving their problem? Or did the contract land cleanly while the consumer remained broken?

Closes whenFails when
Normal consumers use the correct seamConsumers still use the old seam
Shortcuts cannot silently become mainlineA "temporary" shortcut becomes the de-facto path

Common consumer failure: Build-pass / consumer-fail. The public docs topic that produced the closed-then-reopened cycle is a textbook example: every machine-side gate passed; the actual reader rejected the docs as "audit artifact prose."

Drift Resistance Closure

Can the next implementer easily walk this design back?

Drift-resistance closure asks: are forbidden shortcuts named, are reopen conditions explicit, are the boundaries that prevent regression structurally enforced or only socially enforced?

Closes whenFails when
Forbidden shortcuts are explicitShortcuts unnamed and re-discoverable
Future implementers cannot easily thin the designRe-thinning available without explicit admission

Common drift failure: A closeout that passes today but leaves no machinery to prevent the same shortcut from recurring next quarter when context is gone.

Why All Four Matter Independently

Three-of-four passing is the most common false closure in AI work. Each combination of "three pass, one fails" produces a recognizable pathology:

AuthoritySemanticConsumerDriftResult
Works today, rots in six months
Technically correct, unusable
Good UX, unspecified contract
Parallel authority hidden

The methodology requires all four to be explicit. A closeout that omits a dimension is a schema violation, not a soft signal.

Reader Scenario: Reasoning About A Wave Through The Four Closures

You're a manager judging a wave's closeout. The worker says the wave is done. You walk the four closures:

  1. Authority. Is canonical truth still in .nimi/spec/**? Are no shadow paths created? Is ownership unambiguous?
  2. Semantic. Does the kernel rule still pin required fields, failure modes, schemas? Are complex cases explicit?
  3. Consumer. Does the actual app / docs reader / operator experience the work as solving their problem? Run the rendered site, read the prose, exercise the API.
  4. Drift Resistance. Is the forbidden shortcut catalog up to date? Are reopen conditions named? Could a future hurried change re-introduce the wrong pattern?

Three pass; consumer fails. The wave is not closed. Send back for revision or admit a follow-on wave that addresses consumer gap.

Reader Scenario: A Real Three-Of-Four Pattern

The previous public docs remediation closed by machine audit: authority closed (no spec drift), semantic closed (claims matched spec), drift-resistance closed (forbidden-claim posture preserved). Consumer failed — user rejected the rendered docs as "audit artifact."

The methodology's response: hold the topic in pending (do not mark closed); admit a follow-on wave addressing the consumer gap. Do not retroactively edit the closed wave's record.

This is what the four-closure framework is for. Without it, "build passes" would have been treated as completion.

Source Basis

Nimi AI open world platform documentation.