ADR 089 was recorded as accepted before implementation

Observation and provenance

Recorded on 2026-09-25 while opening this workshop. The operator identified the latest ADR as already present in the ADR directory while implementation was only starting. They also reported that agents often make this mistake: the ADR text is needed before implementation, and the normal workaround is to write that draft in the workshop first.

Repository inspection used commit df4ae037b8405e7bb811faa04b76f829932f4a83. The facts below describe that revision, not a claim that the inconsistency persists in a later checkout. Live links identify the artifacts; git show <commit>:<path> recovers their observed versions.

Repository evidence

  • The ADR type says: "Use an ADR for a concrete architectural decision that has been made and implemented". The proposal type describes a proposal as a finished but unadopted design and excludes a decided choice from that role.
  • ADR 089 was already in kb/reference/adr/, with both frontmatter and visible status set to accepted. Its Decision section places tag heads in kb/tags/, replaces the old hub, and retires index_source and index_key from the tag-head type.
  • The tag workshop, kb/work/tag-contract-convergence/README.md (deleted when the workshop closed in b745bcc8; recover it with git show), simultaneously said that ADR 089 adopted the choices, then stated: "No live participation declaration, consumer change, or relocation has landed."
  • The observed git tree had no kb/tags/ directory. It retained kb/notes/tags-README.md, and kb/types/tag-readme.schema.yaml still used index_source and index_key. These are concrete implementation witnesses, not an exhaustive audit of all seven decision clauses.
  • The ADR was introduced in commit c9926a11e8de8b2c474bdfc4678860247729ece8, dated 2026-09-25T22:41:51+02:00, titled "Adopt one tag namespace per KB with heads in kb/tags (ADR 089)". Its most recent edit at observation was commit 81efa680fe96c2b8f0586bbc038b97844edae7d8, at 2026-09-25T23:00:22+02:00. The inconsistency was therefore present in committed records, not only an unsaved draft.

Why this was an easy mistake

The operator's account distinguishes the necessary activity from the error. The intended decision must be written before implementation to guide that implementation. Under the current convention it is drafted in the workshop, then promoted to the ADR directory once implemented. The mistake is putting that useful pre-implementation text directly in its eventual destination.

The working diagnosis is that the document form and destination encourage conflation: an agent is asked to write an ADR, and kb/reference/adr/ is the obvious directory. But in Commonplace that placement asserts more than "the choice has been made". It also asserts that implementation exists. The proposal and ADR contracts do not provide an explicit durable category for the intervening state, so the workshop carries the distinction through practice. This diagnosis explains a plausible error path; no original agent transcript was inspected to establish its actual reasoning.

The ADR schema checks document shape and admits accepted, superseded, or deprecated; those checks alone cannot establish that a decision has been implemented. Adding a status field without clarifying its meaning and evidence would leave the central problem open.

Consequence and evidence limits

A reader treating the ADR as a description of shipped architecture can infer a layout and rules that the observed checkout does not yet implement. The workshop and ADR communicate different lifecycle states unless the reader already knows the drafting convention.

This is one repository-supported occurrence plus an operator report of recurrence. It does not establish an error rate, the original agent's motive, or downstream harm. It also does not show that the decision itself was wrong. Implementation belongs to the tag workshop. This record preserves the earlier inconsistency even if that work later makes the ADR accurate.