054-Add adapted-from and operationalized-from lineage relations

Type: ../types/adr.md · Status: accepted

Status: accepted Date: 2026-07-18

Context

ADR 053 retired distillation with deliberately no successor technical term, on the reasoning that the KB's link grammar (derived-from/abstracted-from) already carried the maintenance-relevant boundary and a bare successor word would fail the minimality bar. Its own Consequences section named the resulting gap: "the KB's most common composite operation — consumer-directed selection plus reshaping — no longer has a one-word name."

The adaptation-lineage-vocabulary workshop's migration audit found that gap was real, not hypothetical. Writers filled it with unauthorized ad-hoc synonyms — "condense"/"condensing" reintroduced as an unacknowledged peer operator in four notes, and a separately coined "Extraction" successor term — inside the same migration that retired distillation for exactly this failure mode ("the control trap can recur under any future attractive label," per ADR 053's own risk note); neither term made it into ADR 053, the migration plan, or AGENTS.md's vocabulary list. Separately, the audit found the migration's own flagship derived-from case (skills-derive-from-methodology.md, cited from ~15 other notes) argues itself into the claim-preserving bucket while its own text admits "a different person reading the same methodology would produce a meaningfully different skill" — evidence of judgment-shaped reworking, not strict derivation.

Decision

Introduce two lineage relations, alongside derived-from/abstracted-from, to name this composite operation without reopening ADR 053's other three decisions:

  • adapted-from / Adapted into: — the general relation for consumer-directed selection-plus-reshaping that is neither claim-preserving (derived-from) nor claim-amplifying (abstracted-from). Truth condition, test, and maintenance consequence: link-vocabulary.md.
  • operationalized-from / Operationalized into: — a narrower, flat authoring-time alternative to adapted-from (never a stacked child edge) for the methodology → procedure pairing specifically, where the target adds ordering, defaults, or stopping conditions the source doesn't fix, without new substantive claims.

derived-from is clarified, not changed in substance: its authorization is tested per source→destination pairing against its entailment condition, declared in the source's COLLECTION.md — not banned by register, and not assumed available either.

Recording direction: the persisted edge is always the source-side footer. Authorization sits with the source collection's COLLECTION.md, per ADR 019's existing outbound-link architecture — link-vocabulary.md is a palette, not itself authoritative.

generated-from (canonical artifact → mechanically-reproducible projection) was considered and deliberately not minted: no live case of a persisted artifact needs it yet.

ADR 053's other three decisions are unchanged: "derive" stays ordinary English with no vocabulary entry; ampliative traffic still routes through the discovery lifecycle; distillation still survives only in external-system and historical-ADR senses.

Implemented: catalogue rows and full lineage-semantics section in link-vocabulary.md; operationalized-from authorized in kb/notes/COLLECTION.md for its kb/instructions/ destination (the one pairing with live corpus evidence — adapted-from stays unauthorized everywhere until a real case identifies its actual source collection); the four-way lineage-tracking test in cp-skill-write/SKILL.md; the hyphenated-identifier-vs-prose convention in AGENTS.md (this repo's terms) and AGENTS.md.template (generic, no hardcoded terms).

Consequences

  • Easier: the composite operation ADR 053 left unnamed now has two relations with distinct maintenance semantics; corpus cases like skills-derive-from-methodology.md can be correctly classified as judgment-shaped rather than forced into derived-from or left to accumulate ad hoc synonyms.
  • Harder: four lineage tests to learn instead of two. Mitigated by keeping cp-skill-write's own statement terse (names the four headings, defers the actual tests to link-vocabulary.md) rather than inlining every test at the write path.
  • Corpus corrections (execution-plan §4, complete): skills-derive-from-methodology.md and write-instruction.md reclassified — the audit showed the latter has two distinct provenance stories (repeated practice → abstracted-from, methodology → operationalized-from), not one mislabeled edge; kb/work/lineage-mechanisms/'s six "ad-hoc distillation" passages resolved per case, not by blanket substitution; a corpus-wide Derived into: sweep for methodology→procedure targets; the Finding 1 ("condense"/"condensing") and Finding 2 (trace-rework language, probe-checklist wording) reconciliations; constraining.md and context-engineering.md updated. The reported 70+ generated "Extraction" footers could not be reproduced — no emitter exists in scripts/, src/, or kb/instructions/, and the cited commit's diff is unrelated rename/backlink maintenance — so no generator fix was made; the claim is retained in migration-audit-findings.md as an unreproduced historical report, not acted on further.
  • Risk: the same recurrence risk ADR 053 named — and it recurred within the very migration that fixed the last instance. Considered and rejected a write-time gate in cp-skill-write against this: the actual failure was ordinary-English peer-operator words ("condense," "Extraction"), which no fixed check can catch prospectively since the next escape word isn't known in advance, and the hyphenated-identifier discipline this ADR extends was never what failed. Enforcement stays with periodic audit — the same mechanism that caught this instance.