053-Retire distillation without a successor term

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

Status: accepted Date: 2026-07-17

Context

Distillation was a load-bearing vocabulary term — in the AGENTS.md vocabulary, a definition note, a tag in learning-theory's covered_by, note titles, and the Distilled into: link label. A corpus audit classified 464 occurrences and found the word covering operations with opposite maintenance semantics: entailment-preserving reshaping (methodology → skill; the output is re-derivable from the retained source) and evidence-to-rule generalization (traces → preference rule; the output must earn authority through testing). The definition note's own instance list spanned the boundary, and the KB's most theoretically load-bearing "distill" usage — the trace→rule ladder — was on the ampliative side. Because the word did not mark which control regime governed an artifact, use-shaped artifacts could borrow the recomputable-copy maintenance story without holding its checkability.

Two theory results shaped the remedy. Vocabulary collisions are prevented at write time, not resolved at read time: prose gives an unqualified technical sense no reliable namespace, co-loaded notes merge colliding senses silently, and a definition note binds too weakly to prevent drift — the distillation definition existed throughout the drift it failed to stop. And a minimality constraint: a technical term that is also a common English word taxes every future occurrence with a which-sense resolution, so a word earns a definition only when its operational consequences cannot be carried by plain prose with a citation. Applied strictly, a successor term derivation fails that bar — ordinary "derive" is close enough, and the strong sense can cite its theory directly.

Decision

Retire distillation as KB vocabulary with no successor technical term.

• "Derive" stays ordinary English. No definition note, no vocabulary entry. A passage that needs the strong sense — entailment-preserving reshaping with matching, fallback, and cache/staleness semantics — cites the theory's home: the two-layer execution system note, which states those maintenance regimes as citable claims.
• The link grammar carries the boundary. derived-from / Derived into: asserts the artifact adds no substantive claims beyond its source and receives recomputable-copy or managed-staleness maintenance; abstracted-from / Abstracted into: asserts the claims exceed the instances and authority is earned by testing. Semantics are specified in link-vocabulary.md. An entangled artifact is classified by its dominant regime as an explicit, falsifiable bet.
• Ampliative traffic routes to the discovery lifecycle. The coined compound discovery lifecycle is the technical term for staged ampliative acceptance; traffic enters at its conjecture phase and never inherits accepted status from the sentence that states it. Bare "discovery" is ordinary English and no note title captures it.
• Distillation survives only in external systems' own command names, the machine-learning knowledge-distillation sense inside captured sources, and historical ADR text.

Consequences

• Easier: an artifact's maintenance regime is readable off its lineage label instead of inferred from an overloaded word; the incoming technical term is a greppable multi-word coinage, so the write-time uniqueness check is a lexical search rather than the multi-agent semantic classification the 464-hit audit required; the derived/abstracted boundary is policed by labels and review gates, which prose demonstrably could not do.
• Harder: the KB's most common composite operation — consumer-directed selection plus reshaping — no longer has a one-word name and must be described in plain language with a citation where the strong sense matters. The resonance with ML knowledge distillation is lost deliberately: that sense is statistical fitting, and the borrowed intuition sat on the wrong side of the exact boundary the KB sense needed to hold. Amended by ADR 054, which names this operation adapted-from/operationalized-from; the other three bullets are unchanged.
• Risk: the control trap can recur under any future attractive label. The guard is the promoted invariant — load-bearing senses ride on schema positions, rare compounds, or link-required clausal binding, never on capturing a common word — and the label semantics that force every lineage edge to declare its regime.