Migration plan: replace distillation with derivation

Goal (operator-set; revised 2026-07-17): retire "distillation" without a successor technical term. "Derive" stays ordinary English everywhere — no definition note, no vocabulary entry. The theory that backed the old term moves into citable notes and the link grammar. Prerequisite unchanged: clean up the KB so the current usages are cleanly separated — some split into their components, some replaced by plain language or routed to discovery. This file sequences that work against the usage audit's evidence.

End state

  • No successor term. The theory of entailment-preserving reshaping (graded entailment, matching, fallback, cache/staleness semantics) lives in the two-layer structure note as canonical citable claims. A note that needs the strong sense cites the structure note or states the specific support inline — the same way any claim cites its grounds.
  • The link grammar carries the boundary: Derived from: / Derived into: semantics are specified in link-vocabulary.md — the label asserts the artifact adds no substantive claims beyond the source; ampliative lineage takes a discovery-side label. Labels and review gates police the DER/AMP boundary; prose never did (the audit shows distillation drifted with a definition in place).
  • The discovery lifecycle owns all ampliative traffic, graded by the polation axis (now in discovery-lifecycle.md's grading section; thread closed): routine induction (extrapolation) at the shallow end, generative insight (hyperpolation) at the deep end. The compound is the technical term (decided 2026-07-17, per the naming rule in prose-has-no-scope); bare "discovery" stays ordinary English.
  • Mixed artifacts carry an explicit dominance bet (the dominant-regime call, now specified in link-vocabulary.md; vocabulary thread closed) instead of a hybrid category.
  • "Distillation" survives only in: external systems' own command names, the ML-knowledge-distillation sense in sources, and historical ADR text.
  • Net vocabulary diff: 0, with a quality upgrade. distillation leaves the AGENTS.md vocabulary; discovery lifecycle enters (there is no existing discovery entry to amend). The incoming term is a greppable multi-word coinage, not a captured common word — the exposure class that produced the migration.

Execution status (2026-07-17)

  • First semantic wave applied and committed (80f47b41): the trace→rule ladder cluster renamed to abstraction vocabulary (the six notes plus agent-memory-README.md), with the note renames trace-derived-memory-earns-authority… → trace-extracted-… and use-trace-derived-extraction → use-trace-extraction-as-meta-learning and corpus-wide backlink fixes. Validation clean. Residues — and the lesson that sweep patterns must match single-l distil — recorded in obvious-distillation-cases.md.
  • Adjacent infrastructure committed (ca1ea328, 5ed02f6a): trace-derived retired in favor of the coined compounds trace-extracted (lineage) and trace-learning (loop tag, type-spec sections) across agent-memory-systems, systems.csv, the renderer, and tests — the naming rule applied to infrastructure vocabulary.
  • Workshop self-references cleaned (8dd16304): the threads no longer use "distillate"/"distillation" as operator vocabulary while arguing for its retirement.
  • Theory promoted (e429e82e): the scope/collision reflection is now the library note vocabulary-collisions-prevented-at-write-time-not-read-time.md — the uniqueness invariant and the multi-word-coinage rule are citable claims the remaining waves build on.
  • Ordering note: these Wave 1-shaped edits deliberately ran ahead of Wave 0, under the staging rule in obvious-distillation-cases.md — only cases whose replacement follows from the note's own argument, needing no receiving vocabulary.
  • Wave 0 item 1 done (7e9a63c8): the structure note is promoted as kb/notes/theory-and-methodology-form-a-two-layer-execution-system.md — one note, one claim; open-endedness argument folded in as grounds; promotion criterion kept in the note (no separate gate artifact); maintenance semantics stated as citable claims; physics terms cited as loans. Stress test passed: all ten hard-case rows in the receiving-vocabulary draft (thread since closed as consumed) classify using the note plus plain language, confirming the later waves are expressible as prose + label + citation. The two-layer thread in this workshop is now source material with a durable home; a review-gate pass over the promoted note is still worth scheduling before Wave 1 leans on it.
  • Wave 0 item 2 done (0dc7fb4c): link-vocabulary.md now defines the two lineage regimes — derived-from/Derived into: (no added claims; recomputable-copy maintenance, citing the structure note) vs abstracted-from/Abstracted into: (claims exceed the instances; authority earned by testing) — with the classifying question, the dominant-regime call for mixed artifacts, and Distilled into: retired for new writing. The edit also fixed a latent collision in the catalogue itself: the old derived-from row defined the label as "abstraction provenance". 15 existing Distilled into: footers plus the pre-existing derived-from edges are flagged as unclassified lineage pending the Wave 1/2 passes. cp-skill-write's footer prescription updated to match.
  • Wave 0 item 3 done (7ac83249) — Wave 0 complete. kb/notes/definitions/discovery-lifecycle.md defines the compound (phases with Peircean analogues, polation grading with the no-geometry caveat, ampliative traffic entering at the conjecture stage, bare "discovery" ordinary English); the co-arising insight note links in as the degenerate case; the discovery tag-README (complete: true) lists it. The draft's six pre-Wave-1 acceptance tests all pass: every non-META audit class has a receiving home (derived/abstracted labels, conjecture stage, accumulation/constraining for authored judgment), the trace→rule ladder already rewrote without a bridge term, conjecture and accepted status are distinguishable, consumer-directedness and the coverage bet live in the structure note's region choice, form-vs-content distance is stated, and every rewrite names its control regime via label or citation.
  • ADR recorded (2026-07-18): ADR 053 records the whole decision — retire without successor, labels carry the boundary, discovery lifecycle owns ampliative traffic — with the remaining Wave 2/3 churn named in its Consequences.
  • Waves 2 and 3 executed (2026-07-18): ~60 files' prose reworded to plain language by three parallel batch agents; skills-derive-from-methodology-through-distillation rewritten around two-layer derivation and renamed to its first clause; the definitions cluster de-termed (constraining's use-shaping 2×2 kept as the salvaged orthogonality table); definitions/distillation.md deleted with all inbound links repointed or dropped and the Faithful-Self-Evolvers caveat salvaged into knowledge-storage-does-not-imply-contextual-activation.md; the distillation tag retired (covered_by now six children, seven member notes redistributed to constraining/deploy-time-learning/agent-memory/discovery, distillation-README.md deleted); the last two distill* filenames renamed; AGENTS.md vocabulary swapped Distillation for Discovery lifecycle; kb/index.md, the tag READMEs, ADR 011's gloss, and the undefined-terms/unearned-generality gate examples migrated. Full-corpus validation: zero failures. Residual distill* occurrences are all sanctioned survivors (ADR history, the collisions note's meta-discussion, link-vocabulary's retired-label documentation, quoted history, the ML sense). Still deferred per plan: the kb/agent-memory-systems type-spec surface, kb/sources/, and the crystallized-reasoning coinage decision.
  • Wave 1 planned and started — full inventory with per-file treatments, gate flags, and lineage-edge classifications in wave-1-worksheet.md. Already executed: distillation-is-transformation-not-selection retired (757d6daa, operator decision — its polemic target dissolved with the term); the COLLECTION.md derived-from glosses fixed and abstracted-from authorized (9e252b41 — both files instructed writers to mis-label). Remaining: five gated rewrites (tradeoff-note retitle, the false "every" in a-derived-copy…, source-tracking convention update, evolving-understanding retitle, memory-management contrast), write-instruction.md's two provenance stories, ~24 derived-from edges mostly flipping to abstracted-from, and 9 Distilled into: footers to classify. Wave 2's obvious library subset (the second-wave table in obvious-distillation-cases.md) runs in parallel — safe batch work for a mechanical executor; three held-back notes reclassified into it (see worksheet). Then the rest of Wave 2 and the Wave 3 META surfaces.

Waves

Ordered so each wave makes the next one mechanical. Validate (cp-skill-validate) after each wave; commit waves separately.

Wave 0 — receiving vocabulary (blocks everything)

No usage can be re-worded until the target terms exist and are constrained.

Minimality constraint (operator, 2026-07-17): technical terms that are also ordinary English words are inherently confusing — marker-scoping localizes the ambiguity but does not remove it, and every dual-sense word taxes every future reader with a which-sense resolution. The mitigation is a minimal technical vocabulary (this is minimum-viable-vocabulary-is-the-naming-set-that-most-reduces applied to the framework's own reflective terms): a word earns a definition note only when its operational consequences — verification and maintenance semantics — differ enough from plain prose that inline support can't carry them. Applied strictly (operator, same day), derivation itself fails the bar: the natural meaning of "derive" is close enough, and a note that needs the strong sense can cite the theory directly. So Wave 0 promotes theory and label grammar, not definitions. The physics loans (matching, cutoff, correspondence) stay cited-in-passing, not coined; the authored-judgment residue reuses accumulation/constraining; polation stays a source citation grading discovery internally. The constraint also ratifies the routing decision — the ampliative operation gets no separate captured word; its technical handle is the compound discovery lifecycle (naming decision of 2026-07-17), with bare "discovery" staying ordinary English.

  1. Promote the two-layer structure note with the theory stated as citable claims: graded entailment-preservation (checkability degrades with source coherence, per a-derived-copy-of-recomputable-truth-must-be-checked-or-absent.md), matching, fallback, cache/staleness semantics. This note is the canonical home a rewritten usage points at when it needs the strong sense; the selection material (consumer-directedness, the coverage bet) folds in here too.
  2. Specify the link-label semantics in kb/reference/link-vocabulary.md (moved up from Wave 3 — the labels are now the only surface where the entailment claim is formally made): Derived from: / Derived into: asserts the artifact adds no substantive claims beyond the source; ampliative lineage takes a discovery-side label. Update cp-skill-write to match.
  3. Amend discovery-is-seeing-the-particular-as-an-instance-of-the-general.md (and/or a new short definition note if the claim note shouldn't carry definitional load) to own the routine-induction end, using the polation grading. The technical term the amendment introduces is discovery lifecycle; bare "discovery" stays ordinary English. The compound needing a citable home strengthens the separate-note option — a definition note titled by the compound, leaving the existing title-as-claim note's insight claim untouched.
  4. Write the bet doctrine down where it can be cited (a section of the structure note, or a small companion note) — mixed artifacts are classified by dominant regime as a falsifiable commitment with stated detection signatures.

Output check: every non-META instance class in the audit must have a receiving home — plain prose plus structure-note citation (DER/SEL), the discovery lifecycle (AMP), or accumulation/constraining (authored judgment) — with none left implicitly bridged by an undefined strong sense of "derive".

Wave 1 — split the entangled usages (the cleanup the rename needs)

The audit's AMP and MIX instances, ~30 in the library. These are semantic edits — each one re-words a passage to name the actual operation, not a token swap:

  • The definition's own instance list (definitions/distillation.md:23,25): move "repeated failures → gate instruction" and "notice patterns" to discovery examples.
  • The trace→rule ladder cluster: relabel the "Distill" rung (trace-extracted-memory-earns-authority..., an-outcome-check-licenses-replay..., an-accepted-edit-verifies-the-change..., abstract-an-experience...) to induction/discovery vocabulary; their oracle theory transfers unchanged.
  • distillation-is-transformation-not-selection.md: split its argument — anti-mere-selection (kept, becomes an argument for derivation over selection) vs its ampliative evidence (re-filed as discovery). Likely outcome: retitle and rewrite, since both the title's subject and its evidence change vocabulary.
  • write-instruction.md: resolve the internal tension — its "repeated manual operations → stable procedure" framing is discovery (extrapolative), its "distilled from methodology" step is derivation; say both, correctly.
  • spec-mining-as-codification.md:19 and similar: one-line re-words (mined regularities are extrapolations, then codified).
  • Two-stage pipeline descriptions (mostly agent-memory-systems prose, but a few library instances): name the stages separately where separable; where entangled, apply the bet and record it.

Wave 1.5 — dissolved (2026-07-17)

An incoming-term audit was planned here while the direction was still "make derivation a defined term": ~518 loose deriv* occurrences across 166 library files would have become ambiguous against a strict technical sense. With no successor term, prose "derive" cannot be a false signal, and the audit collapses to the labeled edges only — the ~27 existing Derived from:/derived-from link labels, each of which gets classified against the Wave 0 label semantics during the Wave 1/2 passes that already touch every footer. Kept as a record of why the wave existed and why it went away.

Wave 2 — the mechanical rename (DER and SEL instances)

~120 instances that are already derivation- or selection-shaped: token-level re-wording, low risk, batched by collection. Includes skills-derive-from-methodology-through-distillation.md → title simplifies to its own first clause.

Wave 3 — META surfaces (the infrastructure)

  • AGENTS.md vocabulary entry: remove distillation, amend discovery; nothing enters (net −1).
  • definitions/distillation.md: delete with backlink fixes once nothing links to it (no successor to redirect to — per repo norms, no backcompat stub); its salvageable material routes to the structure note and the discovery amendment.
  • Tag retirement: the distillation tag has no successor tag — remove it from learning-theory's covered_by, retire its tag-README, and redistribute member notes to the tags matching their post-migration classification (discovery-side or existing learning-theory children) — one sweep, validator-checked.
  • Distilled into: footer: per the audit these span SEL/DER/AMP, so each footer gets classified during Wave 1/2 passes against the label semantics fixed in Wave 0 — Derived into: where no-added-claims holds, a discovery-side label otherwise.
  • Filename renames (5 notes) with backlink rewrites — use the relocation tooling if the move-map engine workshop has landed anything usable.
  • ADR: record the whole decision as an ADR once Waves 0–1 prove the scheme (worked-case-first). ADR 011's hardcoded gloss and the review-gate examples get updated here.

Deferred / out of scope

  • kb/agent-memory-systems type-spec (Distilled form: field, "raw → distilled loop" contract) — self-contained surface, ~100 boilerplate instances; migrate on its own schedule after the library settles, with the field's replacement name drawn from the discovery side (the field's content is usually ampliative).
  • kb/sources/ and external command names — never touched.
  • The crystallized-reasoning note's coinage — separate decision, out of this workshop's scope.

Best next step (analysis)

The single highest-leverage step is Wave 0, items 1–2 together: promote the structure note and fix the label semantics. Reasons:

  • The structure note is now the canonical home everything downstream cites, and the label grammar is the only surface where the entailment claim is formally made — nothing in Wave 1 can be reworded safely until both exist.
  • The stress-test survives the direction change: the receiving-vocabulary draft's hard-case table still has to classify cleanly — but the classifications now decide which label or citation each case gets, not which defined term. If the trace→rule ladder, ad-hoc distillation, and the skill-creator loop can be expressed as plain prose + label + citation on paper, the waves are safe. (Passed; the draft is closed as consumed.)
  • The discovery amendment promotes together with the structure note — the pair is one atomic change (AMP traffic needs its receiving lifecycle the moment the old term stops covering it).

Risk to watch: Wave 1 rewrites load-bearing theory notes (the trace→rule oracle cluster). Those edits should go through review gates, not just token edits — the semantic-review machinery exists for exactly this.

The receiving-vocabulary draft was fully consumed by Wave 0 — its derivation/selection material became source text for the structure note's theory section and the link-vocabulary spec, its discovery amendment became definitions/discovery-lifecycle.md, and its hard-case table passed against the promoted surfaces — and the thread was closed and deleted 2026-07-17.