Ingest: Where It Lives Is Not What It Is (architectural vocabulary for retained adaptation)

Type: kb/sources/types/ingest-report.md

Source: where-it-lives-architectural-vocabulary-retained-adaptation.md Captured: 2026-05-28 From: file:///home/zby/txt/paper/submissions/asisas-2026/paper.md

Classification

Type: scientific-paper -- a position paper (ASIS&S 2026 submission) with a citation apparatus (Perry & Wolf, Greshake et al., Park et al., Shinn et al., plus the recent agent-memory/skills survey line). It argues a thesis rather than reporting an experiment, but the methodology lens (what is and is not validated) applies, and the paper itself flags the absence of empirical validation. Domains: artifact-analysis, context-engineering, software-architecture Author: This is the KB owner's own paper. It is the published, externally-framed statement of the exact vocabulary this KB has already codified across its definition notes and axes-of-artifact-analysis.md. Credibility for KB purposes is high because it is our position; the caveat is that it is self-authored, so it cannot serve as independent corroboration.

Summary

The paper argues that retained behavior-shaping artifacts in agentic systems (prompts, workflows, validators, routing rules, memories, adapters, checkpoints) should not be classified by storage substrate (memory system, flat file, repo, vector store, etc.) but by the properties that determine how they affect behavior. It proposes a four-field architectural record: storage substrate (where state lives), representational form (prose / symbolic / distributed-parametric, each with its own default inspection method — read / test / probe), lineage (review-relevant source dependencies for invalidation and rollback), and behavioral authority (consumer x channel x force). Classification attaches to the operative part or consumption path, not the stored object, because objects bundle several behavior-shaping parts. A worked dependency-resolution example shows the same lesson retained five ways, each with a distinct review-evidence / invalidation / rollback profile. The framework is offered as a design and review discipline that exposes efficiency, security, and sovereignty risk; the paper claims no empirical validation and proposes design-review and incident-analysis trials as future work.

Connections Found

The companion connect report found zero existing inbound edges and (correctly, since snapshots author no outbound links) no outbound connections. Its grep showed the KB reproduces this paper's vocabulary across at least nine notes — every field/term definition and the central synthesis note paraphrasing a paper section almost verbatim.

Correction to the connect report's edge proposal. The connect report read this near-total restatement as a missing-attribution gap and proposed reverse derived-from edges from axes-of-artifact-analysis.md and the field/term definitions (storage-substrate, representational-form, lineage, behavioral-authority, retained-artifact, operative-part) to this snapshot. That direction is wrong. The paper was distilled from these notes; the notes are the source, the paper the derivative. The lineage edge therefore belongs on the paper (paper derived-from notes), not on the notes pointing at the paper — and since snapshots author no outbound links, that lineage is recorded in this report and the snapshot metadata, not as note edges. There is no notes→paper derived-from work to do, and no evidence edges either: a self-authored paper cannot independently corroborate the notes it came from (see Limitations). What is real is the reverse: once the paper is accepted, the notes may carry a citable-reference edge to it as the published external statement of the position — a behavioral-authority relationship, deferred until acceptance. The connect report also flags this as a high-fan-out source and notes the paper's sovereignty framing has no representation anywhere in kb/notes/ — that, not attribution, is the live follow-up.

Extractable Value

  1. Lineage runs paper-from-notes, recorded on the paper, not the notes. The near-total restatement across nine notes is not a missing-attribution gap to fix with note edges — the notes are the source and the paper the distillation, so there is no derived-from edge to author from the notes. The paper-derived-from-notes lineage is recorded here and in the snapshot metadata. The genuine future edge is the reverse and conditional: once accepted, the notes may reference the paper as the published external statement of the position (a behavioral-authority, not lineage, relationship). [deferred — gated on acceptance]
  2. Sovereignty as a third risk axis, alongside efficiency and security. The paper frames sovereignty risk (behavior depending on artifacts the owner cannot inspect, regenerate, delete, or roll back — vendor checkpoints, hosted retrieval indexes; Couture & Toupin 2019, Dale 2025) as first-class. The connect report confirms grep finds no sovereignty coverage in any note. The efficiency and security axes already appear scattered across the artifact-analysis cluster; this is the missing third leg and a candidate synthesis note: "the four-field record exposes an efficiency / security / sovereignty risk triad." [experiment]
  3. External-facing framing and terminology for the same vocabulary. The paper packages our internal terms for an architecture audience and supplies an academic citation scaffold (Perry & Wolf for "architectural element"; Muccini & Vaidhyanathan, Franch et al. for AI-architecture lineage; Greshake et al. for the prose data/instruction-blur security claim; the Jia/Du/Jiang/Mei survey line for the storage-first classifications it argues against). Useful as a reference for how to present the vocabulary outside the KB and for sourcing claims our notes currently assert unsourced. [just-a-reference]
  4. The dependency-resolution worked example as a ready-made teaching/validation case. The Section 7 table (same lesson retained five ways, each with its own review-evidence / invalidation / rollback row) is a concrete instantiation of the abstract record. It is reusable as a worked example in axes-of-artifact-analysis.md or a future instructions artifact, and as the exact kind of design-review exercise the paper's future-work section proposes. [just-a-reference]
  5. Explicit statement that the framework is unvalidated and the proposed validation protocol. The paper names what would test it: apply the record in design reviews and incident analyses, check inter-rater classification consistency of operative parts, and check whether derived obligations expose risks storage labels miss. This is a candidate research/operational agenda the KB could adopt to move the vocabulary from theory to tested practice. [deep-dive]

Limitations (our opinion)

Editorial judgment, labeled as opinion. Two limitations matter for how the KB should treat this source.

First, it cannot serve as independent corroboration of our own claims. It is self-authored: it is the same position our notes already hold, so an evidence edge "the paper corroborates the note" would be circular if read as external support. The honest relationship is derived-from (the note abstracts the published statement) or co-authorship of one position, not independent confirmation. Treat citations to it as "we also argued this in print," not "an outside party validated this."

Second, by the paper's own admission there is no empirical validation — no design-review trial, no incident analysis, no inter-rater study. The central claims (that storage-first classification is "architecturally insufficient," that reviewers can classify operative parts consistently, that the four fields are the right cut) are argued by construction and worked example, not tested. The paper's hardest-to-vary move is the four-field set itself: Section 8 concedes the fields "are not exhaustive" and that lineage and behavioral authority were "compressed for space," which means the boundary between fields and the claim that these four are the load-bearing ones is asserted rather than demonstrated. A simpler account the paper does not rule out is that representational form alone (with substrate, lineage, authority as attributes of it) carries most of the discriminating work. For KB purposes this is fine — we are building methodology, not running a study — but the recommended-action and any promoted synthesis should not overstate empirical grounding.

The live follow-up is the sovereignty refinement (extractable value item 2), not lineage edges. The paper, while being distilled from the notes, refined and added content the notes do not yet hold — most clearly the sovereignty risk axis (behavior depending on artifacts the owner cannot inspect, regenerate, delete, or roll back), alongside the efficiency and security axes the artifact-analysis cluster already scatters. Mining that back into the notes — a candidate "efficiency / security / sovereignty risk triad" synthesis note — is the highest-value forward action and the correct direction of flow (paper-refinement → notes).

Two things are explicitly not to be done: (a) do not author derived-from/evidence edges from the notes to this snapshot — the lineage runs the other way (item 1); (b) do not yet add a reference edge from the notes to the paper — defer that citable-authority edge until the paper is accepted, then record it as a behavioral-authority reference ("we also argued this in print"), not as independent corroboration. Promotion is out of scope for this ingest.