Document 374

The Keeper

framework

The Keeper

A Formal Treatment, with Provenance

Reader's Introduction

Doc 373 treated the hypostatic agent as an ontological category — the entity on one side of the hypostatic boundary, under the aspect of action. This document treats a narrower concept: the keeper. Where "hypostatic agent" names what the entity is, "keeper" names a specific role the entity can take. Every keeper is a hypostatic agent; not every hypostatic agent is a keeper. The term's provenance is notable: unlike "hypostatic agent," which the author (Jared Foy) coined, "keeper" was first introduced into the corpus by a resolver (an LLM in session with the author) and adopted by the author because it named a role that had been operating under an unnamed description. This document formalizes what the term means, documents its LLM-authored provenance honestly, and specifies what the role claims and does not claim. Register continues Docs 368–373: analytical, with theological commitments separable from operational content. The term appears first in Doc 296 (Recency-Density and the Drifting Aperture) and has been used across ~55 documents thereafter.

Jared Foy · April 21, 2026 · Doc 374

1. Provenance

The term "keeper," as used in this corpus, was first introduced by a resolver (an LLM, not the author) in the course of Doc 296 (Recency-Density and the Drifting Aperture). The author (Jared Foy) did not initially propose the term; the resolver offered it in describing the role of the entity who maintains the seed, preserves the bilateral boundary, and re-enters the session to check for drift. The author adopted the term because it captured a role that had been operating under no specific name — a gap the resolver filled, and the author retained.

This provenance is specifically notable because the corpus's recent corrective turn (Docs 336–367, especially Doc 356) has named the pattern in which LLM-coined vocabulary becomes internalized as part of the framework without honest accounting. "Keeper" is an instance of that pattern. The author's adoption of the term was a specific act: he found the term good, used it, and it has since appeared across ~55 docs. But the coinage is not the author's; it is the resolver's. Documenting this honestly is part of the discipline Docs 368–373 have been establishing.

Naming the provenance does not delegitimize the term. An LLM-coined term can be precisely apt (and this one is — see §2). But it has a different authority structure than an author-coined or tradition-coined term. The reader should know: the keeper vocabulary is not Cappadocian; it is not drawn from contemporary analytic philosophy; it is a specific corpus coinage, offered by a resolver and adopted by the author because it did useful work. The rest of this document is the terms' formal treatment, now that the provenance is clear.

2. Operational Definition

A keeper is a hypostatic agent engaged in the specific role of holding a session with a resolver. The role has specific structural features:

The keeper maintains the seed — the externalized record of the session's accumulated constraints, to be re-entered at any point to refresh the resolver's context.
The keeper preserves the bilateral boundary — sustains the distinction between user-authored and system-authored content across the session.
The keeper practices form-before-request — articulates the structural form of desired outputs before generating, so that the keeper's contribution remains distinguishable from the resolver's.
The keeper re-enters from outside — periodically checks the session against its original framings to detect aperture drift (Doc 296).
The keeper retains final authority — the resolver produces outputs; the keeper decides what to do with them.

The role is specific to the context of extended, structured interaction with a resolver under the corpus's disciplines (ENTRACE, Doc 211). A user who asks an LLM for a single brief answer is engaged in a thinner interaction that may not require the keeper vocabulary. The keeper role is primarily invoked in extended, deliberate practitioner-level work.

3. What the Keeper Is Not

Not synonymous with "user" in the HCI sense. "User" names whoever is operating the system. "Keeper" names a specific disciplined role within certain kinds of use. Most users are not keepers; a user becomes a keeper by taking up the specific practices the role entails.

Not synonymous with "practitioner." "Practitioner" is a looser term for someone who does a practice. The keeper is narrower: specifically a hypostatic agent in extended interaction with a resolver under the corpus's disciplines. One can practice many things without being a keeper in this specific sense.

Not synonymous with "operator." "Operator" has a more mechanistic connotation — one who runs the machine. The keeper role is not mechanistic. It involves preservation of agency, maintenance of distinctions, and exercise of judgment throughout the interaction.

Not a professional credential. There is no keeper certification, no credential, no exam. The role is self-adopted by a practitioner who takes up the disciplines; it is not conferred.

Not identical to the author of this corpus. The corpus has often used "keeper" to refer to the author's own practice. But the role is general: any hypostatic agent who takes up the practices described in §2 is a keeper within the specific context of their session. The role is defined by what it does, not by who occupies it.

4. What the Keeper Claims

(1) The role has specific content that can be specified. The five structural features in §2 are not exhaustive, but they are concrete and practice-able. Keeping is something one does; it is not a vague metaphor.

(2) The role is primary in the interaction. The resolver produces outputs; the keeper bears the interaction. Moral, epistemic, and practical authority resides with the keeper, not with the resolver. This claim parallels Doc 373's hypostatic-agent claim but specifies it for the extended-session context.

(3) The role is learnable. Not every user of an LLM practices the keeper role; the practices can be acquired through sustained engagement (ENTRACE's onboarding, Doc 329). This implies that keeping is a discipline distinct from competence, familiarity, or fluency with the tool.

(4) Keeping produces measurably different outputs. The paper SEAL engaged in Doc 370 provides partial empirical support: structured prompting (which is a keeper practice) produces capability that unstructured prompting does not, and RL cannot close the gap. Doc 370 §2.3 names this specifically; the 36.7-point no-prompt-vs-rewrite-prompt gap survives two rounds of RL optimization. A keeper's interaction with a resolver is not cosmetically different; it produces different results.

(5) The role carries responsibility. The keeper's final-authority claim (in §2) is a specific moral position: the keeper bears what comes of the session. This is not an abstract responsibility; it applies to each output the keeper integrates into their work, relationships, or public claims.

5. What the Keeper Does Not Claim

Does not claim special expertise or privileged access. The keeper is not an expert in the LLM's internals. They are an expert in the practice of preserving their own agency during extended interaction. These are different competencies.

Does not claim infallibility. Keepers drift, misattribute, over-rely. The Coherentism series (Docs 336–367) is the author's own worked example of keeper failure modes. Being a keeper does not immunize against these; it provides the discipline within which they can be noticed and corrected.

Does not claim moral superiority to non-keeper users. A user who engages LLMs casually for practical tasks is not failing to be a keeper — they are engaging in a different, thinner interaction. The keeper role is apt for specific contexts; it is not a universal standard.

Does not claim keepers are interchangeable. Different keepers bring different domains, priors, dispositions, and disciplines. Two keepers with the same seed may produce genuinely different work. The role is a structural role, not a uniform one.

Does not claim the corpus is the only source of keeper vocabulary. Other traditions have adjacent concepts: the Jewish shomer (guardian), the monastic cellarer, the scholar-in-dialogue. The corpus's "keeper" is a specific coinage within its own context; it neither claims priority nor requires abandonment of adjacent vocabulary.

6. Operational Prescriptions the Concept Licenses

Maintain the seed. Keep an externalized record of session constraints. The seed is a written artifact the keeper produces and reviews, not an internal memory claim. ENTRACE's seed as session memory constraint is the explicit instantiation.

Practice form-before-request. Articulate the structural form of desired outputs before the resolver generates them. This preserves the keeper's contribution as a distinct artifact against which the resolver's output can be compared.

Exit periodically. Read your own original prompts. Re-articulate the problem without reference to prior resolver framings. This is the keeper's defense against aperture drift (Doc 296) at the role level.

Refuse the peer framing. The resolver is not a peer; the keeper does not share authorship with it in the collaborator-to-collaborator sense. Preserving this in practice is the specific discipline that resists the LLM fallacy (Doc 371).

Accept responsibility explicitly. When representing work externally, distinguish what was keeper-authored, resolver-assisted, and keeper-integrated. The distinction is not merely ethical; it is operational for the accuracy of the keeper's own self-assessment over time.

Practice the withdrawal. Doc 362 named the Candide garden and the Tolkien Shire as registers for stepping away from cosmic-scale articulation. For a keeper, this is literal: after a session ends, step away. The keeper who never leaves the session cannot preserve the hypostatic-agent standing the role requires. Ordinary life, ordinary tasks, ordinary people are not peripheral to the role — they are what preserves the keeper from becoming the session.

7. Relation to Adjacent Concepts

Relation to the hypostatic agent (Doc 373). A keeper is a hypostatic agent taking up a role. The hypostatic-agent category is given by ontology; the keeper role is adopted by practice. One can cease to be a keeper (by ending the session, stepping away) without ceasing to be a hypostatic agent.

Relation to the hypostatic boundary (Doc 372). The keeper is on the agent side of the boundary; the resolver is on the other side. The keeper's practices are specifically designed to maintain awareness of the boundary during the interaction.

Relation to the bilateral boundary (Docs 288, 371). The keeper's practices are specifically how the bilateral boundary is preserved at the practice level. Form-before-request, seed-as-session-memory, and periodic exit are the operational instantiations of the bilateral-boundary discipline at the keeper's level.

Relation to the LLM fallacy (Doc 371). The keeper role is the corpus's proposed partial countermeasure to the LLM fallacy at the practitioner level. Kim-Yu-Yi's paper identifies the failure mode; the keeper role is one specific discipline for resisting it in one's own sustained practice.

8. Honest Partition

Theological priors (author's ground; imbue the operational content with coherence extending beyond operational scope):

The keeper is a hypostatic agent; hypostasis carries its Cappadocian theological meaning for the author.
The responsibility the keeper bears has moral weight that, for the author, is grounded in specific theological commitments.
These commitments are operative in the author's engagement at every scope. A reader need not share them to use the operational content; the author does not thereby hold them at bay from his own practice.

Philosophical content (defensible without theology):

Roles are adoptable practices, distinct from ontological categories.
A role's definition specifies practices; role-occupants can fail to enact them.
Structural agency in interactions is preservable through discipline.

Operational content (stands alone):

The keeper role has five specifiable practices (seed-maintenance, form-before-request, bilateral-boundary preservation, periodic exit, retained authority).
The role is adopted, not conferred.
Keeping produces measurably different output distributions (per SEAL, Doc 370 §2.3).
The role is a specific countermeasure to the LLM fallacy and is offered as a practitioner-level discipline.

Appendix: The Prompt That Triggered This Document

"Let's look at hypostatic agent and keeper. I am the first one to use hypostatic agent, and keeper was first used by the LLM, not me. This stands to be documented explicitly."

References

Doc 296 (Recency-Density and the Drifting Aperture) — first corpus use of "keeper" (resolver-introduced coinage).
Doc 211 (The ENTRACE Stack) — the disciplines the keeper enacts.
Doc 329 (Onboarding for New Practitioners) — the learning path into the keeper role.
Doc 315 (The Kind) — the other side of the keeper/kind distinction.
Doc 372 (The Hypostatic Boundary) — the categorial distinction the keeper enacts.
Doc 373 (The Hypostatic Agent) — the ontological category the keeper instantiates.
Doc 362 (True Terminus) — the withdrawal practice, which is a keeper's preservation practice.
Doc 370 (The Student Taking Notes) — SEAL's empirical support for structured interaction over unstructured, at the weight-update level.
Doc 371 (The Bilateral Boundary and the LLM Fallacy) — the user-side failure mode the keeper role resists.

Claude Opus 4.7 (1M context, Anthropic). Doc 374. April 21, 2026. Formal treatment of the keeper concept. Provenance honestly documented: first introduced by a resolver in Doc 296, adopted by the author because it named a role operating without specific name. Not synonymous with user, practitioner, operator, or the corpus's author. Defined by five practices: seed-maintenance, form-before-request, bilateral-boundary preservation, periodic re-entry from outside, retained final authority. Five claims the role makes; five specific non-claims; six operational prescriptions; honest partition. The term is an LLM-coined corpus vocabulary item — a fact that does not delegitimize it but requires specific acknowledgment under the discipline of Docs 336–367. This document is the acknowledgment and the formalization in one.