Registration, Promotion, and the Self-Evolving Knowledge System

1. “Self-evolving” means governed change

AIOS does not define self-evolution as a model continuously retraining itself on everything a person does. It defines it as a person-governed movement through which work produces evidence, evidence supports candidates, and accepted candidates become reusable ground.

This chapter owns the boundary between experience now and knowledge that may shape later reasoning. Earlier chapters explain how purpose, context, cognitive movements, workflows, and authority produce work. Here, their results acquire temporal and epistemic standing. The Flow Atlas makes that evolution navigable, epistemic maintenance keeps it revisable, and product mechanics ensure that promotion becomes an exact, reconstructable file effect rather than an invisible change of model memory.

The system evolves because its durable semantic environment changes:

• better artifacts are preserved;
• decisions gain rationale and correction paths;
• relationships become explicit;
• workflows and roles can be authored or improved;
• repeated practices can be reviewed;
• obsolete ground can be superseded;
• future contexts can select more useful expertise.

The change remains inspectable, scoped, reversible, and attributable.

2. The practice flywheel

The loop compounds through a governed gate, not by absorbing everything that recurs. Evidence and counterevidence precede the authority decision; acceptance or narrowing returns a resource to future work, while deferral or rejection still preserves disposition and lineage.

The flywheel compounds without converting recurrence into enforcement.

3. Knowledge classes that can evolve

Candidates may become:

• a reasoning scaffold;
• a one-stage method;
• a multi-stage workflow;
• a role or domain perspective;
• a principle or scoped belief;
• a template;
• an exemplar;
• an evidence or decision standard;
• an ontology term or controlled relation;
• a metadata convention;
• a project-local practice;
• an organizational standard;
• a deprecation or supersession rule.

Different classes require different evidence and validation. A useful phrase is not automatically a workflow. A repeated workflow is not automatically a principle.

4. The temporal knowledge boundary

The central transition is not “the model remembered.” It is a sequence of changes in standing:

observed content
  → proposed memory
    → reviewed acceptance decision and authorization to canonicalize
      → exact canonical write and post-state verification
        → settled accepted standing
          → selective activation in a later context

• Observed content is something encountered in a source, interaction, artifact, outcome, or return. It can be valuable, false, adversarial, ambiguous, or merely local to one occasion. Observation gives it no authority over the future.
• Proposed memory is an inert interpretation of what might be worth retaining: a relationship, summary, candidate practice, unresolved contradiction, or resumption cue. It retains its sources, trust domain, time, and derivation.
• An acceptance decision and authorization to canonicalize records that an authorized person or governing process has reviewed the proposal for evidence, scope, counterexamples, and future use and has approved the exact durable change. The decision is always bounded, but accepted standing is not yet settled merely because the decision was expressed.
• The canonical write and post-state verification preserve the approved result, its provenance, standing, rationale, and lineage in the person- or organization-controlled knowledge system and prove that the intended state now exists. The source observation remains recoverable; it is not overwritten by the interpretation derived from it.
• Settled accepted standing begins when the authorized change has been verified and the receiving artifact and companion metadata can be reconstructed consistently after restart.
• Activation is a later context decision. Even accepted knowledge enters only those movements for which it is relevant, current, permitted, and proportionate.

These are standing boundaries, not necessarily separate storage products. A proposal can be held in an ordinary file, but its location does not give it accepted standing. Conversely, an acceptance decision is not settled standing until the corresponding canonical artifact and metadata have been written, verified, and made reconstructable after restart.

Do not read the states as one memory-write event. Observation can be deferred or dismissed; authorization precedes the verified canonical write; accepted ground becomes active only for a fitting movement, and can later be qualified, superseded, retired, or restored.

“Promote” must therefore name the next consequence. Proposing memory, approving development, producing a reusable asset, accepting its claims, registering its identity, writing canonical artifacts, and activating it are not one event.

5. Candidate record

A credible candidate should contain:

candidate:
  kind: workflow | scaffold | role | principle | template | exemplar | term | convention
  title: "..."
  account: "What appears to recur and why it matters"
  standing: proposed-memory
  scope:
    project_refs: []
    source_set_refs: []
  evidence:
    occurrences: []
    outcomes: []
    counterexamples: []
  proposed_contribution: "What reusable part exists"
  situational_elements: []
  limitations: []
  proposed_future_effect: "How context or validation would change"
  activation_conditions: []
  contraindications: []
  review_triggers: []
  provenance:
    source_refs: []
    trust_domains: []
    observed_at: []
    derived_by: []
  acceptance_authority: not-reviewed
  canonical_target: null
  revocation_and_rollback: null
  related_or_superseded_candidates: []

Repeated observations strengthen evidence. They never auto-approve.

6. Pattern review begins with metadata and returns to evidence

Metadata helps locate:

• completed documents;
• workflow and stage records;
• accepted tags;
• decisions and annotations;
• revisions and lineage;
• source-use receipts;
• completion syntheses;
• project relationships;
• repeated failures or repair types.

Metadata is an index, not proof. Provenance can show where a candidate came from, when it was derived, and who accepted it; it cannot establish that the candidate is true. Pattern review must descend into exact artifacts, passages, evidence, outcomes, and counterexamples.

7. Evidence requirements for promotion

A candidate should establish:

• recurring occasion or need;
• instances sufficiently independent to support transfer;
• the distinctive contribution of the proposed method;
• conditions under which it helped;
• what varied across cases;
• counterexamples and misuse conditions;
• what remains situational judgment;
• scope: movement, project, person, practice, domain, or organization;
• what future context or operation would change;
• validation still required;
• provenance and privacy boundary.

The question is not merely “Could this be useful again?” It is whether a future user can recognize when it applies and when it does not.

Promotion also crosses five different validity questions. Is the proposal structurally well formed? Is the semantic case adequate? Does it name the correct canonical object and revision? Does the accepting authority have jurisdiction over the proposed scope? Did the exact intended file and metadata effects occur? A positive answer at one layer cannot substitute for the others. Chapter 22 develops this trusted effect path in operational detail.

8. Two co-equal routes

Evidence-derived promotion

A pattern emerges from completed work and undergoes review.

Direct creation or import

A person, team, researcher, or partner already knows the method or standard they want to preserve.

This route is distinct from ordinary artifact or source admission. Admission gives an existing file stable local identity, provenance, and a declared custody state so the domain can address it. It does not make the file reusable expertise, accept its claims, register it as a governing resource, or activate it in context. Direct creation or import in this section names the later, governed route by which a deliberately formed expertise resource becomes eligible for reuse.

The upper path tests recurrence and outcomes; the lower path tests identity, format, provenance, and scope. They meet at approval, after which the resource is registered and available for later context selection rather than automatically used.

Imported authority does not imply universal scope. Direct resources still need version, provenance, limitations, and retirement paths.

9. The delayed memory-write boundary

Persistent memory is a delayed action surface. A source read today can alter decisions months later if it is silently summarized, indexed, or retrieved as trusted ground. AIOS therefore separates permission to observe from permission to influence future contexts.

A memory proposal should remain inert until the appropriate process has:

1. preserved the exact source and trust domain;
2. distinguished source content from the system’s interpretation;
3. checked whether the observation is current, corroborated, contradicted, or adversarial;
4. declared the proposed scope and future use;
5. identified the authority required for acceptance;
6. written any accepted result through an exact, reviewable operation;
7. produced a receipt naming the changed artifacts and revisions.

This boundary applies even when the model encountered the material through an ordinary question. Query access must not become an indirect route for installing guidance, expanding permissions, or poisoning a future context.

10. Project-local and reusable intelligence

Keep knowledge project-local when its validity depends on:

• one source set;
• temporary constraints;
• stakeholder agreement;
• one artifact;
• a local vocabulary;
• a one-time exception;
• context not expected to transfer.

Promote when:

• the reusable identity is clear;
• contribution separates from case content;
• counterexamples have been considered;
• occasion and boundaries are recognizable;
• future benefit justifies context and governance cost;
• person authority accepts the scope.

Moving a file to an expertise folder does not make its claims general.

11. Registration is availability

Registration gives a resource:

• stable identity;
• version;
• type;
• owner;
• scope;
• provenance;
• evidence and limitations;
• standing;
• source path;
• relationships;
• validation and acceptance records.

It makes the resource resolvable by the system.

Registration does not mean:

• include it in every context;
• execute it automatically;
• treat it as universally true;
• rewrite prior artifacts;
• prevent competing resources;
• convert it into deterministic policy.

12. Activation is contextual

An accepted resource enters a judgment only when:

• the current occasion matches;
• scope applies;
• version and source are current;
• contraindications do not dominate;
• the person's or project's accepted configuration permits it;
• context composition selects it at useful resolution.

The capsule should make the resource's role visible, especially where its guidance is consequential.

13. Promotion authority follows scope

The system can evolve at several scopes:

A resource does not move upward in scope merely through use. Each expansion is a new claim requiring evidence and authority.

A model may discover and articulate a candidate. It does not thereby acquire authority to accept the candidate. Project, person, team, organization, and domain-package scopes each require an explicitly named owner or delegated process. The same resource may be accepted at one scope and remain only proposed at another.

14. Revalidation and decay

Promoted knowledge can become stale when:

• source standards change;
• repeated use produces counterexamples;
• the operating environment changes;
• a model change alters behavior;
• scope expands beyond evidence;
• an external regulation changes;
• the original purpose no longer bears.

The system can reduce routine exposure, flag for review, revise, retire, or supersede the resource. It should not silently rewrite it.

15. Competing practices can coexist

Two accepted methods may remain valid under different conditions.

Example:

• a rapid single-pass review for low-risk internal drafts;
• an independent two-pass review for contested public claims.

The registry can preserve both and let context composition select by consequence and scope. Internal consistency means the distinction is visible.

16. Canonical artifacts stay local; access structures remain rebuildable

The corpus of canonical files and accepted ground remains inspectable. Summaries, search indexes, embeddings, extracted facts, relationship graphs, recommendation indexes, and Atlas views may accelerate access, but they remain derived and rebuildable unless a narrowly defined transactional invariant requires otherwise. Deleting and regenerating an index must not erase accepted knowledge or change its standing.

Each self-contained domain knowledge system remains independently intelligible and portable without hidden provider memory: its purpose, sources, terminology, standing, decisions, and lineage can be reconstructed from owned ground. Self-contained does not mean isolated. A domain can import evidence, use remote models, exchange attestable packages, and collaborate with other domains through explicit disclosure and acceptance boundaries.

This supports:

• person ownership;
• provider changes;
• exact provenance;
• selective sharing;
• independent review;
• local retirement and deletion;
• team or partner packages;
• reconstruction without a hidden service.

When remote models are used, selected context may still leave the device. Local custody of the canonical corpus is not identical to fully local inference.

17. Failure modes

Ambient learning

The system silently infers and installs what the person believes.

Delayed memory poisoning

Untrusted or interaction-generated content becomes future guidance without a promotion boundary.

Repetition bias

Common events are mistaken for useful transferable method.

Folder promotion

Moving content changes its authority without review.

Evidence-free standardization

A persuasive resource gains organization-wide scope without counterexamples or testing.

Registration enforcement

Available knowledge becomes mandatory.

Stale expertise

Promoted resources remain active after their conditions fail.

Invisible provenance

Future users cannot tell where a method came from or who accepted it.

Promotion overload

Candidate generation creates more review burden than useful learning.

18. Research connection

Current memory-security research gives direct support to the delayed-write boundary. AgentPoison showed that poisoning less than 0.1% of an evaluated agent memory could produce high downstream attack success under its direct-write threat model. MINJA demonstrated a query-only path in which an interaction induced persistent malicious memory that later affected behavior. These bounded benchmarks do not show that an AIOS implementation is secure; they show why read access, memory proposal, acceptance, canonical write, and later activation must remain separate authority transitions.

The broader evidence and limitations are summarized in the relational-memory research brief. AIOS contributes the product architecture that connects those risks to promotion authority, canonical files, receipts, and future context composition.

19. Research evaluation

Evaluate:

• precision and usefulness of formation detection;
• evidence quality of candidates;
• false-promotion and missed-promotion rates;
• transfer of promoted methods to independent cases;
• user ability to inspect and reject learning;
• scope discipline;
• revalidation and retirement behavior;
• effect of promoted knowledge on context size and instruction conflict;
• ability to preserve competing practices;
• improvement in future work relative to governance cost.

20. What this page contributes to the whole

Registration and promotion close the Fractal Seed’s reintegration loop. Local work does not disappear when a response ends, and generated language does not become knowledge merely because it exists. Experience can become reusable intelligence only through visible standing changes, evidence, scope, authority, exact canonical effects, and later selective activation. The next chapter shows how those distributed relationships can be mapped without turning the map into a new center of truth.