Claims Register, Open Questions, and Next Research

1. The register keeps the publication answerable to evidence

A complete architectural account can easily sound like a complete empirical claim. This register prevents that collapse. It identifies what governs the design, what the system is building toward, what has been observed or demonstrated, what remains hypothetical, what could disconfirm it, and which research comes next.

The register is designed to change through explicit evidence and authority. A closed item records what closed it. A narrowed claim preserves the reason and prior scope. A failed hypothesis becomes part of the architecture's learning history.

2. Claim interpretation rules

1. RATIFIED establishes architecture, not empirical superiority.
2. TARGET establishes intended behavior, not current implementation.
3. OBSERVED requires direct source or runtime evidence.
4. DEMONSTRATED requires a declared end-to-end proof.
5. INFERENCE marks a reasoned first-party interpretation capable of error.
6. CANDIDATE identifies developed future work awaiting acceptance.
7. HYPOTHESIS states a falsifiable predicted effect.
8. OPEN requires a founder decision, implementation result, or experiment.
9. Failure can narrow one claim without invalidating the full architecture.
10. One favorable task result never establishes general intelligence.

3. Foundational claim register

4. Context, memory, and artifact claims

5. Work, agency, and governance claims

6. Economic and sovereignty claims

7. Research-positioning claims

8. Capability register

9. Reference journey register

10. Future requirement register

11. Epistemic and decision register

12. Priority open questions

Context and cognition

1. What operational test defines the smallest complete context?
2. How should standing, relevance, recency, source quality, and diversity trade off?
3. Can the context compiler be evaluated separately from the model?
4. Which cognitive movements need distinct ground?
5. Does the Fractal Seed improve work beyond ordinary hierarchy and expert authorship?
6. How can material loss remain visible without burdening routine work?

Memory and correction

1. Which memory forms are necessary in product rather than only conceptually useful?
2. When should accepted guidance decay or enter revalidation?
3. How should conflicting accepted claims coexist before resolution?
4. What consequence justifies proactive contradiction or drift notice?
5. How does source descent remain usable across many synthesis generations?
6. What balance among active ground, dormancy, archive, and deletion best protects continuity?

Human authority and agency

1. How should sub-agents surface prepared actions and context as the person's attention moves?
2. How can authorization remain expressive without constant confirmation?
3. What makes a return sufficiently independent and complete?
4. When should background work surface, wait, or expire?
5. How do recommendations remain useful without narrowing exploration?
6. Which interaction makes negative authority understandable?

Files, metadata, and Atlas

1. What is the minimum companion memory needed for reconstruction?
2. How are artifact-shadow conflicts repaired without hidden authority?
3. Which relationships require exact registration?
4. How can semantic review candidates avoid appearing exact?
5. Which trace formats permit external inspection without changing canonical ownership?
6. How should schema evolution preserve long-lived local knowledge?

Economics and sovereignty

1. How much knowledge-work cost comes from repeated context reconstruction?
2. Which bounded tasks permit less expensive models to benefit from mature ground?
3. How much switching cost is semantic rather than infrastructural?
4. How should context authoring and maintenance labor be valued?
5. Which privacy benefits survive actual remote-model use?
6. Do semantic traces reduce assurance effort or relocate it?
7. Which institutions can govern attestable domain packages without gatekeeper capture?
8. How can decentralized exchange preserve provenance, privacy, incentives, and pluralism?

Research contribution

1. Which existing systems have the nearest full-stack relationship?
2. Which contribution survives exact claim charting?
3. What minimum implementation supports independent replication?
4. Which ablations reveal genuine interaction effects?
5. How much result quality comes from founder or expert curation?
6. Does the architecture transfer beyond research, writing, planning, and software work?

13. Prioritized research program

Program A — Minimum complete composition. Build and independently review Journey A from clean sources through provider-visible context, artifact and shadow settlement, interface state, termination, and reload.

Program B — Context composition and renewal benchmark. Compare transcript, human summary, lexical retrieval, embedding retrieval, broad context, one continuous extended trace, and AIOS composition on identical tasks, sources, models, and token budgets. For sustained multi-faceted tasks without a native verifier, hold the total token allowance constant while comparing one static-context trace with staged judgments whose accepted intermediate products renew the next situation; test whether any difference grows with horizon length.

Program C — Correction across interruption. Run Journey H after a meaningful delay with credible contradiction; test source recovery, scope, consumer precision, acceptance, supersession, and later judgment.

Program D — Attention-following asynchrony. Compare synchronous work, unmanaged parallel work, and bounded commissions on blocked attention, review burden, duplication, return quality, integration, and control.

Program E — Practice compounding. Compare no reuse, static templates, ungoverned reflection, and governed promotion across repeated work, including harmful carryover and maintenance cost.

Program F — Provider portability and task amplification. Move one mature project among at least two eligible remote models and one eligible local model; measure adaptation, data conversion, quality, cost, continuity, and revalidation.

Program G — Atlas and audit reconstruction. Ask independent reviewers to reconstruct a consequential decision and change impact from canonical files and derived views; measure accuracy, time, missing lineage, and false candidates.

Program H — Organizational and decentralized pilots. Use the existing workflow and domain-package library to test shared practice, local variation, updates, revocation, and privacy-preserving return of candidate improvement.

14. Evidence thresholds

15. Research data and reproducibility

Where privacy permits, every study preserves its task and acceptance contract, source corpus or reproducible equivalent, authority and standing map, model identity and settings, exact provider-visible context, prompt and response contract, outputs and failures, person actions, artifact revisions, tokens and costs, reviews, interruption or reload state, baselines, ablations, rater protocol, analysis, limitations, null results, and adverse outcomes.

Data excluded for privacy remains described by type, reason, and consequence for replication. Reproducibility never requires surrendering personal ground beyond its consent and authority boundary.

16. Publication maintenance cycle

Every revision keeps architecture, implementation, demonstration, and hypothesis standing distinct. Review evidence stays tied to the version it examined.

17. Completion condition for the publication draft

The recommissioned publication reaches structural completion when the Overview, Core Thesis, Chapters 01–26, Addendum Chapters 27–28, diagrams, source authority, claims, and open questions all exist in order; architectural vocabulary is consistent; external research has not redefined the system; and originality, diagram, and source checks pass.

Structural completion does not confer founder approval, resolve external review, establish current implementation, or prove the economic and research hypotheses. Those are separate gates with their own authority and evidence.

The claims register ends the publication where the architecture itself continues: with explicit purpose, developed ground, visible standing, and a concrete program for learning what is actually true.