flowchart LR Intent["Human intent"] --> HumanControl["Human control-envelope check"] HumanControl --> Constitution["Alignment and constitution"] Constitution --> Authority["Governance and authority"] Authority --> Plan["Planning and semantic IR"] Plan --> WorldModel["Qualified world-model branches"] WorldModel -. "prediction, uncertainty, discrepancy; no authority" .-> Plan Plan --> Memory["Virtual context memory"] Memory --> Route["Routing and specialist cores"] Route --> Reason["Reasoning, claims, and verification"] Reason --> Execute["Labor OS and runtime adapters"] Execute --> Artifacts["Artifacts, audit logs, and replay"] Execute --> Operations["Governed operations and incident control"] Operations -. "degrade, hold, recover, or decommission" .-> Execute Artifacts --> BehavioralEvidence["Behavioral evidence and benchmark ratchets"] Reason -. "scoped internal observations" .-> WhiteBox["White-box evidence packets"] BehavioralEvidence --> Evidence["Claim and evidence transition"] WhiteBox -. "preserve, restrict, escalate, reject, or expire; never grant" .-> Evidence Execute -. "independent observations" .-> RealityResidual["Reality-residual records"] RealityResidual -. "re-estimate, replan, fallback, review, or hold" .-> WorldModel Operations --> Evidence Evidence --> Improve["Governed self-improvement"] Improve --> Training["Governed training-run transaction"] Training --> BehavioralEvidence Training -. "identity-bound candidate; no release authority" .-> Evidence Improve --> Authority Authority --> Privacy["Purpose-bounded information use and data rights"] Privacy -. "rights, privacy, and residual constraints" .-> Memory Authority -. "permission ceilings" .-> Execute Memory -. "source-bound packets" .-> Reason Evidence -. "promotion or rollback" .-> Route Evidence -. "qualification records" .-> Improve
The ASI Stack
A Governed Systems Architecture for Advanced AI, with ASI as the Stress Case
Living Book Baseline
Current canonical metrics (generated from machine records): 87 manifest chapters; 491 public-safe records; 87 chapter-core claims; 87/87 chapters externally positioned; 0 promoted core claims; 115/115 accepted transitions identity-resolved (25 direct, 61 subclaim, 29 proxy; 0 parent movements).
The ASI Stack: A Governed Systems Architecture for Advanced AI, with ASI as the Stress Case is a living technical book by Corben Sorenson.
Choose the product you need
| Product | Best for | Start here |
|---|---|---|
| Narrative book | A 22-unit thesis-to-method route over the complete 87-chapter reference, with live research scaffolding hidden. | Open the narrative route |
| Architecture reference | Interfaces, invariants, failure routes, protocols, tests, proofs, and implementation horizons across all 87 working-manifest chapters. | Open the architecture index |
| Evidence registry | Current commit/count state, claim support, sources, tests, proofs, releases, and residuals. | Open the evidence registry or inspect the claim matrix |
| Original paper library | Corben’s exact source manuscripts and architecture papers, separated from the book’s later synthesis. | Read the papers and their chapter lineage |
Read the current book
This live site is the canonical current 87-chapter book. Use Human view for the reader-facing projection or the 87-chapter architecture index for direct technical lookup. The generated reader pipeline currently derives all 87 chapters and passes the reader-spine contract. Older downloadable artifacts remain in the GitHub release archive, but they are not the complete current book and therefore are not presented here as the current edition.
All three products derive from one canonical source tree; their density, navigation, artifact, and review boundaries are defined in docs/product_contracts.md, with executable projection details in docs/product_projection_artifacts.md. The Post-v2.3 Claim Proof, Causal Validation, and SOTA-Challenge Roadmap is complete with an exact terminal no-public-release record. The Post-v2.3 Evidence Competence, Transfer, and Publication Roadmap is the sole active successor, with machine authority in roadmap_records/post_v2_3_maintenance_transfer_and_publication_status.json, an N0–N5 claim-bearing experiment standard, and a complete accepted-transition identity crosswalk. All 115 accepted transitions now resolve through 25 exact atom, 61 bounded subclaim, and 29 proxy relations with zero indirect parent support movement. The competence contract prevents weak implementations, invalid instruments, bad proxies, and underpowered tests from becoming negative evidence about an architecture. The live working tree has 87/87 chapter-core claims at argument, zero chapter-core promotions, zero external reproductions, and no SOTA support; the local X synopsis source is current, but its older unpublished platform draft is stale and must be refreshed before publication. v2.3.0 remains the latest immutable public living-book release.
Current status
Book state v2.3.0 is the latest completed immutable HTML release
Sources 491 public-safe records; 87/87 chapters externally positioned
Claims 87 chapter core claims remain argument-level
Tests 25 narrow non-core evidence transitions accepted; no chapter-core promotion
Visual edition preview
No visual abstracts are currently linked from the living book. The first-generation 1–12 predecessors were withdrawn from the current projection and preserved as historical custody while generation-two replacements are reviewed. Chapters 13–18 have no YouTube object yet; generation-two 19–24 are unlisted candidates, not current embeds.
60-Second Trust Surface
What this is. This is a public living-book research program for governed advanced-AI systems, with ASI used as the extreme stress case. Its strongest current contribution is the evidence discipline: manifest-driven chapters, source queues, claim/support-state separation, Lean hooks, schemas, harnesses, release records, and reader-edition gates.
What this is not. This is not a validated ASI implementation, not a deployed safety system, not a benchmark-proven architecture, and not a claim that the chapter theses are already externally proven.
Auditable current state. The inventory has 491 public-safe records; 87/87 chapters are externally positioned with 0 explicit external-baseline exceptions. All 87 chapter core claims remain at argument; the core-claim disposition ledger records 87 per-chapter core-claim dispositions, 22 accepted no-change transition dispositions, 65 accepted no-promotion dispositions, and 0 promoted core claims. The 25 accepted non-core upward evidence transitions are recorded in the non-core evidence ledger, alongside 61 accepted blocks_promotion decisions and three historical refuted labels. The competence audit classifies all 90 accepted negative/no-change records as 1 N0, 15 N1, 74 N2, and 0 N3–N5, so no exact, broad, parent, or chapter-core refutation currently follows.
Boundaries readers should keep in view. Appendix C is the claim/support-state ledger. Appendix G separates Corben-authored, Corben-supplied, and local-project sources from Appendix H, which holds external literature; external positioning is not exhaustive literature synthesis. The original paper library makes 50 digest-bound Corben manuscripts readable in the live book and shows which chapters they informed. Those pages preserve intellectual lineage; they do not make a historical paper’s claims current, correct, reproduced, novel, or evidence-promoted. Novelty positioning is not proof of novelty. Human view is a convenience projection, not a reviewed reader-release manuscript; its current heuristic queue is in the reader continuity audit. By author decision, no external-human review or outreach is a prepublication gate; the preserved specialist packets may be used only after completion, and no independent review is claimed in the review ledger.
Fast audit path. Start with the machine-readable canonical public status, the active Evidence Competence, Transfer, and Publication Roadmap, its experiment-competence standard, and its machine status; then the completed Claim Proof, Causal Validation, and SOTA-Challenge Roadmap, its machine status, and terminal record; then the v2.3.0 completion declaration, exact release record, historical v1.x roadmap, v1.0 candidate status, claim ledgers, and Appendices C, G, and H. All 87 chapter-core claims remain at argument. CI rejects contradictory counts, release or roadmap identity drift, manifest/rendered navigation drift, duplicate chapter links or titles, missing chapter pages, and multiple-H1 chapter pages.
Core thesis
Efficient ASI should not be treated as one giant opaque model. It should be treated as a governed cognitive stack: alignment, governance, planning, memory, reasoning, execution, routing, compression, evidence, and recursive self-improvement cooperating through typed boundaries, verification, artifact memory, and bounded authority.
The central design move is simple: never let intelligence, authority, memory, evidence, and execution collapse into one undifferentiated loop. A capable system can generate, plan, retrieve, reason, route, and act; a governed system also records which layer had permission to do each thing, which artifact carried the decision, which evidence justified it, and which rollback path remains available.
Architecture at a glance
Every arrow is meant to become an interface, not just a metaphor. The book is organized so each chapter owns one boundary: what enters it, what artifact it emits, what invariants must hold, what can fail, and what evidence would justify stronger claims.
The human-control arrow is deliberately non-authorizing. Its packet can show that a named reviewer lacks information, comprehension opportunity, workable load, time, authority, alternatives, or an effective intervention path and can therefore force added capacity, reduced autonomy, safe hold, or abstention. It cannot infer a person’s internal state, assign moral responsibility, approve an effect, or certify meaningful control.
The operations loop is equally non-promotional. Its control packet can narrow capability, data, tools, population, and duration; reconcile declared internal state and external-effect dispositions; expire emergency authority; and keep an unknown effect in safe hold. It cannot prove that the inventory is complete, that an effect was reversed, that fallback is useful, or that recovery is safe.
Dependency spine
The chapter list is dynamic, but the argument should remain stable when chapters are inserted, moved, split, or merged. The living spine is the dependency order below: governance defines what may happen, operational layers lower intent into work, evidence decides what can improve, and the release layer keeps the book itself honest.
flowchart TD Unit["Stack as unit of analysis"] --> AuthorityModel["Boundaries, authority, and failures"] AuthorityModel --> EvidenceLanguage["Evidence states and claim discipline"] EvidenceLanguage --> IntentLayer["Human intent, constitution, agency, and value conflict"] IntentLayer --> ChangeRights["Governance rights, stable fields, replacement, security, and RSI gates"] ChangeRights --> ControlSpine["Intent contracts, commands, plans, semantic IR, and context ABI"] ControlSpine --> EpistemicSpine["Context adequacy, claim ledgers, proof-carrying claims, and tribunal review"] EpistemicSpine --> ExecutionSpine["Typed jobs, artifacts, adapters, and procedural memory"] ExecutionSpine --> CapabilitySpine["Routing, replaceable substrates, governed training, readiness, and personal compute hives"] CapabilitySpine --> EfficiencySpine["Compact generation, fast generation, deliberation, compression, and resource budgets"] EfficiencySpine --> SubstrateSpine["Mathematical/search substrates, Circle contracts, coil memory, and cyclic mixers"] SubstrateSpine --> EvidenceSpine["Executable specs, benchmarks, white-box evidence, thresholds, adversarial evaluation, safety cases, and operations"] EvidenceSpine --> IntegrationSpine["Integrated reference architecture, Theseus reference path, prototype roadmap, living method, and research backlog"] IntegrationSpine --> EvidenceLanguage
Precision follows the same spine rather than becoming a hidden tensor setting. The compression owner defines a protected-behavior contract and complete executable package; generation routes base, residual, reference, or abstention; resource accounting closes physical and assurance costs; executable specifications bound transformations; custody binds derivatives and platforms; and readiness expires or revokes the resulting certificate. The live book therefore treats precision as a governed cross-layer resource, not a universal bits-per-weight fact.
Chapter roles
Breadth does not mean that all 87 chapters make equally central or equally supported claims. The current reader-role audit assigns every chapter one primary job:
| Role | Chapters | Reader contract |
|---|---|---|
| Thesis-bearing | 11 | Advances a central architectural claim and states the strongest alternative, failure case, and evidence that could change it. |
| Load-bearing reference | 57 | Owns an interface, invariant, artifact, or failure family needed to implement or audit the thesis. |
| Implementation case | 7 | Shows how contracts compose in artifacts, runtime effects, reference traces, prototypes, operations, or the living-book workflow. |
| Speculative research | 12 | Preserves promising mechanisms and research programs in the current book without presenting argument-level possibilities as established architecture. |
flowchart LR Thesis["Thesis-bearing chapters<br/>why the architecture should exist"] --> Reference["Load-bearing reference<br/>what each boundary must own"] Reference --> Cases["Implementation cases<br/>how boundaries compose and fail"] Reference --> Research["Speculative research<br/>what must still earn stronger evidence"] Cases --> Evidence["Observed effects, residuals, and revision"] Research --> Evidence Evidence --> Thesis
The role is editorial, not evidentiary. A thesis-bearing chapter is not thereby proved; a reference chapter is not secondary in implementation; an implementation case does not imply deployment; and a speculative chapter is not rejected or deferred from the manuscript. The complete current classification is recorded in evidence_quality/current_chapter_role_map.json; the P7.1a-W2 audit remains the immutable historical baseline for its 60-chapter editorial pass.
Reading path
The shortest path through the book is:
- Why a stack: ASI must be decomposed into governed layers before it can be made efficient, testable, or corrigible.
- How work moves: human intent becomes command contracts, plans, context packets, claims, jobs, artifacts, and evidence records.
- How capability grows: routing, compression, procedural memory, benchmarks, and stable capability fields let the system improve without erasing its safety boundary.
- How the book stays alive: claims, proofs, schemas, tests, source notes, and public renders form the book’s own evidence ratchet.
How to read
Start with Part I for the architecture thesis, boundary model, and governance substrate. Read Part II as the operational spine: intent, planning, memory, reasoning, execution, and artifacts. Part III covers routing, replaceable substrates, governed training, compression, representation, mathematical substrates, and cyclic/proof-carrying mechanisms. Part IV explains evidence, implementation, Project Theseus as a reference path, and the living-book process.
The appendices are part of the architecture:
- Appendix A tracks source assignments.
- Appendix C tracks claim support states.
- Appendix D tracks protocol schemas.
- Appendix E tracks planned Codex tests.
- Appendix F tracks living updates.
Evidence discipline
Every major claim should carry a support state. The default state is argument. Claims move only when sources are actually read, prototypes are inspected, proofs are checked, or tests are actually run. This book should never pretend that source ingestion or test evidence exists before it does.
Living status
The mutable root site and /latest/ are the canonical current publication surfaces and contain all 87 manifest chapters. Versioned tags and GitHub Releases are immutable historical snapshots for exact citation and reproduction; they do not override the current chapter count, order, source inventory, or claim state.