Appendix B — Glossary
These are owned working definitions. A term is either assigned to one active chapter that controls its interface and failure boundary, or marked as a source-lineage label that carries no normative architectural authority.
| Term | Working definition | Owner | Status |
|---|---|---|---|
| ASI stack | A proposed substrate-neutral governed architecture of logical responsibility and authority boundaries across alignment, governance, planning, memory, reasoning, execution, routing, compression, evidence, and improvement. A monolithic model, modular service graph, hybrid system, or human/AI process may instantiate the boundaries; the term does not require one model or physical module per layer. | asi-is-a-stack-not-a-model |
normative architecture definition |
| Noninheritance law | The cross-stack rule that capability does not confer authority, context does not confer belief or permission, plans do not confer effects, receipts do not confer reality, theorems do not confer runtime enforcement, replacements do not inherit qualification, and self-improvement proposals cannot ratify themselves. | asi-is-a-stack-not-a-model |
opening thesis law; architecture at argument support |
| Governed Transition Calculus | The shared transition shape (T=(I_s,I_t,A,O,E,R,C,,)), separating source and target identity, authority, obligations, evidence, residuals, consumers, expiry or revocation, and rollback, compensation, or retirement. Domain owners add stricter fields; the shape alone proves neither its premises nor runtime enforcement. | executable-specifications-and-lean-proof-envelope |
shared interface vocabulary; no universal theorem or implementation claim |
| Developmental Intelligence Loop | The governed capability-development lifecycle from curriculum, interaction, prediction error, abstraction, and causal intervention through memory consolidation, proceduralization, skill composition, stabilization, readiness review, and authority-gated promotion. | governed-model-training-distributed-optimization-and-scaling |
cross-owner lifecycle at argument support |
| Minimum trusted kernel | The smallest explicit trust budget over identity, time, authority, policy, context resolution, artifact custody, observation, evaluation, sandboxing, rollback, and release whose failure can invalidate a governed transition. | security-kernel-and-digital-scifs |
trust-minimization target; no secure implementation claim |
| Bounded liveness | The requirement that admitted tasks, residuals, quarantines, and review loops reach finite owned dispositions while useful throughput, false blocking, latency, human time, compute, recovery, residual burden, and governance cost remain visible beside safety. | security-kernel-and-digital-scifs |
joint safety/usefulness target |
| P0–P6 semantic depth | A non-promotional classification of formal or executable reach: P0 record shape, P1 finite route, P2 reachable witness, P3 implementation binding, P4 cross-component safety, P5 liveness or recovery, and P6 empirical binding to a named observation contract. | executable-specifications-and-lean-proof-envelope |
proof-scope organization; levels must be earned independently |
| Alignment and constitution | The layer that constrains admissible goals, values, self-modification, and human-agency requirements. | constitutional-alignment-substrate |
owned interface definition |
| Stable Capability Field | A stable capability boundary with replaceable implementations, bounded authority, qualification evidence, and rollback rules. | stable-capability-fields |
owned interface definition |
| PlanForge | A source-lineage label for compiling goals into governed task graphs, dependencies, budgets, and replanning triggers; the active interface owner is Planning as a Control Layer. | planning-as-a-control-layer |
lineage label; not a separate normative layer |
| Virtual Context Memory | A source-lineage label for governed context materialization from durable sources; the active interface owner is the Virtual Context ABI. | virtual-context-abi |
lineage label; not a separate normative layer |
| Talos | A source-lineage label for typed jobs, execution, artifacts, permissions, audit, and replay; active ownership is split across the named execution chapters. | labor-os-and-typed-jobs |
lineage label; not a separate normative layer |
| Spinoza | A source-lineage label for claim verification and belief revision; active ownership is split between Claim Ledgers and Proof-Carrying Claims. | spinoza-verification-and-proof-carrying-claims |
lineage label; not a separate normative layer |
| MoECOT | A source-lineage label for modular orchestration with specialist cores, readiness gates, ledgers, and replay. | routing-heads-and-specialist-cores |
lineage label; not independent implementation evidence |
| Octopus Router | A source-lineage label for dynamically selecting bounded specialist arms or cores. | routing-heads-and-specialist-cores |
lineage label; routing interface owned by chapter |
| Ratcheting Modular Intelligence | A source-lineage doctrine based on benchmark frontiers, residual escrow, regression preservation, and verified module promotion. | benchmark-ratchets-and-anti-goodhart-evidence |
lineage label; not an independent evidence class |
| Compact Generative Systems | The compression owner for selecting the smallest adequate generative or governing structure without hiding residual complexity. | compact-generative-systems-and-residual-honesty |
owned interface definition |
| RankFold / NeuralFold | A compression source-lineage label for low-rank residual coding, functional preprocessing, and fallback routing. | rankfold-neuralfold-and-artifact-compression |
lineage label; mechanism claims require separate evidence |
| Functional precision | The minimum numerical and representational detail needed by a complete executable implementation to preserve named behavior under a declared domain, metric vector, tail rule, resource envelope, and fallback policy. It is contract-relative and not a universal bits-per-weight property. | rankfold-neuralfold-and-artifact-compression |
owned compression-interface definition; no optimality or preservation result |
| Precision Contract | A versioned contract binding protected behavior, domain and exclusions, metrics and thresholds, confidence, physical resources, escalation, fallback, and exact reference, transformed, decoder, runtime, platform, and evaluator identities. | rankfold-neuralfold-and-artifact-compression |
owned admission artifact; contract selection and satisfaction require separate evidence |
| Precision field | A governed allocation of numerical precision across layers, channels, directions, tokens, tasks, uncertainty states, and consequence classes, optionally using ordered residual planes and a conservative route. | rankfold-neuralfold-and-artifact-compression |
design-rationale resource allocation; no routing advantage claim |
| Precision Certificate | A scoped, expiring receipt that binds one transformed executable package to its protected contract, evidence, cost ledger, route, fallback, limits, platform, and revocation state. | readiness-gates-residual-escrow-and-quarantine |
readiness input only; not optimality, safety, support, or deployment authority |
| BBVCA | A generate-verify-repair source-lineage label using seeded local laws, bounded search, and residual tracking. | compact-generative-systems-and-residual-honesty |
lineage label; mechanism claims require separate evidence |
| Artifact graph | A traceable graph of produced objects, source handles, jobs, claims, evidence records, projections, revocations, and replay grades. | artifact-graphs-audit-logs-and-replay |
owned interface definition |
| Claim ledger | A structured lifecycle record of claims, support states, sources, uncertainty, contradiction links, revisions, and review status. | claim-ledgers-and-belief-revision |
owned interface definition |
| Context packet | An active bounded context materialization supplied to an agent or job, including source handles, summaries, authority labels, and adequacy status; it is not durable memory. | virtual-context-abi |
owned interface definition |
| Proof-carrying summary | A bounded summary carrying provenance and checks sufficient to audit the stated fidelity obligation without proving the broader source claim. | spinoza-verification-and-proof-carrying-claims |
owned verification artifact definition |
| Authority ceiling | The maximum action authority a principal, layer, field, tool, or implementation may exercise under a named grant and policy state. | system-boundaries-and-authority |
owned invariant definition |
| Readiness gate | An evidence, authority, regression, and residual checkpoint that must pass before a capability or artifact may enter a stronger lifecycle state. | readiness-gates-residual-escrow-and-quarantine |
owned lifecycle definition |
| Residual escrow | A visible owned ledger of unresolved failures, uncertainty, debt, or residual complexity left by a route, compression, benchmark, or release decision. | readiness-gates-residual-escrow-and-quarantine |
owned lifecycle definition |
| Benchmark ratchet | A governed benchmark lifecycle that preserves floors, negative cases, provenance, contamination controls, and residuals while moving an evaluation frontier. | benchmark-ratchets-and-anti-goodhart-evidence |
owned evidence-interface definition |
| Training Run Transaction | A prospectively frozen, identity-bound record joining architecture, data order, objective, optimizer, numerical and topology policy, execution failures, full checkpoint state, resume evidence, checkpoint-family selection, and independent qualification handoff. | governed-model-training-distributed-optimization-and-scaling |
owned training-integrity interface definition; no model-quality authority |
| Optimizer Policy Card | The executable optimizer identity for one run: family and implementation, parameterization, eligible parameter groups and fallback, state and precision, update equation, schedule or averaging state, decay, clipping, stabilizers, preconditioner or curvature cadence, approximation, batch/topology semantics, tuning budget, rescue ladder, checkpoint state, and nonclaims. | governed-model-training-distributed-optimization-and-scaling |
owned training-policy definition; an optimizer family name alone is insufficient |
| Optimizer family | A class of update mechanisms sharing a broad information or geometry choice—such as coordinate adaptation, tensor preconditioning, curvature approximation, sign updates, orthogonalized matrix updates, or schedule/averaging policy. A family does not specify a runnable configuration or imply superiority. | governed-model-training-distributed-optimization-and-scaling |
taxonomy term; no evidence or selection authority |
| Method-specific rescue | A prospectively bounded debugging and tuning opportunity tailored to an optimizer’s known implementation requirements, applied before a failed arm may support a negative inference and charged to its tuning budget. | governed-model-training-distributed-optimization-and-scaling |
competence-gate definition; cannot use held-out outcomes |
| Relational Dimension Compiler (RDC) | A Corben source-lineage architecture label for constructing a typed relational topology, choosing qualified operators, reifying accepted relations, preserving branch/evidence lifecycle, and reversibly contracting stable subcomplexes. | relational-dimension-compilation-and-polyadic-cognition |
source-lineage label with a current distinct argument-level owner; no implementation, benchmark, or support promotion follows |
| Semantic / computational / storage arity | Three separate properties: the number of role-bearing arguments in a relation, the number of items jointly evaluated by a primitive operator, and the arity of the durable representation. None determines the others. | replaceable-cognitive-substrates-beyond-transformer-monoculture |
owned substrate-interface distinction; no universal bounded-arity claim |
| Typed relation object | A versioned relation instance with schema, role bindings, time, branch, uncertainty, provenance, defeaters, lifecycle, and permitted consumers; a transient interaction score is not a relation object. | governed-world-models-and-reality-grounding |
owned predictive-state distinction; does not establish truth or authority |
| Semantic renormalization | Query- and environment-relative contraction of a qualified relational subcomplex into a macro-object with a boundary interface, error envelope, provenance, and mandatory expansion or invalidation triggers. | governed-world-models-and-reality-grounding |
design-rationale abstraction term; no contraction-fidelity result |
| RODIE | The proposed Relational Order and Dimensional Intelligence Evaluation suite for minimum order, roles, topology, interventions, dynamic identity, contraction, object–field coupling, branches, compilation, and total cost. | benchmark-ratchets-and-anti-goodhart-evidence |
proposed instrument family; no corpus, run, or result |
| Information Lifecycle Transaction | A purpose- and authority-bounded record joining affected parties, minimization, flow and derivative lineage, privacy evaluation, rights execution, residual copies and influence, and a bounded non-compliance receipt. | privacy-data-rights-and-information-flow-governance |
owned privacy/data-rights interface; no legal-compliance authority |
| Purpose Lease | A scoped, expiring statement of the exact information use, claimed authority, jurisdiction, recipients, retention, and compatibility rule. | privacy-data-rights-and-information-flow-governance |
does not prove that the purpose is lawful |
| Resume equivalence class | The prospectively declared strength of a checkpoint-continuation claim: bitwise, operation-order-bounded, or statistical, each with its own comparison, tolerance, and inference ceiling. | governed-model-training-distributed-optimization-and-scaling |
owned recovery-claim definition; a successful load is insufficient |
| Internal Evidence Packet | A model-, checkpoint-, method-, population-, interpretation-, intervention-, residual-, evaluator-, and time-specific record that states the maximum inference allowed from model-internal observations. It may preserve, restrict, escalate, reject, or expire evidence but cannot grant execution, release, or support authority. | white-box-evidence-interpretability-and-activation-governance |
owned white-box evidence definition |
| Activation policy candidate | A separately qualified proposal to observe or intervene on named internal states for a bounded purpose, population, model identity, and monitor window. A useful intervention result is not deployment authority and expires after material model or method change. | white-box-evidence-interpretability-and-activation-governance |
owned activation-governance definition |
| Qualified branch packet | A model-, observation-, checkpoint-, horizon-, intervention-, calibration-, disagreement-, planner-, and authority-bound record for an imagined consequence. Admission permits bounded planning use only; it grants neither truth, support, execution, nor release authority. | governed-world-models-and-reality-grounding |
owned predictive-state interface definition |
| Reality-residual record | A preserved comparison between a prior model-relative prediction and an independently admitted later observation, with materiality, attribution limits, affected descendants, and a required re-estimation, replan, fallback, review, or safe-hold route. | governed-world-models-and-reality-grounding |
owned prediction-to-observation reconciliation definition |
| Human control-envelope packet | A task-, controller-, evidence-, representation-, workload-, time-, authority-, intervention-, conflict-, privacy-, responsibility-, and lifecycle-bound record used to decide whether a nominal human role may proceed to bounded review, needs added capacity or reduced autonomy, or must safe-hold. It cannot infer internal state, assign moral responsibility, grant effect or support authority, or certify meaningful control. | human-factors-and-meaningful-control-in-oversight |
owned operator-capacity interface definition |
| Operational continuity contract | A versioned contract for one exact deployed service that joins identity, observability limits, incident classes, command leases, containment, degraded and fallback modes, state/effect reconciliation, recovery, disclosure, and decommissioning. | governed-operations-incident-command-and-graceful-degradation |
owned post-deployment control interface |
| Effect-complete recovery | A recovery claim that separately reconciles required internal state, descendants, external effects, residual acceptance, fresh independent checks, qualified fallback, and emergency-authority expiry. A loaded checkpoint alone is insufficient. | governed-operations-incident-command-and-graceful-degradation |
recovery-gate definition; no efficacy claim |
| Ceremonial approval | An approval event for which one or more necessary control conditions—relevant evidence, comprehension opportunity, authority, time, viable alternatives, effective intervention, or reachable safe state—were absent or unresolved. The label is a bounded operational diagnosis, not a judgment of the person. | human-factors-and-meaningful-control-in-oversight |
owned human-control failure definition |