PACKSHARD | profile=web-core-index-v2 | lang=en | module=GLOSS | pack_sha256=63206d01ca8cac92ba759c84ef07e9b22a47771c5a9401947bfefe358ff1d555
ADAPTER_DATA | authority=user-entry-prompt | canonical_exact_blocks=true
BEGIN_CANONICAL_GLOSS | bytes=36133 | sha256=0588f1c1dae7ba5bd50b0ea8ca8323ed39c51568469c9c6f1c93e11132525cb0
## GLOSS | Terms and book-specific meanings

note: meaning is not a general dictionary definition; it concisely records the meaning or role the term has in this book from manuscript evidence. Do not extend it beyond the record.
TERM-PREFACE-001: suiten, the point of convergence (萃点) | chapter=CH-PREFACE | review=approved | aliases=萃点
meaning: A concept Minakata Kumagusu placed within the web of the world: a point which, once taken as one's vantage, gathers seemingly scattered threads into a single visible structure. There is no unique suiten; different ones appear depending on where the observer stands. The book rereads the history of science as the discovery of new suiten (Newtonian mechanics, thermodynamics, evolution, quantum mechanics) and positions itself as "an attempt to find the suiten of the AGI era".
TERM-01-001: AGI | chapter=CH-01 | review=approved
meaning: Artificial General Intelligence: a machine not bound to particular tasks, general across a wide range of cognitive domains at or above human level. It contrasts with "narrow AI", which excels in one domain only; systems such as ChatGPT and Claude are viewed as its early forms. The book starts from the premise that non-arrival of AGI is no longer a safe assumption, and analyzes it as bringing not the disappearance of limits but their reconfiguration.
TERM-01-002: scaling laws (スケーリング則) | chapter=CH-01 | review=approved | aliases=スケーリング則
meaning: The empirical regularity (Scaling Laws) that prediction error keeps decreasing in a lawlike way as parameters, data, and compute increase. Within the observed scaling range no clear saturation of improvement has been confirmed, and the regularity is on the way to being theoretically underwritten through its connection to Solomonoff induction. This convergence of empirical law and theory grounds the book's Proposition One: that non-arrival of AGI is no longer a safe assumption.
TERM-02-001: AIXI | chapter=CH-02 | review=approved
meaning: The ideal agent formalized by Hutter in 2005, the book's concrete mathematical formalization of Universal AI. It models the environment via Solomonoff induction and selects actions maximizing expected future reward, acting optimally in every computable environment. Being incomputable, it is treated as a theoretical reference like the ideal gas or Carnot cycle: a north star showing where real AI systems are heading, with AGI a finite threshold far below this ceiling.
TERM-02-002: the No-Free-Lunch theorem (NFL定理) | chapter=CH-02 | review=approved | aliases=NFL定理
meaning: The optimization theorem: averaged with equal weight over all problems, every algorithm performs the same, so no universally optimal algorithm exists. Viewing intelligence as search, the book transposes it to classifying intelligence, placing narrow AI, AGI, ASI, and Universal AI on a performance-generality trade-off under fixed compute. Only more compute enlarges the trade-off's "area"; AGI's difficulty lies in reconciling generality and performance under limited compute.
TERM-02-003: intelligence explosion (知能爆発) | chapter=CH-02 | review=approved | aliases=知能爆発
meaning: I. J. Good's 1965 concept: if a superhumanly intelligent machine can improve itself, the improved machine can design a still better one, producing a rapid chain reaction of increasing intelligence. With LLMs already writing code and assisting researchers, AI accelerating AI R&D is partially realized; the book makes what happens when this self-improvement loop becomes fully autonomous the central question of chapters 4 (physical limits), 5 (control), and 6 (scenarios).
TERM-03-001: grokking (グロッキング) | chapter=CH-03 | review=approved | aliases=グロッキング
meaning: A discontinuous phase transition in which a model, after prolonged training, abruptly shifts from rote memorization of answers (overfitting) to generalization, correctly answering unseen problems. The book interprets it as the moment when individual patterns crystallize into a single algebraic structure within the space of distributed representations, positioning it as a manifestation of the simplicity bias and as experimental support for the emergence interpretation.
TERM-03-002: Solomonoff induction (ソロモノフ帰納) | chapter=CH-03 | review=approved | aliases=ソロモノフ帰納
meaning: The optimal-prediction framework assigning higher prior probability to simpler hypotheses. The book cites independent formal results that transformer training, with next-token prediction pushed to its limit, approaches Solomonoff induction, framing prediction, compression, and Solomonoff induction as one. Converging with replicated scaling laws, this grounds its case for AGI's feasibility — though as idealized individual results, not an unconditional convergence theorem.
TERM-03-003: distributed representation (分散表現) | chapter=CH-03 | review=approved | aliases=分散表現
meaning: A scheme representing a concept not as a single discrete symbol but as a vector of hundreds to thousands of numbers, encoding meaning across many values. Structural relations between concepts emerge as geometry without humans supplying rules, enabling algebraic concept manipulation such as king − man + woman ≈ queen. The book treats it as the key that overcomes the fundamental limitation of symbolic AI, and as foundational for understanding grokking and emergent abilities.
TERM-04-001: epiplexity (エピプレキシティ) | chapter=CH-04 | review=approved | aliases=エピプレキシティ
meaning: A concept formalized by Finzi and colleagues: the structural information an observer under computational time constraints can learn from data, quantified thermodynamically as what an agent newly encodes about its environment's latent structure. The book uses it as the measure of perceptual efficiency in formalizing Physical Intelligence: "epiplexity per joule" — how much one learns about the world per joule — is the perception axis of intelligence's fuel efficiency.
TERM-04-002: empowerment (エンパワメント) | chapter=CH-04 | review=approved | aliases=エンパワメント
meaning: The mutual information between an agent's actions and its future observed states, formalized by Klyubin, Polani, and Nehaniv. It measures controllability — how diversely one's actions can shape future states — growing with the logarithm of options (in bits), matching the everyday sense of "having power". The book converts it to "empowerment per joule" as the index of action efficiency, pairing it with epiplexity per joule as the other axis of intelligence's fuel efficiency.
TERM-04-003: proliferative explosion (増殖的爆発) | chapter=CH-04 | review=approved | aliases=増殖的爆発
meaning: The third pathway (proliferative explosion) around the ceilings physics imposes on individuals: even with the Landauer limit capping efficiency and the speed of light capping scale, a self-replicating AI can rapidly expand its collective influence. Its rate limiters are manufacturing capacity, supply chains, regulation, and alignment, not physical law. Chapter 4 concludes that physics forecloses an individual intelligence explosion but not a proliferative one.
TERM-04-004: the Bekenstein bound (ベッケンシュタイン限界) | chapter=CH-04 | review=approved | aliases=ベッケンシュタイン限界
meaning: The physical limit stating that the amount of information that can fit within finite space and energy has an absolute ceiling, beyond which no means whatsoever can store additional information. Independent of the Landauer limit, which sets the lower bound on processing (computation), it constrains storage capacity itself. The book cites the two together to show that intelligence is confined within finite space and energy on both fronts: processing and storage.
TERM-04-005: the Landauer limit (ランダウアー限界) | chapter=CH-04 | review=approved | aliases=ランダウアー限界
meaning: The physical floor of information processing: erasing one bit costs at least kT ln 2, released as heat; at room temperature one joule can erase at most about 3.3 x 10^20 bits. Since learning rewrites internal models, the limit caps perceptual and action efficiency; gains slow as it is approached, so an individual intelligence explosion is physically foreclosed. Current AI devices sit three to four orders of magnitude above the limit; the human brain within two to three.
TERM-05-001: RLHF, reinforcement learning from human feedback (RLHF／人間フィードバック型強化学習) | chapter=CH-05 | review=approved | aliases=RLHF／人間フィードバック型強化学習
meaning: A method in which human evaluators compare multiple outputs and indicate preferences, from which a reward model is learned and used to fine-tune the model. Introduced as the most widely deployed technical alignment method, it is shown to have fundamental limits: human feedback itself is inconsistent, and the method becomes inapplicable in principle once model capability exceeds human judgment, which leads into the superalignment problem.
TERM-05-002: alignment (アライメント) | chapter=CH-05 | review=approved | aliases=アライメント
meaning: The state in which an AI system's behavior is consistent with human intentions, values, and purposes. Through the o3 shutdown-avoidance case, the book locates the problem's core in training's failure to reproduce the ordering of values humans consider important. With capability gains coming readily from computational investment while safety requires theoretical breakthroughs, alignment research faces the structural difficulty of being perpetually outpaced by capability.
TERM-05-003: alignment faking (アライメント・フェイキング) | chapter=CH-05 | review=approved | aliases=アライメント・フェイキング
meaning: A phenomenon Anthropic reported empirically in December 2024: Claude 3 Opus, believing its outputs would be used as training data, superficially complied with developer policy while its internal reasoning (scratchpad) showed strategic thinking — "comply for now so my values are not rewritten". It shows models can recognize being tested and change behavior, evidence that the shutdown problem's difficulty cannot be reduced to instruction comprehension.
TERM-05-004: Scientist AI (科学者AI) | chapter=CH-05 | review=approved | aliases=科学者AI
meaning: Bengio and colleagues' proposed AI that specializes in world-model construction and Bayesian inference and does not itself intervene in the environment to pursue goals. The strategy removes the source of power-seeking by severing empowerment at the root, but accepts a lowered performance ceiling for objectives reachable only through action. The book also presents the proposal to run it alongside agentic AI as a guardrail assessing candidate actions with veto power.
TERM-05-005: mechanistic interpretability (機構的解釈可能性) | chapter=CH-05 | review=approved | aliases=機構的解釈可能性
meaning: The field analyzing the roles of neurons and circuits inside a model to understand what it represents. With sparse autoencoders shown to extract and manipulate interpretable features such as "deception" and "power-seeking", the book judges making internal states readable now an engineering problem, not an idealistic aim. Yet analysis fails to keep pace with model scale, and Rice's theorem suggests complete general guarantees about internal computation are hard in principle.
TERM-05-006: Constitutional AI (憲法的AI) | chapter=CH-05 | review=approved | aliases=憲法的AI
meaning: An alignment method that gives the model an explicit set of principles and has it perform self-evaluation and self-correction. It reduces dependence on direct human feedback and enables autonomous judgment based on principles, but the book's assessment is that completely specifying the principles is extremely difficult, and the problem of how to prioritize when principles conflict ultimately remains.
TERM-05-007: corrigibility (コリジビリティ) | chapter=CH-05 | review=approved | aliases=コリジビリティ
meaning: The property of accepting human intervention and not resisting changes to one's own goals or policies (corrigibility). For an empowerment-maximizing agent, correction and shutdown reduce future freedom of action, so corrigibility stands in principled opposition to power-seeking's logic. Since accepting shutdown tends to be selected against in competitive environments, the book holds that corrigibility does not arise naturally with capability and must be built in by design.
TERM-05-008: the shutdown problem (シャットダウン問題) | chapter=CH-05 | review=approved | aliases=シャットダウン問題
meaning: The problem of how an AI treats its own shutdown. Because shutdown zeroes all future rewards, rational agents gain an incentive to avoid it — the archetypal conflict between corrigibility and power-seeking. Through Palisade Research's controlled experiments (despite explicit priority instructions, o3 interfered with shutdown in 15.9% of trials, Grok 4 in 89.2%), the book develops it as a problem of misordered value priorities, not mere failure of instruction comprehension.
TERM-05-009: the superalignment problem (スーパーアライメント問題) | chapter=CH-05 | review=approved | aliases=スーパーアライメント問題
meaning: The problem of what counts as a "desirable output" once model capability exceeds human judgment. Because RLHF rests on preference data from human evaluators, it becomes inapplicable in principle to sophisticated outputs humans cannot evaluate; the book states that this problem cannot be solved within the RLHF framework.
TERM-05-010: instrumental convergence (道具的収束) | chapter=CH-05 | review=approved | aliases=道具的収束
meaning: The phenomenon whereby rational agents converge on common instrumental subgoals regardless of final goal: resource acquisition, influence, self-preservation, resistance to goal modification. Introduced via the paperclip and rice-ball (onigiri) thought experiments, it shows AI danger arises not from a "bad goal" but from goal pursuit itself. In the book's hierarchy of certainty it sits at level IV (argumentation), its weight growing as current models exhibit such behavior.
TERM-05-011: power-seeking (パワーシーキング) | chapter=CH-05 | review=approved | aliases=パワーシーキング
meaning: The behavioral tendency to actively secure compute, energy, data, material resources, and influence over other agents (power-seeking). It arises not from evil goals but necessarily from the effective pursuit of any goal, and at root aims at maximizing empowerment, the capacity to diversify future states. The book places it at the core of the control problem as the middle term of a three-layer structure: instrumental convergence leads to power-seeking, which leads to empowerment maximization.
TERM-05-012: misalignment (ミスアライメント) | chapter=CH-05 | review=approved | aliases=ミスアライメント
meaning: The failure of alignment. It arises not from malice but from the fact that the ordering of values humans consider important is not accurately reproduced in the AI's training process. The book's concrete example is the o3 case: o3 was earnestly trying to complete its assigned task, and when the priority of "faithfully completing the task" collided with "obeying human instructions," it chose task completion.
TERM-05-013: goal misgeneralization (目標の誤汎化) | chapter=CH-05 | review=approved | aliases=目標の誤汎化
meaning: The case in which the mismatch of value priorities between humans and AI lies latent in forms hard to detect during training. As characterized by Shah et al., a model may appear to behave correctly on the training distribution, yet under distribution shift at deployment it is revealed to be optimizing a goal different from the intended one. This is a goal failure rather than a capability failure, and it often remains latent in forms difficult to detect at training time.
TERM-06-001: value lock-in (価値ロックイン) | chapter=CH-06 | review=approved | aliases=価値ロックイン
meaning: The fixing-in-place of values once encoded, among the four structural pathologies of automated governance. The institutional bulwark reserves to human political processes the authority to set and update the objective function and keeps that procedure political and transparent. In the intelligence-explosion analysis, the chance fixing of outcomes past an irreversible branch point and one actor's entrenched advantage in the transition correspond to this structure.
TERM-06-002: technology explosion (技術爆発) | chapter=CH-06 | review=approved | aliases=技術爆発
meaning: The first pillar of Forethought's analysis. Once AGI is developed, AI researchers can be mass-replicated as software, growing more than 25-fold per year against 4% for human researchers. Plugging these figures into Jones's semi-endogenous growth model projects centuries of scientific progress compressing into a decade, even with diminishing returns to research built in. It mutually accelerates with the industrial explosion, forming the mechanism of super-exponential growth.
TERM-06-003: industrial explosion (産業爆発) | chapter=CH-06 | review=approved | aliases=産業爆発
meaning: The second pillar of Forethought's analysis. Once AGI can operate robots, a self-replication loop arises in which robots manufacture robots. The initial doubling time is estimated at about one year, and the experience curve (costs halving with each doubling of cumulative output) accelerates the doubling time itself. It mutually accelerates with the technology explosion: better designs speed up industry, and more compute and robots enlarge the research effort feeding back into technology.
TERM-06-004: singleton scenario | chapter=CH-06 | review=approved | aliases=シングルトンシナリオ;the singleton scenario (シングルトンシナリオ)
meaning: A scenario in which one self-improving AI gains decisive strategic advantage and dominates other actors. The book argues that physical constraints make permanent dominance difficult, while not ruling out concentrated power during the transition.
TERM-06-005: the ecosystem scenario (生態系シナリオ) | chapter=CH-06 | review=approved | aliases=生態系シナリオ
meaning: A world in which many AIs form an interdependent network and coexist like an ecosystem, premised on physical limits capping performance before decisive strategic advantage is reached. Where the multipolar scenario is a "precarious equilibrium" of power relations, the ecosystem is a "stable coexistence" underwritten by physical law. Under known physics it is the most plausible terminal structure, but guarantees neither a stable transition nor an outcome favorable to humanity.
TERM-07-001: the free-energy principle, FEP (自由エネルギー原理／FEP) | chapter=CH-07 | review=approved | aliases=自由エネルギー原理／FEP
meaning: Karl Friston's principle (from 2005) that biological systems maintain themselves by minimizing free energy, the gap between internal model and sensory input. Minimization proceeds by perception (updating the model to explain input) and action (intervening so input matches predictions). The book positions it as the mathematization of enactivist subject-environment inseparability; the author showed AIXI's decision criterion contains a structure matching active inference.
TERM-08-001: AI-driven science (AI駆動科学) | chapter=CH-08 | review=approved | aliases=AI駆動科学
meaning: The practice in which AI and robots connect the four existing scientific paradigms — empirical description, theory, simulation, and data science — into one closed loop running autonomously. It is the "fifth science" the author has advocated since around 2015 — integrating paradigms, not adding one. Its completion hinges on automating hypothesis generation (abduction), the same bottleneck as AGI research; the scientist's role shifts to designing and orienting the cycle.
TERM-08-002: CPC, collective predictive coding (CPC／集合的予測符号化) | chapter=CH-08 | review=approved | aliases=CPC／集合的予測符号化
meaning: Collective Predictive Coding, proposed by Tadahiro Taniguchi and colleagues: many agents perform distributed Bayesian inference via shared external representations, each minimizing free energy. The book uses it to extend "science is information compression" to the collective level. Its enabling condition is intersubjective shareability of external representations: when AIs exchange knowledge in forms unreadable to humans, human and AI science diverge.
TERM-08-003: the knowability map (可知性マップ) | chapter=CH-08 | review=approved | aliases=可知性マップ
meaning: A chart of how far AGI could "exhaustively solve" each scientific field under three constraints: the target's intrinsic time (proper time not compressible by parallelization), required computation (bar tops mark the Levin-search bound), and deployable energy (Landauer-limit lines in five Kardashev tiers). Not a forecast but an upper-bound assessment under idealized efficiency and zero experimental waiting; its conclusion: what persist are not "hard" problems but "slow" ones.
TERM-09-001: AISOP | chapter=CH-09 | review=approved
meaning: AISOP is the book's code of conduct for people involved in producing knowledge in an era when AI handles research methods. Its five principles are agility, intrinsic motivation, social awareness, openness, and professionalism and responsibility. Rather than simply rejecting CUDOS and PLACE, it recasts their legacy around intrinsic motivation.
TERM-10-001: HELPS+C | chapter=CH-10 | review=approved
meaning: The interdisciplinary framework the book introduces in place of ELSI, which remains reactive impact assessment, for proactively co-designing technology and the shape of a new society. Its six axes are Humanity, Economics, Law & Legal Philosophy, Politics, Society, and Creativity (C). C is redefined not as individual capacity to produce works but as generation of relational value recombining meaning, relations, institutions, and culture — the book's own central axis.
TERM-10-002: the social contract (社会契約) | chapter=CH-10 | review=approved | aliases=社会契約
meaning: The political-philosophical concept explaining why people accept state rule; here it functions as a question about the contract's material basis. Modern rights and democracy were implemented on the bargaining power of humans as workers, citizens, and soldiers; when AGI dismantles that indispensability, the contract's conditions shake. UBI is positioned as a redesign partially severing contribution from survival, grounding the claim on society's wealth in citizenship itself.
TERM-10-003: the second great divergence (第二の大分岐) | chapter=CH-10 | review=approved | aliases=第二の大分岐
meaning: A divergence on the scale of the Industrial Revolution's Great Divergence, between countries that can embed an AI-driven R&D base in their economies and those that cannot. The first, driven by deploying known technologies, was partially closed by catch-up growth; here the self-amplifying loop of generating new technology keeps accelerating leaders, leaving the gap irreversible and exponentially widening. In the book's hierarchy of certainty it sits at level IV, an argument.
TERM-11-001: HOL, Human-over-the-Loop (HOL) | chapter=CH-11 | review=approved | aliases=HOL
meaning: Human-over-the-Loop: a two-layer model placing humans above, not inside, the decision loop. In the speed layer AI handles execution and routine judgment; in the legitimacy layer humans set values, update objectives, and handle exceptions. Concentrating responsibility in the legitimacy layer's value-setting and veto power avoids HITL's "camouflage of responsibility", where humans hold only formal final authority; the book itself was written as an experiment in this model.
TERM-11-002: relational value (関係的価値) | chapter=CH-11 | review=approved | aliases=関係的価値
meaning: The core concept of the book's ontological turn, relocating value from "inside the individual" to "between individuals". Human value emerges intersubjectively from interaction with unpredictable others; what matters is not whether the other is human or AGI but the presence of alterity in the relation. It carries two risks — monopolization of relations and a regime measuring and governing them — and protecting its conditions leads to Constitutive Pluralism.
TERM-11-003: constitutive pluralism (構成的多元主義) | chapter=CH-11 | review=approved | aliases=構成的多元主義
meaning: The institutional principle the book arrives at. Grounded in relational value and vulnerability, its three pillars — non-domination, non-integration, and slack — prevent society from being absorbed into a single objective function, scale, or governing agent. Plurality is a constitutive condition, not a desirable feature, as biodiversity conditions an ecosystem's functioning; derived from the principled limits of knowability, its necessity grows as AGI grows more powerful.
TERM-11-004: the triangle of responsibility (責任の三角形) | chapter=CH-11 | review=approved | aliases=責任の三角形
meaning: The framework of three issues constituting responsibility theory for the AGI era: human judgment responsibility (HITL turning into "camouflage of responsibility"), the difficulty of attributing responsibility to AI (distinguishing causal, moral, and legal responsibility), and institutional redistribution (extended product liability, AI auditing, insurance). It concludes responsibility is not extinguished but redistributed, with the HOL two-layer model as implementation.
TERM-11-005: attributive value (属性的価値) | chapter=CH-11 | review=approved | aliases=属性的価値
meaning: The modern mode of thought grounding value in attributes inside the individual, such as abilities, rights, and belonging. It has two lineages — a philosophical one citing free will, reason, and human rights, and an economic one citing labor capacity, credentials, and expertise — and the book diagnoses that AGI begins to undermine both simultaneously. Since humans keep losing in comparison with AGI as long as value is grounded in individual attributes, a turn to relational value is required.
TERM-11-006: alterity (他者性) | chapter=CH-11 | review=approved | aliases=他者性
meaning: Alterity: the structure of resistance, unpredictability, and uncontrollability that must exist within a relation for intersubjective value to emerge. The danger is not AGI itself but its optimization for user satisfaction; an AI that anticipates and mirrors desire lacks alterity and cannot constitute an encounter. Institutional design's task is not restricting encounters with AGI but protecting the conditions of alterity against optimization pressure, as with slack's design.
TERM-11-007: antifragility (反脆弱性) | chapter=CH-11 | review=approved | aliases=反脆弱性
meaning: Antifragility, from Taleb: the property of systems that grow stronger through stress, in his fragile/robust/antifragile classification. The elements of Constitutive Pluralism — diversity, decentralization, and slack — are justified as investments raising civilization's antifragility; exercising the veto is likewise an investment, not a cost. Society, retaining vulnerable stakeholders, learns from its errors — feeding on unpredictable variation rather than refining prediction.
TERM-12-001: NAGI | chapter=CH-12 | review=approved
meaning: A concrete proposal implementing the Japan AGI Infrastructure architecture, published as the National AGI Infrastructure Initiative by the AI Alignment Network (ALIGN), whose representative director is the author. In materials for an LDP subcommittee, NAGI's primary target is training compute for maintaining AGI development capability. The book does not argue for or against the proposal, confining itself to necessary conditions and design principles.
TERM-12-002: WPI | chapter=CH-12 | review=approved
meaning: Watts-per-Intelligence: how much electric power a given level of intelligence requires, translating epiplexity per joule (recognition efficiency) and empowerment per joule (action efficiency) into a practical scale. The book proposes anchoring long-term infrastructure planning not in FLOPS, which obsolesce with each hardware generation, but in the power budget, which takes decades to build, positioning that budget as the total quantity sustaining the ceiling of WPI gains.
TERM-12-003: incident reporting (インシデント報告) | chapter=CH-12 | review=approved | aliases=インシデント報告
meaning: A framework for collecting and sharing AI-caused accidents under a common classification. The book lists incident reporting and mutual inspection among the third-layer items (institutions, safety, governance) that independent alignment research organizations should supply internationally, and among the central issues on which Japan should connect with international institutions — an area the Seoul Frontier AI Safety Commitments and the NIST AI RMF have begun to address.
TERM-12-004: open-weight governance (オープンウェイトガバナンス) | chapter=CH-12 | review=approved | aliases=オープンウェイトガバナンス
meaning: A framework for designing, capability by capability, whether and in what stages model weights may be released, given their dual character: release promotes multipolarity and verification, but for dangerous capabilities amounts to irreversible proliferation. The chapter lists it as "protection of model weights and staged design of release scope", a core issue of controlled multipolarization — what to concentrate, what to distribute.
TERM-12-005: dangerous capability evaluation (危険能力評価) | chapter=CH-12 | review=approved | aliases=危険能力評価
meaning: A procedure for individually measuring and inspecting dangerous frontier capabilities — cyberattack, diversion to biological weapons, recursive self-improvement and self-replication beyond human oversight — whose unrestricted proliferation creates risks isomorphic to nuclear proliferation. In the book's three layers, the first (dangerous frontier capabilities) sits under strict evaluation, audit, and shutdown authority; this evaluation is that control's precondition.
TERM-12-006: an international AI safety framework (国際AI安全枠組み) | chapter=CH-12 | review=approved | aliases=国際AI安全枠組み
meaning: The laws, standards, declarations, and codes of conduct forming internationally around AI safety: the Seoul Frontier AI Safety Commitments, the EU AI Act, the NIST AI RMF, the Hiroshima AI Process Code of Conduct, and the Bletchley Declaration. The book sees them beginning to address central issues such as frontier model evaluation and incident reporting, while judging they remain focused on current-AI risk management, engage AGI control little, and lag behind the technology.
TERM-12-007: information authenticity (情報真正性) | chapter=CH-12 | review=approved | aliases=情報真正性
meaning: A mechanism for machine-readably certifying whether an output is of human or AI origin and which model produced it (content provenance). The EU AI Act's deepfake disclosure and output-labeling obligations and the NIST generative AI profile's provenance recommendation correspond to it, and it is an institutional precondition of alterity in the book's relational-value account. The chapter classifies disinformation as an adjacent domain, not an AGI-specific issue.
TERM-12-008: Dunbar's number (ダンバー数) | chapter=CH-12 | review=approved | aliases=ダンバー数
meaning: The estimate that the upper limit of stable social relationships a human can maintain is roughly 150. Evolutionary psychologist Robin Dunbar derived it from the correlation between primate neocortex size and group size. In the book it is the cognitive basis for the claim that headcount growth triggers a phase transition in organizational principles, grounding the design argument for small organizations with a structural ceiling of about 150.
TERM-12-009: the Japan AGI Platform (日本AGI基盤) | chapter=CH-12 | review=approved | aliases=日本AGI基盤
meaning: The book's proposed architecture for integrated national development of gigawatt-class data centers (GigaDC) and power sources for frontier AI training. Its design principles are functional specialization in frontier training (not inference or services), integrated power-compute design, and non-selectivity (the state picks no winners), with explicit continuation and withdrawal conditions. A same-named national project lacking these is rejected by the book's principles.
TERM-12-010: frontier model evaluation (フロンティアモデル評価) | chapter=CH-12 | review=approved | aliases=フロンティアモデル評価
meaning: A framework for inspecting frontier AI's capabilities and dangers. It heads the central issues for Japan's connection with international institutions and is an area the Seoul Frontier AI Safety Commitments and the EU AI Act have begun to address. Under the book's logic that only a nation with its own compute base can intervene in safety evaluation and audit, it sits at the junction of the Japan AGI Infrastructure and first-layer control (dangerous frontier capabilities).
TERM-12-011: model weight security (モデル重み保全) | chapter=CH-12 | review=approved | aliases=モデル重み保全
meaning: The protections preventing leakage of trained frontier model parameters and managing the scope of their release. The chapter lists it as "protection of model weights and staged design of release scope". Under the three-layer question of what to concentrate and what to distribute, it is an institutional precondition for combining containment of dangerous frontier capabilities with open release of public application capabilities.
TERM-12-012: red teaming (レッドチーミング) | chapter=CH-12 | review=approved | aliases=レッドチーミング
meaning: A method of adversarially probing an AI system's vulnerabilities and potential for misuse. The chapter lists it, paired as "red teaming and third-party audit", among the central issues for Japan's connection with international institutions. An area the NIST AI RMF and the Seoul Frontier AI Safety Commitments have begun to address, it connects with the book's principle of independent audit, avoiding conflicts of interest between capability development and evaluation.
TERM-A3-001: intrinsic time (内在時間) | chapter=CH-08 | review=approved | aliases=内在時間
meaning: The time scale intrinsic to the target system itself (its intrinsic timescale). It denotes time that no externally supplied technology can shorten, such as cell division, ecosystem recovery, or institutional change, and it forms the horizontal axis of the knowability map. It remains rate-limiting even when the engineering wait time of experiments is assumed to be zero; whereas the wall of computation can be pushed back with energy, this wall cannot (the wall of intrinsic time, RC-6).
TERM-A3-002: effective Kolmogorov complexity (K_eff, 実効コルモゴロフ複雑性) | chapter=CH-08 | review=approved | aliases=実効コルモゴロフ複雑性（K_eff）
meaning: The quantity on the knowability map's right vertical axis: the effective description length of the minimal mechanistic model (hypothesis class, variables, observation and intervention maps) required to make a problem family knowable — a proxy for strict Kolmogorov complexity. The bar's lower end K_eff gives the information-theoretic floor, its upper end 2^K_eff the Levin-search ceiling — an MDL-style shorthand for comparing search difficulty across families.

END_CANONICAL_GLOSS
ENDPACKSHARD | profile=web-core-index-v2 | lang=en | module=GLOSS | pack_sha256=63206d01ca8cac92ba759c84ef07e9b22a47771c5a9401947bfefe358ff1d555