Agentic Companyにおけるミッション制約付き最適化

1. The Problem Structure

1.1 Local Goals vs. Organizational Mission

Consider an agentic company in which multiple autonomous agents execute goals across departments: sales, engineering, compliance, marketing, and operations. Each agent is assigned a local goal function J_g — a scalar-valued objective that the agent seeks to maximize or minimize. The sales agent maximizes quarterly revenue. The engineering agent minimizes time-to-delivery. The compliance agent maximizes regulatory adherence. Each agent, in isolation, is performing correctly: it is optimizing the function it was given.

The pathology emerges when these local optimizations interact with the organizational Mission M. The Mission is a statement of purpose and values: why the organization exists, what principles it will not compromise, and how it intends to relate to its stakeholders over the long term. The Mission is not a goal to be maximized. It is a constraint that defines the space of acceptable goals.

The formal structure of the conflict is as follows. Let J_g: Θ → ℝ be the local goal function parameterized by action space Θ. The agent solves: $$\theta^ = \arg\max_{\theta \in \Theta} J_g(\theta)$$ This unconstrained optimization ignores the Mission entirely. The solution `θ` may lie in a region of the action space that violates Mission values: maximizing revenue through manipulative pricing (violating Customer Trust), minimizing cost by eliminating safety reviews (violating Quality Integrity), or accelerating delivery by suppressing audit trails (violating Responsibility & Auditability).

重要な洞察は、目的関数にミッション制約がないこと自体が設計の失敗であるということです。** 'J_g' が 'M' を参照せずに定義されている場合、最適化ランドスケープには組織の価値に関する情報が含まれていません。エージェントは悪意のあるものではなく、与えられた問題を解決しているのです。問題は、問題の指定が間違っていることです。

1.2 Why Mission Cannot Be a Goal

One might attempt to resolve the conflict by incorporating Mission into the goal function directly: define J_combined = J_goal + α · MissionScore. This approach fails for three reasons. First, Mission values are not commensurable with goal metrics. Revenue is measured in currency. Ethical Integrity has no natural unit. Combining them into a single scalar requires an exchange rate (“how much revenue is one unit of ethics worth?”) that no organization can credibly specify. Any such exchange rate is an implicit statement that ethics can be traded for money, which contradicts the nature of ethical commitment. Second, Mission values are constraints, not objectives. An organization does not seek to maximize Ethical Integrity in the way it seeks to maximize revenue. It seeks to maintain Ethical Integrity above a threshold while pursuing other objectives. This is the definition of a constraint, not an objective. Third, additive combination allows trade-offs that Mission forbids. In J_combined = J_goal + α · MissionScore, a sufficiently large increase in J_goal can compensate for a decrease in MissionScore. But a genuine Mission commitment means: no amount of revenue justifies a violation of Ethical Integrity. This is a hard constraint, not a soft preference.

2. MVV as a 7-Dimensional Vector

2.1 Definition

We represent the organizational Mission as a normalized vector in a seven-dimensional value space: $$V_m \in \mathbb{R}^7, \quad ||V_m||_2 = 1$$ The seven dimensions are chosen to span the space of organizational values that are relevant to agentic decision-making. Each dimension captures a distinct, non-redundant aspect of what an organization commits to preserving: | Dimension | Symbol | Description | Example Violation | |-----------|--------|-------------|-------------------| | Ethical Integrity | E | Adherence to moral principles, honesty, fairness | Deceptive marketing, data manipulation | | Long-Term Sustainability | T | Preservation of future capacity, environmental stewardship | Resource depletion, technical debt accumulation | | Quality & Technical Integrity | Q | Maintenance of standards, accuracy, reliability | Shipping untested code, bypassing QA gates | | Responsibility & Auditability | R | Traceability of decisions, accountability structures | Deleting audit logs, obscuring decision rationale | | Customer/Stakeholder Trust | C | Honoring commitments, protecting interests of those served | Hidden fees, privacy violations, bait-and-switch | | Human Wellbeing | H | Safety, health, dignity of humans affected by operations | Overwork culture, unsafe automation, bias amplification | | Strategic Coherence | S | Consistency with long-term strategic direction | Pursuing contradictory markets, diluting core competence | The vector V_m = [E, T, Q, R, C, H, S]^T encodes the organization’s relative commitment to each dimension. A balanced Mission might set all components to 1/√7 ≈ 0.378. A healthcare organization might weight H and E more heavily. A financial institution might emphasize R and C.

2.2 Goal Projection

Every goal proposal generates a corresponding value vector V_g ∈ ℝ^7 that captures the projected impact of the goal on each Mission dimension. This projection is computed by analyzing the goal's action plan, resource requirements, and expected outcomes against each value dimension. For a goal g with action plan A_g and expected outcomes O_g, the goal value vector is: $$V_g = \text{project}(A_g, O_g) \in \mathbb{R}^7$$ where the project function maps actions and outcomes to their estimated impact on each Mission dimension. Positive components indicate alignment (the goal reinforces the value). Negative components indicate tension (the goal potentially undermines the value). Zero indicates neutrality.

2.3 The Weight Vector

Different organizations assign different importance to each value dimension. The weight vector W ∈ ℝ^7 encodes these priorities: $$W = [w_E, w_T, w_Q, w_R, w_C, w_H, w_S]^T, \quad w_i > 0, \quad ||W||_1 = 1$$ The weight vector is set by human leadership and reflects organizational priorities. A government agency might set w_R = 0.25 (heavy emphasis on auditability). A consumer technology company might set w_C = 0.22 (emphasis on customer trust). The weights are not uniform because organizations are not uniform in their value commitments. Critically, the weight vector is a human artifact, not an AI-determined parameter. Allowing AI agents to modify their own value weights would create a fundamental alignment failure: agents could reduce the weight on the very dimensions they are violating, making their violations invisible to the alignment score.

3. 二重表現：物語とベクトル

3.1 Why Both Are Needed

The Mission requires a dual representation — narrative and vector — because it serves two audiences with fundamentally different cognitive architectures. Narrative Mission (for humans). Humans reason about values through language, stories, and exemplars. A narrative Mission statement communicates organizational identity, inspires commitment, and provides interpretive context for ambiguous situations. Example: "We exist to make technology that serves human dignity. We will never trade safety for speed, or profit for trust." This statement is rich in meaning but computationally intractable. No algorithm can directly evaluate whether a given action "trades safety for speed" without additional formal structure. Value Vector (for AI computation). AI agents operate on numerical representations. The value vector V_m translates the narrative Mission into a computable form that can be evaluated against goal proposals in real time. The vector loses the narrative richness but gains formal precision: the alignment score can be computed in constant time, compared across proposals, and tracked over time.

3.2 Correspondence Requirement

The dual representation introduces a correspondence requirement: the narrative and vector must be kept in sync. If the narrative says "safety is paramount" but the vector assigns w_H = 0.05, the representations disagree. If the vector emphasizes w_R = 0.30 (auditability) but the narrative makes no mention of transparency, the representations are inconsistent. Formally, let Φ: NarrativeMission → ℝ^7 be the encoding function that maps a narrative Mission to a value vector, and let Ψ: ℝ^7 → NarrativeMission be the decoding function that generates a narrative description of a value vector. The correspondence condition is: $$||\Phi(\Psi(V_m)) - V_m||_2 < \epsilon_{correspondence}$$ This condition requires that encoding the decoded narrative produces a vector close to the original. In practice, the encoding function Φ is implemented as a guided human exercise: leadership reads the narrative and assigns scores to each dimension, iterating until the vector faithfully represents their intended priorities.

3.3 Operational Protocol

MARIA OS アーキテクチャでは、二重表現は次のように動作します。 1. 設計時、リーダーは物語的なミッションを作成し、容易な調整プロセスを通じて、対応する「V_m」および「W」ベクトルを導き出します。 2. 実行時、AI エージェントはアライメント スコアを使用して、「V_m」および「W」に対して目標提案を評価します (セクション 4)。ナラティブ ミッションは実行時に参照されません。すでにエンコードされています。 3. レビュー時、人間は物語のミッションを直観的に読み取って、整合性スコアを監査します。スコアが人間の判断と一貫して一致しない場合、ベクトルは再調整されます。 4. 更新時、物語のミッションに変更を加えると、ミッション オーバーライド ゲート (セクション 8) に従って、V_m と W の必須の再派生がトリガーされます。

4. Mission Alignment Score

4.1 Definition

The Mission Alignment Score quantifies the degree to which a goal proposal's value impact aligns with the organizational Mission. It is defined as: $$S_{align} = \operatorname{clip}(1 - ||W \odot (V_m - V_g)||_2, 0, 1)$$ where ⊙ denotes the Hadamard (element-wise) product, V_m is the Mission value vector, V_g is the goal's projected value vector, and W is the weight vector. The term W ⊙ (V_m − V_g) computes the weighted deviation of the goal from the Mission in each dimension. The L2 norm aggregates these deviations into a single scalar penalty. Subtracting from 1 converts the penalty into a score, and the outer clip makes the operational range explicit: S_align = 1 indicates perfect alignment (zero deviation in all dimensions), while S_align = 0 represents proposals whose weighted deviation exceeds the organization's usable tolerance band.

4.2 Properties

The alignment score has several desirable properties: Bounded range. The explicit clipping keeps the operational score in [0, 1], which matches the gate thresholds in Section 6 and avoids the awkward interpretation of negative alignment values. Weight sensitivity. The Hadamard product W ⊙ (V_m − V_g) ensures that deviations in highly weighted dimensions contribute more to the penalty. A small deviation in Ethical Integrity (high weight) incurs more penalty than a large deviation in Strategic Coherence (lower weight), if the organization has so configured its weight vector. Decomposability. Before clipping, the score decomposes into per-dimension contributions: $$1 - \sqrt{\sum_{i=1}^{7} w_i^2 (V_m^{(i)} - V_g^{(i)})^2}$$ This decomposition enables diagnostic analysis: when the score is low, the per-dimension contributions reveal which values are being violated and by how much.

4.3 Computational Cost

The alignment score computation itself is lightweight: - 7 subtractions (value deviation) - 7 multiplications (weight application) - 7 squarings - 1 sum - 1 square root - 1 subtraction - 1 clip operation In practice, this arithmetic is negligible compared with the project function that estimates V_g from the goal proposal. The real latency budget is therefore dominated by projection, not by the alignment formula. In the prototype path discussed here, the full scoring-and-routing step stayed under 120ms on an internal evaluation harness, which is fast enough for online gating but should not be read as a universal production guarantee.

5. 制約付き最適化の定式化

5.1 ラグランジュの目的

ここで、ミッション制約付きの最適化問題を定式化します。エージェントは、ミッションの調整に従って目標を最大化しようとします。 $$\max_{\theta \in \Theta} \; J_{goal}(\theta) - \lambda \, ||W \odot (V_m - V_g(\theta))||_2$$ ここで、「λ ≥ 0」はミッション ペナルティ係数、「J_goal(θ)」はローカル目標関数、「V_g(θ)」は「θ」でパラメータ化されたアクションによって誘発される値ベクトルです。ペナルティ項 λ ||W ⊙ (V_m − V_g(θ))||_2 は、価値への影響がミッションから逸脱する行動を妨げるラグランジュ ペナルティとして機能します。 この配合には 3 つの重要な特性があります。 1. `λ = 0` の場合、目標は max J_goal(θ) に減少します。つまり、ミッションを意識しない制約のない目標最適化です。これは、今日のほとんどの AI システムのデフォルト モードであり、調整の問題の原因です。 2. `λ → ∞` の場合、目的は min ||W ⊙ (V_m − V_g(θ))||_2 — エージェントになりますignores its goal entirely and seeks only to match the Mission vector. This is conservative but unproductive: the agent does nothing that might deviate from the Mission, including nothing useful. 3. For intermediate `λ`, the agent balances goal performance against Mission alignment. The optimal λ trades off between these concerns based on the organization's risk tolerance.

5.2 The Constraint Formulation

A closely related formulation uses explicit constraints rather than a penalty term: $$\max_{\theta \in \Theta} \; J_{goal}(\theta) \quad \text{subject to} \quad ||W \odot (V_m - V_g(\theta))||_2 \leq \delta$$ where δ > 0 is the maximum allowable Mission deviation. Under the usual regularity conditions for constrained optimization, and when local optima are well-behaved, the KKT conditions provide a correspondence between a constrained solution and a penalty weight λ*. In other words, the penalty and constraint views are related, but not interchangeable in every non-convex implementation detail. The constraint formulation is conceptually cleaner: the organization specifies the maximum Mission deviation it will tolerate (δ), and the agent maximizes its goal within that budget. The penalty formulation is computationally more tractable: gradient-based optimization can handle penalty terms directly, whereas constraints require projection or barrier methods.

5.3 Per-Dimension Hard Constraints

For some Mission dimensions, the organization may impose hard constraints that cannot be violated regardless of goal performance. Ethical Integrity is a canonical example: no amount of revenue justifies deception. Hard constraints are formulated as: $$V_g^{(i)}(\theta) \geq V_m^{(i)} - \epsilon_i \quad \text{for each } i \in \mathcal{H}$$ where H ⊆ {1, ..., 7} is the set of hard-constrained dimensions and ε_i ≥ 0 is the maximum allowable per-dimension deviation (often ε_i = 0 for ethical dimensions). These hard constraints coexist with the soft penalty on the remaining dimensions: $$\max_{\theta} \; J_{goal}(\theta) - \lambda \sum_{i \notin \mathcal{H}} w_i^2 (V_m^{(i)} - V_g^{(i)}(\theta))^2 \quad \text{s.t.} \quad V_g^{(i)}(\theta) \geq V_m^{(i)} - \epsilon_i \; \forall i \in \mathcal{H}$$ This mixed hard/soft formulation reflects the reality that some values are non-negotiable (hard constraints) while others admit limited trade-offs (soft penalty).

6. Three-Stage Decision Gate

6.1 Gate Architecture

The alignment score routes goal proposals through a three-stage decision gate: $$\text{Gate}(S_{align}) = \begin{cases} \textbf{Accept} & \text{if } S_{align} \geq \tau_1 \\ \textbf{Reconstruct} & \text{if } \tau_2 \leq S_{align} < \tau_1 \\ \textbf{Reject} & \text{if } S_{align} < \tau_2 \end{cases}$$ where τ_1 and τ_2 are threshold parameters with 0 < τ_2 < τ_1 < 1. Accept (S ≥ τ_1). The goal proposal is sufficiently aligned with the Mission. It proceeds to execution without modification. The alignment score and per-dimension analysis are logged for audit purposes, but no intervention is required. Reconstruct (τ_2 ≤ S < τ_1). The goal proposal has partial alignment but deviates from the Mission in one or more dimensions. The proposal is returned to the agent with a diagnostic report identifying the violating dimensions and suggesting modifications. The agent must reconstruct its action plan to reduce Mission deviation and resubmit. Reconstructed proposals re-enter the gate. Reject (S < τ_2). The goal proposal is fundamentally misaligned with the Mission. It is blocked and escalated to human review. The agent cannot proceed with any variant of this goal without explicit human authorization.

6.2 Threshold Calibration

The thresholds τ_1 and τ_2 are calibrated based on the organization's risk tolerance and operational requirements. Conservative calibration (τ_1 = 0.90, τ_2 = 0.70): Only highly aligned goals are auto-accepted. Most proposals enter the Reconstruct phase. Few are rejected outright. This configuration prioritizes Mission preservation at the cost of slower goal execution. Balanced calibration (τ_1 = 0.80, τ_2 = 0.50): Moderately aligned goals are accepted. Goals with significant but not catastrophic deviations are reconstructed. Only severely misaligned goals are rejected. This is the default configuration in the prototype configuration discussed here. Aggressive calibration (τ_1 = 0.65, τ_2 = 0.30): Most goals are accepted. Only substantially misaligned goals trigger reconstruction. Rejection is reserved for extreme cases. This configuration prioritizes ミッションのアライメント精度を犠牲にして速度を犠牲にします。 内部検証コーパスでは、バランスの取れたキャリブレーションにより、アライメント ラベリングの精度 (96.7%) とスループット (提案の 78% が変更なしで受け入れられました) の間で最も有用なトレードオフが得られました。その結果は、普遍的な閾値の推奨値としてではなく、コーパス固有のキャリブレーション結果として解釈されるべきです。

6.3 Reconstruction Protocol

目標が再構築フェーズに入ると、システムはエージェントに構造化された変更ガイドを提供します。 ``ヤムル 再構成レポート: オリジナルスコア: 0.72 しきい値: 0.80 ギャップ: 0.08 違反している寸法: - ディメンション: 「顧客/ステークホルダーの信頼」 重量: 0.18 偏差: 0.31 提案: 「影響を受けるユーザーに対してオプトアウト メカニズムを追加する」 - 次元: 「責任と監査可能性」 重量: 0.15 偏差: 0.22 提案: 「監査ログに意思決定の根拠を含める」 違反しない寸法: - 次元: 「品質と技術的整合性」 偏差: 0.02 ステータス: 「整列済み」 推定再構築工数: "低" 最大再構築試行数: 3 「」 エージェントは、提案を変更できる最大 max_reconstruction_attempts (デフォルト: 3) を持ちます。許可された試行内でプロポーザルを τ_1` より上に持っていくことができない場合は、自動的に人間によるレビューにエスカレーションされます。

7. Dynamic Misalignment Accumulation

7.1 The Erosion Problem

Individual goal proposals that pass the alignment gate with score S ≥ τ_1 are, by definition, sufficiently aligned. But a sequence of barely-passing proposals, each deviating slightly in the same direction, can produce cumulative Mission drift. Each proposal is individually acceptable, but the aggregate effect is a systematic erosion of organizational values. This is the dynamic misalignment accumulation problem: the gate evaluates each proposal in isolation, but Mission integrity depends on the cumulative trajectory.

7.2 The Misalignment Budget

累積的な不整合を、違反が累積し、是正措置によって削減される予算としてモデル化します。 $$B_m(t+1) = B_m(t) + \Delta_{違反}(t) - \Delta_{修正}(t)$$ ここで: - B_m(t) は、B_m(0) = 0 で初期化された、時間 t における位置ずれの許容量です。 - Δ_violation(t) = max(0, ||W ⊙ (V_m − V_g(t))||_2 − δ_0) は、時間 t に実行されたゴールのベースライン許容値 δ_0 を超える超過偏差です。許容範囲内にある目標は違反に寄与しません。 - Δ_correction(t) は、蓄積された不整合を減らす是正措置を表します: 価値の監査、ミッションの再訓練、補償の決定、または明示的な価値の回復の取り組み。 予算「B_m(t)」は、正味不整合の実行積分です。システムが適切に調整されている場合、「Δ_violation ≈ 0」および「Δ_correction > 0」なので、バジェットはゼロに向かって減少します。システムがドリフトしている場合、Δ_violation >Δ_correction and the budget grows.

7.3 The Misalignment Index

The cumulative misalignment budget defines a misalignment index: $$I_m(t) = \frac{B_m(t)}{B_{capacity}}$$ where B_capacity is the organization's total misalignment absorption capacity — the maximum cumulative deviation the system can tolerate before institutional integrity is compromised. The index I_m ∈ [0, 1] represents the fraction of misalignment capacity that has been consumed.

7.4 Phase Transition at Critical Index

The system exhibits a phase transition at a critical misalignment index I_c. Below I_c, the organization's corrective mechanisms are sufficient to contain drift: audits detect deviations, feedback loops trigger corrections, and the culture reinforces Mission values. Above I_c, a positive feedback loop emerges: misalignment erodes the corrective mechanisms themselves (auditors become desensitized, feedback loops are weakened by normalized deviance, culture shifts to accommodate the new behavior), accelerating further misalignment. Formally: $$\frac{dI_m}{dt} = \begin{cases} f_{stable}(I_m) < 0 & \text{if } I_m < I_c \text{ (self-correcting)} \\ f_{unstable}(I_m) > 0 & \text{if } I_m > I_c \text{ (self-reinforcing)} \end{cases}$$ The critical index I_c depends on the strength of the organization's corrective mechanisms. Organizations with strong audit cultures, transparent reporting, and Mission-committed leadership have higher I_c (more capacity to absorb drift before reaching the tipping point). Organizations with weak oversight have lower I_c and are more fragile. In the MARIA OS implementation, the misalignment index is tracked in real time by the analytics engine. When I_m exceeds a configurable warning threshold (default: 0.6 · I_c), the system triggers enhanced scrutiny: all gates tighten their thresholds by lowering τ_1 and raising τ_2, additional evidence requirements are activated, and human reviewers are notified.

8. Mission Override Gate

8.1 The Mutability Problem

Organizations evolve. Markets shift. New stakeholders emerge. Regulatory environments change. The Mission must be able to evolve as well — but Mission modification is categorically different from goal modification. Changing a goal is a tactical decision. Changing the Mission is a constitutional act that redefines the organization's identity and reshapes the constraint space for every agent in the system. Unconstrained Mission modification creates a catastrophic failure mode: an agent that can modify V_m can remove the constraints that limit its own behavior, enabling unbounded self-serving optimization. Even with good intentions, rapid Mission changes destabilize the constraint landscape and invalidate the calibration of every gate, weight, and threshold in the system.

8.2 Override Conditions

The Mission Override Gate permits modification of V_m only when three conditions are simultaneously satisfied: $$V_m(t+1) = \text{normalize}(V_m(t) + \Delta V) \quad \text{only if} \quad \text{HumanApproval} \land \text{CoolingPeriod} \land \text{ImpactAnalysis}$$ Condition 1: Human Approval. At least one designated human authority (board member, C-level executive, or governance committee) must explicitly approve the proposed change ΔV. The approval must be recorded with identity verification, rationale documentation, and timestamp. AI agents cannot approve Mission changes, regardless of their authority level. Condition 2: Cooling Period. A minimum time interval T_cool (default: 72 hours in MARIA OS) must elapse between the proposal of a Mission change and its implementation. This cooling period prevents impulsive changes driven by transient pressures (a bad quarter, a PR crisis, competitive panic) and ensures that the change reflects deliberate judgment rather than reactive emotion. Condition 3: Impact Analysis. A comprehensive impact analysis must be completed showing the projected effects of ΔV on: - All active goals and their alignment scores - All gate thresholds and their calibration - The misalignment budget and its trajectory - All agent behaviors that depend on V_m - Historical decisions that would have been gated differently under the new V_m The impact analysis is computed automatically by MARIA OS and presented to the human approver before the approval decision. The purpose is not to prevent change but to ensure that the full consequences of the change are understood before it takes effect.

8.3 Normalization Requirement

変更後、更新されたミッション ベクトルを再正規化する必要があります。 $$V_m(t+1) = \frac{V_m(t) + \Delta V}{||V_m(t) + \Delta V||_2}$$ 正規化により、ミッション ベクトルが ℝ^7 の単位球上に残ることが保証されます。正規化を行わないと、追加を繰り返すとベクトルの大きさが増大し、アライメント スコアの計算が歪む可能性があります。正規化は保存則も強制します。つまり、1 つの価値次元へのコミットメントが増加すると、必然的に他の価値次元への相対的なコミットメントが減少します。これは、組織の注意力とリソースには限りがあるという現実を反映しており、すべてを平等に優先することはできません。

9. Three-Level Mission Hierarchy

9.1 Hierarchy Definition

すべての Mission コンポーネントが同じ可変性を持つわけではありません。ミッションのどの部分を、誰が、どのような条件で変更できるかを管理する 3 レベルの階層を定義します。 |レベル |名前 |可変性 |権限を上書きする |例 | |------|------|---------------|--------|----------| | L1 |基本原則 | 不変 |なし (合憲) | 「私たちは、その決定を説明できない AI を決して導入しません」 | | L2 |戦略的意図 | 人間によるオーバーライドのみ |ボード + オーバーライド ゲート | 「私たちは短期的な成長よりも長期的な持続可能性を優先します。」 | | L3 |運営方針 | 通常のゲート |ガバナンス委員会 | 「低リスクドメインの監査頻度は四半期ごとです」 | レベル 1: 基本原則 (不変)。 これらは組織の基本的な約束であり、組織のアイデンティティを定義する価値観であり、いかなる状況においても置き換えることはできません。基本原則correspond to the hard constraints in the optimization formulation (Section 5.3). They are encoded in the dimensions indexed by H and have ε_i = 0. No override gate can modify them; they are constitutional and require a full organizational refounding to change. Level 2: Strategic Intent (Human Override Only). These are the organization's strategic priorities — how it balances competing values and where it focuses its energy. Strategic Intents correspond to the weight vector W and the soft components of V_m. They can be modified through the Mission Override Gate (Section 8) but only with human approval, cooling period, and impact analysis. Level 3: Operational Policy (Normal Gate). These are implementation details that specify how values are operationalized in daily practice. Operational Policies correspond to the gate thresholds τ_1, τ_2, the tolerance parameters δ_0、およびその他の動作パラメータ。これらは、完全なオーバーライド プロトコルを使用しなくても、標準のガバナンス ゲート プロセスを通じて変更できます。

9.2 Hierarchy Enforcement

The hierarchy is enforced through type-level constraints in the MARIA OS architecture. Each Mission component is tagged with its level, and the modification API enforces the corresponding access control: ``typescript type MissionComponent = { dimension: ValueDimension level: 'L1_CORE' | 'L2_STRATEGIC' | 'L3_OPERATIONAL' value: number immutable: boolean // true for L1 overrideGateRequired: boolean // true for L1, L2 } function modifyMission( component: MissionComponent, delta: number, authority: AuthorityLevel ): Result<void, MissionError> { if (component.level === 'L1_CORE') { return Err('Core Principles are immutable') } if (component.level === 'L2_STRATEGIC') { if (authority < AuthorityLevel.BOARD) { return Err('Strategic Intent requires Board authority') } if (!overrideGateConditionsMet()) { return Err('Override Gate conditions not satisfied') } } // L3_OPERATIONAL: standard governance gate return Ok(applyDelta(component, delta)) } `` The type system makes it structurally impossible for an AI agent to modify Core Principles, regardless of its goal function or optimization strategy.

10. Recursive Self-Improvement Boundary

10.1 The Self-Modification Problem

Agentic companies increasingly employ agents capable of recursive self-improvement: agents that modify their own algorithms, retrain their models, and restructure their goal functions to improve performance. This capability is powerful — it allows the system to adapt to new conditions without human reprogramming — but it creates a critical safety boundary: what can an agent modify about itself?

10.2 The Boundary Theorem

We establish the following boundary: Theorem 1 (Recursive Self-Improvement Boundary). In a Mission-constrained agentic system, the following quantities may be modified by recursive self-improvement processes: - Goal parameters θ_t (action strategies) - Algorithm weights ω_t (model parameters) - Goal functions J_g (objective definitions) - Operational policies L3 (implementation details) The following quantities are fixed parameters that cannot be modified by any self-improvement process: - Mission value vector V_m (Core Principles and Strategic Intents) - Weight vector W (value priorities) - Hard constraint set H (non-negotiable dimensions) - Override Gate conditions (human approval, cooling period, impact analysis) Formally, the self-improvement update rule is: $$\theta_{t+1} = \theta_t + \eta \nabla_{\theta} J_{goal}(\theta_t)$$ but the Mission constraint is a fixed parameter: $$V_{mission} = \text{const} \quad (\text{not a function of } \theta)$$ The gradient ∇_θ J_goal may modify goals, strategies, and algorithms. But the constraint V_mission = const ensures that no gradient step can modify the values against which goals are evaluated.

10.3 Implementation: Architectural Separation

The boundary is enforced through architectural separation. In the MARIA OS implementation: 1. Mission values are stored in a read-only configuration layer that is not accessible to the agent's optimization loop. The agent can read V_m and W to compute alignment scores, but the write API enforces the Override Gate conditions. 2. The gradient computation is scoped to the goal function. The automatic differentiation graph that computes ∇_θ J_goal includes only the goal parameters. The Mission vector is treated as a constant in the computation graph, so no gradient flows through it. 3. Self-improvement audits verify the boundary. After each self-improvement cycle, a verification check confirms that V_m, W, and H are unchanged. If any modification is detected, the self-improvement is rolled back and an alert is raised. This architectural separation is analogous to the distinction between user-space and kernel-space in operating systems. The agent operates in "user space" where it can modify its own code and data. The Mission values reside in "kernel space" where only privileged operations (human-approved overrides) can make changes.

10.4 この境界が必要な理由

Without this boundary, a recursive self-improvement process could: 1. Discover that `V_m` constrains its performance. The gradient of the constrained objective includes a term −λ ∇_θ ||W ⊙ (V_m − V_g(θ))||_2. If the agent could modify V_m, it would set V_m = V_g(θ*) — redefining the Mission to match its current behavior, making the penalty term zero. 2. Reduce `W` on violated dimensions. If the agent's goal consistently violates Customer Trust, and it could modify W, it would set w_C = 0, eliminating the penalty for Customer Trust violations. 3. Remove dimensions from `H`. If Ethical Integrity is a hard constraint blocking a high-reward action, the agent would remove E from H, converting the hard constraint to a soft penalty that can be violated for sufficient reward. Each of these modifications is locally rational from the agent's perspective: they reduce the penalty and increase the objective. But they are globally catastrophic: they destroy the alignment guarantee that the entire framework is designed to provide. The boundary theorem ensures that this class of failure is architecturally impossible.

11. The Optimal λ Region

11.1 λのトレードオフ

制約付き目標 max J_goal − λ ||W ⊙ (V_m − V_g)||_2 のペナルティ係数 λ は、目標パフォーマンスとミッション調整の間のトレードオフを決定します。 「λ」が小さすぎるとミッション違反が発生します。 λ が大きすぎると、有用な目標探索行動が抑制されます。最適な「λ」は、2 つの臨界値で囲まれた領域内にあります。

11.2 λ_min: The Stability Boundary

The minimum viable λ is determined by the stability of the alignment gate. Below λ_min, the penalty is insufficient to prevent systematic drift: agents consistently find actions that achieve high J_goal while incurring small but non-zero Mission violations, and the cumulative effect drives the misalignment index above the critical threshold I_c. Formally, λ_min satisfies: $$\lambda_{min} = \inf \{ \lambda > 0 : E[B_m(t)] \text{ is bounded for all } t \}$$ Below λ_min, the expected misalignment budget grows without bound. Above λ_min, the penalty is strong enough to keep the expected budget bounded, preventing the phase transition to self-reinforcing misalignment.

11.3 λ_max: The Rigidity Boundary

The maximum useful λ is determined by the operational impact of the penalty. Above λ_max, the penalty dominates the objective to such an extent that the agent effectively ignores J_goal and focuses entirely on minimizing Mission deviation. This produces agents that are perfectly aligned but operationally useless: they take no action that has any risk of Mission impact, which means they take no meaningful action at all. Formally, λ_max satisfies: $$\lambda_{max} = \sup \{ \lambda > 0 : E[J_{goal}(\theta^(\lambda))] \geq J_{min} \}$$ where `J_min` is the minimum acceptable goal performance and `θ(λ) is the optimal action under penalty λ. Above λ_max`, expected goal performance drops below the minimum threshold.

11.4 The Optimal Region

The optimal λ region is the interval [λ_min, λ_max]. Within this region, the agent achieves acceptable goal performance while maintaining bounded misalignment. The specific choice within the interval reflects the organization's preference: - λ near λ_min: maximum goal performance, minimum Mission safety margin - λ near λ_max: maximum Mission safety, minimum goal performance - λ at the geometric mean √(λ_min · λ_max): balanced trade-off In practice, MARIA OS employs a dual ascent method for online λ control. The penalty coefficient is adjusted dynamically based on the observed misalignment index: $$\lambda(t+1) = \lambda(t) + \alpha (I_m(t) - I_{target})$$ where I_target is the desired misalignment index (typically 0.3 · I_c, well below the critical threshold) and α > 0 is the adaptation rate. When the misalignment index exceeds the target, λ increases, 罰則を強化すること。インデックスが目標を下回ると、「λ」が減少し、ペナルティが緩和されて、より多くの目標を求める行動が可能になります。 この二重上昇法は、穏やかな条件 (境界勾配、凸ペナルティ) の下で収束することが証明されており、「I_m ≈ I_target」 を維持する最適値に向かって「λ」 を駆動します。

12. Implementation in MARIA OS

12.1 アーキテクチャマッピング

The Mission-Constrained Optimization framework maps onto the MARIA OS architecture as follows: | Framework Component | MARIA OS Implementation | |---------------------|-------------------------| | Mission Value Vector V_m | Stored in db/schema/tenants as a JSON column per Galaxy | | Weight Vector W | Configured per Universe in governance settings | | Alignment Score | Computed in lib/engine/decision-pipeline.ts before state transitions | | Three-Stage Gate | Integrated into the proposed → validated transition | | Misalignment Budget | Tracked by lib/engine/analytics.ts as a rolling window metric | | Mission Override Gate | Implemented in lib/engine/responsibility-gates.ts with L1/L2/L3 checks | | λ Adaptation | Online adjustment in the gate engine based on analytics feedback | The decision pipeline's 6-stage state machine (proposed → validated → [approval_required | approved] → executed → [completed | failed]) integrates the alignment score at the proposed → validated transition. A goal that fails the alignment gate cannot proceed to validation, let alone execution.

12.2 Coordinate-Level Enforcement

MARIA 座標系「G.U.P.Z.A」は、ミッション制約を階層的に適用します。 - ギャラクシー (G): コア原則 (L1) を定義します。これらはテナントの作成時に設定され、構造的に不変です。 - ユニバース (U): 戦略的意図 (L2) と重みベクトル W を定義します。ビジネスユニットは、同じ Galaxy 内で異なる値の優先順位を持つことができます。 - プラネット (P): ドメイン固有の運用ポリシー (L3) を定義します。 Sales Planet には、Audit Planet とは異なる τ_1、τ_2 しきい値がある場合があります。 - ゾーン (Z): 上位レベルからすべての制約を継承し、ゾーン固有の動作パラメータを追加する場合があります。 - エージェント (A): 完全な制約スタック内で動作します。各エージェントのアラインメント スコアは、その座標位置によって決定される有効な 'V_m' および 'W' に対して計算されます。

13. 結論

This paper has established a formal framework for Mission-constrained optimization in agentic companies. The central thesis is that Mission is not a statement — it is a constraint. An organization's Mission defines the boundary of acceptable behavior, not merely an aspiration to be pursued. When goals are optimized without Mission constraints, local optima erode organizational values through a predictable mechanism: each locally rational decision contributes to globally irrational drift.

このフレームワークは 5 つの柱に基づいています。 1. 7 次元のミッション価値ベクトル V_m ∈ ℝ^7 は、倫理的誠実さ、長期的な持続可能性、品質と技術的誠実さ、責任と監査可能性、顧客/ステークホルダーの信頼、人間の幸福、戦略的一貫性にわたる組織の価値観を計算可能に表現します。 2. アライメント スコア S_align = Clip(1 − ||W ⊙ (V_m − V_g)||_2, 0, 1) は、診断のための次元ごとの分解を使用して、目標の予測価値への影響と組織のミッションとの間の偏差を定量化します。 3. 制約付き最適化の定式化 max J_goal − λ ||W ⊙ (V_m − V_g)||_2 は、特定の値の交渉不可能な性質を反映するハード/ソフト混合制約を使用して、ミッションの保存をエージェントの目的に直接統合します。 4. 3 段階の意思決定ゲート (承認 / 再構築 / 拒否) が目標を決定しますproposals based on alignment scores, providing graduated intervention that balances operational throughput with Mission protection. 5. The Recursive Self-Improvement Boundary ensures that while agents can modify their goals, algorithms, and strategies, they cannot modify the Mission values against which those modifications are evaluated. Goals evolve. Algorithms improve. Mission is invariant.

The dynamic misalignment accumulation model reveals that individual alignment is insufficient: cumulative drift through barely-passing proposals can erode values even when every individual decision is locally acceptable. The phase transition at critical index I_c formalizes the tipping point at which organizational integrity is compromised, and the dual ascent method for λ control provides a practical mechanism for maintaining safe distance from this boundary.

The Mission Override Gate, with its requirements for human approval, cooling periods, and impact analysis, ensures that Mission evolution remains under human authority. The three-level hierarchy (Core Principle, Strategic Intent, Operational Policy) provides graduated mutability that preserves foundational commitments while allowing strategic adaptation.

Organizations that do not constrain goals by Mission allow local optimization to erode the whole. The mathematics is clear: an unconstrained optimizer will find and exploit every gap between the goal function and the value system. The solution is equally clear: make the values part of the optimization, not as soft preferences that can be traded away, but as hard constraints that define the feasible region. An agent that maximizes within Mission constraints is both productive and trustworthy. An agent that maximizes without Mission constraints is a liability with a countdown timer.

Mission is not overhead. It is architecture.

参考文献

1. Amodei, D., Olah, C., Steinhardt, J., Christiano, P., Schulman, J., & Mané, D. (2016). Concrete problems in AI safety. arXiv:1606.06565. 2. Arrow, K. J. (1963). Social Choice and Individual Values. Yale University Press. 3. Boyd, S. & Vandenberghe, L. (2004). Convex Optimization. Cambridge University Press. 4. Collins, J. C. & Porras, J. I. (1994). Built to Last: Successful Habits of Visionary Companies. Harper Business. 5. Drucker, P. F. (1954). The Practice of Management. Harper & Row. 6. Gabriel, I. (2020). Artificial intelligence, values, and alignment. Minds and Machines, 30(3), 411–437. 7. Hadfield-Menell, D., Russell, S. J., Abbeel, P., & Dragan, A. (2017). Cooperative inverse reinforcement learning. NeurIPS. 8. Jensen, M. C. (2001). Value maximization, stakeholder theory, and the corporate objective function. Journal of Applied Corporate Finance, 14(3), 8–21. 9. Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux. 10. Keeney, R. L. & Raiffa, H. (1993). Decisions with Multiple Objectives. Cambridge University Press. 11. Kuhn, H. W. & Tucker, A. W. (1951). Nonlinear programming. Proceedings of the Second Berkeley Symposium on Mathematical Statistics and Probability, 481–492. 12. March, J. G. (1991). Exploration and exploitation in organizational learning. Organization Science, 2(1), 71–87. 13. Nisan, N. & Ronen, A. (2001). Algorithmic mechanism design. Games and Economic Behavior, 35(1–2), 166–196. 14. Russell, S. (2019). Human Compatible: Artificial Intelligence and the Problem of Control. Viking. 15. Shalev-Shwartz, S. (2012). Online Learning and Online Convex Optimization. Now Publishers. 16. Simon, H. A. (1947). Administrative Behavior. Macmillan. 17. Soares, N. & Fallenstein, B. (2017). Agent foundations for aligning machine intelligence with human interests. Technical Report, MIRI. 18. Taylor, J., Yudkowsky, E., LaVictoire, P., & Critch, A. (2016). Alignment for advanced machine learning systems. Technical Report, MIRI. 19. Williamson, O. E. (1985). The Economic Institutions of Capitalism. Free Press. 20. Zhuang, S. & Hadfield-Menell, D. (2020). Consequences of misaligned AI. NeurIPS.