KĀ'EO AI Policy Framework

• KĀ‘EO AI Policy Framework is a multi-layered governance model unifying technical, ethical, and cultural controls for responsible AI deployment in Indigenous contexts.
• It incorporates machine-readable Policy Cards, trust evaluation modules, and human oversight to ensure compliance with cultural and educational standards.
• The framework operationalizes community-centered values and formal audit processes, enabling transparent and accountable AI integration.

The KĀ‘EO AI Policy Framework is an operationalized, multi-layered governance construct designed to ensure culturally respectful, fair, and accountable integration of artificial intelligence systems into Indigenous-language assessment contexts. Based on community-centered values and actionable policy structures, it unifies technical, ethical, and institutional controls for responsible deployment, centralizing cultural authority and human expertise throughout the AI lifecycle. Its architecture spans core principles, governance mechanisms, trust modeling, technical enforcement layers, and operational toolsets, synthesizing insights from Policy-as-a-Service, Actionable Principles for AI, and machine-readable standardization methodologies (Morris et al., 2020, Kūkea-Shultz et al., 19 Dec 2025, Mavračić, 28 Oct 2025, Stix, 2021).

1. Foundational Principles and Ethical Constraints

At its core, the KĀ‘EO AI Policy Framework v4.0 specifies seven non-negotiable values for all AI activity within the Kaiapuni Assessment of Educational Outcomes program:

• Stewardship of student data: asset-specific, controlled access.
• Protection of linguistic integrity: the Hawaiian language as sacred trust, requiring cultural authority on all textual analysis.
• Equitable access: system responsiveness to learner and community differences.
• Appropriate pedagogical integration: strict prohibition of decontextualized automation.
• Transparent communication: disclosure of AI’s capabilities and limits.
• Environmental proportionality: computational use constrained by educational benefit.
• Human expertise precedence: explicitly subordinating technological efficiency to expert judgment.

These align with CARE Principles for Indigenous Data Governance and mandate formal project approval protocols, zero-data-retention prompting, written audit records, and mandatory "humans as the loop" control at all AI decision junctures (Kūkea-Shultz et al., 19 Dec 2025).

2. Policy and Governance Architecture

The governance stack is modular, supporting linked technical, organizational, and human control layers. Direct adaptation from Policy-as-a-Service (PaaS) (Morris et al., 2020), the architecture incorporates:

Layer	Role	Mechanisms
Policy Repository & Versioning	Store, version, and geo-tag all machine-readable policies	Query APIs, policy tagging
Interpreter & Enforcement	Resolve, activate, enforce policies at runtime	Semantic reasoner, rule-to-action map
Trust Evaluation Module	Compute, report multidimensional trustworthiness metrics	Log mining, performance scoring
Human–Machine Interface (HMI)	Human oversight, feedback, runtime exceptions, explanations	Dashboards, alerts, decision tracing
Supporting Services	Audit, logging, simulation, update notification	CI/workflow, audit logging

Governance processes include project-level pre-approvals, tiered operator roles (restricted Lab operators, lead psychometricians), session management (ephemeral environments), and cultural review gates with explicit sign-off before any AI-generated content advances to development (Kūkea-Shultz et al., 19 Dec 2025).

3. Trust Model and Socio-Technical Integration

Trust in KĀ‘EO’s framework is evaluated multi-dimensionally, using the PaaS aggregation schema (Morris et al., 2020):

• Technical trust: reliability proxies (e.g., mean time between failures, intervention rates).
• Social trust: human override frequency, sentiment analysis, user surveys.
• Regulatory/legal trust: compliance/audit pass rates, validity certification.

A composite score $T$ is computed:

$$T = w_1 T_{\text{perf}} + w_2 T_{\text{social}} + w_3 T_{\text{reg}}$$

where each $T_i \in [0,1]$ and $w_i$ are context- and policy-dependent weights. These scores dynamically gate operational modes (e.g., enforce stricter operational controls when $T < T_{\min}$ by reducing system autonomy or requiring operator confirmation). Feedback loops connect trust metrics to policy review and system reconfiguration cycles.

4. Operational Policy Encoding and Enforcement

KĀ‘EO’s framework incorporates machine-readable policy artifacts modeled as Policy Cards (Mavračić, 28 Oct 2025), which encode all relevant operational, regulatory, and ethical constraints:

$$\text{PolicyCard} ::= \{\,\mathit{meta},\,\mathit{scope},\,\mathit{applicable\_policies},\,\mathit{controls},\,\mathit{obligations},\,\mathit{monitoring},\,\mathit{kpis\_thresholds},\,\mathit{change\_management},\,\mathit{assurance\_mapping},\,\mathit{references}\,\}$$

Each Policy Card specifies a set of "action rules" as Attribute-Based Access Control (ABAC) tuples:

$$\mathit{ActionRule} ::= \langle\,\mathit{id},\,\mathit{subject},\,\mathit{action},\,\mathit{resource},\,\mathit{condition},\,\mathit{effect}\rangle$$

with $$\mathit{effect} \in \{\mathrm{allow}, \mathrm{deny}, \mathrm{require\_escalation}\}$$.

Formal enforcement semantics:

• $$r.\text{effect} = \mathrm{allow} \land r.\text{condition} \implies \mathrm{permit}(r.\text{subject}, r.\text{action}, r.\text{resource})$$
• $$r.\text{effect} = \mathrm{deny} \land r.\text{condition} \implies \lnot \mathrm{permit}(\ldots)$$
• $$r.\text{effect} = \mathrm{require\_escalation} \land r.\text{condition} \implies \mathrm{escalate}(r)$$

Obligations are encoded as $$\langle \mathit{id}, \mathit{trigger\_condition}, \mathit{actions}, \mathit{evidence\_fields}\rangle$$ and automate evidence recording under specified circumstances.

Validation enforces schema compliance (JSON Schema 2020-12), regex patterns, and critical KPI thresholds. Cards are versioned (semver), subject to CI pipeline checks, and auditable in alignment with NIST AI Risk Management, ISO/IEC 42001, and EU AI Act (Mavračić, 28 Oct 2025).

5. Stakeholder Roles, Human-in-the-Loop, and Accountability

Mandatory human-in-the-loop operation is central: every analytic phase (data ingestion, evidence synthesis, cross-item aggregation, narrative translation, cultural and linguistic review) requires explicit review and sign-off before output is integrated or disseminated (Kūkea-Shultz et al., 19 Dec 2025).

• Lab operators: data access, AI prompt control.
• Psychometricians: verify metric accuracy, analytic integrity.
• Cultural and linguistic authorities: enforce language and context validity, prevent misinterpretation of culturally valued forms.
• Management/oversight: draft, approve, and audit project-level plans, conduct periodic external audits, control formal workflow progression.

No AI output is accepted without dual verification from both psychometric and cultural authorities. Zero-data retention, locked access environments, and audit logs are mandatory at all workflow steps.

6. Technical Workflow and Metrics

KĀ‘EO’s AI-augmented review process is defined by a deterministic, staged workflow:

1. Reference document ingestion (NotebookLM, closed/ephemeral mode).
2. Flagged item analysis by structured prompt (psychometric indices, text paraphrase, flagging for ambiguity, DOK, overload, DIF).
3. Human verification of every claim.
4. Cross-item aggregation for systemic pattern detection.
5. Narrative translation into developer-facing briefs (Claude 3.5 Sonnet, structured-output, training-data opt-out).
6. Final cultural and linguistic gate.

Key psychometric indices formalized in LaTeX include:

• Difficulty: $p_i = \frac{n_{\text{correct},i}}{N}$
• Point-biserial (discrimination): $$r_{pb,i} = \frac{\overline{X}_{\text{correct},i} - \overline{X}_{\text{incorrect},i}}{\sigma_X} \sqrt{p_i q_i}$$
• Differential Item Functioning: $$\Delta_{\text{MH},i} = -2.35 \times \ln\left(\frac{\sum_{\text{group=F}} A_i/B_i}{\sum_{\text{group=M}} C_i/D_i}\right)$$
• DOK Misalignment: $$DOK\_AI = \frac{1}{I} \sum_{i=1}^I |\mathrm{DOK}_{\text{assigned},i} - \mathrm{DOK}_{\text{observed},i}|$$
• Linguistic Ambiguity Index: $$\mathrm{LAI}_i = \frac{\text{near-synonymous alternatives}}{\text{lexical tokens in stem}}$$
• Structural Overload Score: $$\mathrm{SOS}_i = \max(0, \text{MeasuredSteps}_i - \text{ExpectedSteps}_{\mathrm{DOK}_i})$$

Pseudocode explicitly outlines each workflow control, verification, and exception handling mechanism (Kūkea-Shultz et al., 19 Dec 2025).

7. Adaptation and Extensibility via Actionable Policy Modules

The KĀ‘EO Architecture is structured for recursive adaptation and cross-domain extensibility as formalized by the three-pillar "KĀ‘EO synthesis function" (Stix, 2021):

$P = \mathbb{F}(L, S, T)$

where $L$ = landscape data, $S$ = stakeholder inputs, $T$ = toolbox of mechanisms.

These pillars—Preliminary Landscape Assessment, Multi-stakeholder Participation, Implementation/Operationalizability—combine with module-based templates (scope, inputs, outputs, timelines, feedback), supporting scaling (local, national, supra-national) and continuous policy integration with clear accountability metrics.

The Policy Cards model enables modular overlays for global deployments, automated synthesis from model metadata, integration with policy engines (e.g., Open Policy Agent), and the use of cryptographic attestations for privacy-preserving compliance. CI/CD pipelines and audit dashboards operationalize continuous assurance at system and organizational levels (Mavračić, 28 Oct 2025).

---

References:

• Morris et al., Policy-as-a-Service (PaaS) (Morris et al., 2020)
• Policy Cards governance (Mavračić, 28 Oct 2025)
• KĀ‘EO AI Policy Framework v4.0 and AI-augmented item analysis (Kūkea-Shultz et al., 19 Dec 2025)
• Actionable Principles for AI Policy (Stix, 2021)