Zluri Responsible AI Framework

1. Purpose and scope

This framework sets out how Zluri builds, deploys, and operates artificial intelligence responsibly. It is Zluri's AI management system governance document and applies across the company as a durable, general-purpose framework rather than a description of any single feature.

Scope. This framework applies to every Zluri capability that uses artificial intelligence, machine learning, foundation models, or autonomous agents, whether that capability is customer-facing or internal but processing customer data. This includes, for example, catalog and data intelligence, identity-risk scoring, conversational assistance, and agentic automation. These are illustrative uses; the framework applies to any AI capability Zluri operates now or in future, and the current set of AI capabilities is maintained in the AI System Register (Section 6) rather than fixed in this document.

Out of scope. Internal developer and staff productivity tooling that never touches customer data or customer-facing output. That use is governed by the Acceptable Use of AI policy referenced in Section 15.

AI supply posture. Zluri deploys AI capabilities using foundation model providers and cloud infrastructure approved through its AI governance process. The current production model architecture, providers, and deployment regions are maintained in the AI System Register and supporting technical documentation. Zluri does not train foundation models using customer data. Use of Zluri AI is optional and is not required to access Zluri's core platform. Controls in this framework represent Zluri's required target state. Where a control is not yet fully operational, it is tracked to implementation through the AI Governance Committee, and current implementation status is recorded in the AI System Register and Statement of Applicability.

‍

2. Zluri's Responsible AI principles

Zluri commits to five principles. They are the foundation of every control in this framework, and every operating dimension in Section 8 traces back to one or more of them.

2.1 Transparency. People affected by Zluri's AI should be able to understand that AI is being used, what it does, and its limitations. Zluri discloses when a person is interacting with an AI system, explains in plain terms what its AI capabilities do and do not do, surfaces the evidence behind AI-generated outputs where feasible, and publishes its responsible-AI position externally. Transparency is implemented through documented disclosures, user communications, governance records, and appropriate customer documentation that shows up as user-facing disclosures, model documentation, and clear communication to customers and regulators.

2.2 Accountability. Every AI capability has a named human owner who is answerable for it, and no AI system operates without a clear line of human responsibility. Zluri maintains a governance structure with defined roles, decision rights, and escalation paths, so that when an AI system behaves unexpectedly it is always clear who owns the response. Accountability also means Zluri records its decisions: which controls apply to a system, why, and who accepted any residual risk.

2.3 Fairness and bias mitigation. Zluri's AI supports decisions about identity and access, so unfair or biased output can have real consequences for people. Zluri commits to evaluating its AI for biased or inequitable behavior, defining fairness measures appropriate to each use, setting thresholds that a system must meet before release, and keeping a human in the loop for decisions that materially affect a person. Where data protection controls limit direct testing on protected attributes, Zluri applies compensating evaluation and human review rather than treating the limitation as an exemption.

2.4 Privacy and data protection. AI capabilities must uphold the same privacy and security standard as the rest of Zluri's platform, with no exception made because a feature is AI. Zluri minimizes the personal data exposed to AI systems, protects it with tokenization, encryption, and access controls, keeps customer data within defined boundaries, and does not allow customer data to be used to train foundation models. Privacy is enforced through Zluri's existing information security and privacy management systems, which this framework extends rather than duplicates.

2.5 Safety and reliability. Zluri's AI should behave predictably, resist misuse, and fail safely. Zluri constrains what its AI can do, guards against adversarial manipulation, tests for accuracy and robustness before release, monitors behavior in production, and bounds the authority of autonomous agents so that consequential or irreversible actions require human control. A capability that cannot meet its safety and accuracy bar is not shipped.

‍

3. How this framework is grounded

This framework is organized around Zluri's own principles and controls, not around any external standard's clause structure. Recognized standards are the source and reference for why each control was selected, not the skeleton of the document. Specifically, the controls here are informed by and mapped to the following standards:

Standard	What it informs
ISO/IEC 42001	The AI management system
NIST AI Risk Management Framework	Risk functions and generative-AI risk
ISO/IEC 27001 and ISO/IEC 27701	Zluri's existing certifications, which the AI controls inherit rather than re-create
EU AI Act	Regulatory obligations
OWASP Top 10 for Large Language Model Applications	AI-specific threats
CSA AI Controls Matrix	Cloud-AI control coverage

The detailed crosswalk from Zluri's dimensions to these standards is in Annex E, and the formal record of which controls apply to which system, and why, is the Statement of Applicability maintained by the AI Governance Committee. This keeps the framework readable and durable while remaining defensible to an auditor.

‍

4. Governance and accountability

Governance operates in three tiers so that responsibility scales from individual systems to the board.

Tier	Level	Role
Tier 1	AI Governance Committee	Sets policy, owns AI risk at the organizational level, and approves new and higher-risk AI use. Described in Section 5.
Tier 2	Responsible AI review gate	The operational control (Section 11) through which every AI capability passes before release, staffed by engineering leadership with security and privacy input.
Tier 3	Named accountable owner per capability	Each AI capability in the register has exactly one named accountable owner, recorded and dated. Ownership is auditable and cannot be assigned to a committee or to a generic title.

Zluri also maintains a confidential channel through which staff, contractors, and third parties can raise concerns about any AI system without reprisal. Concerns are logged, triaged by the committee, and closed with a recorded disposition.

‍

5. The AI Governance Committee

The AI Governance Committee is Zluri's top-level body for responsible AI. It is defined here by role and mandate. Its current membership is maintained as a separate controlled roster, so that personnel changes do not require a revision of this framework.

Composition (by seat, not by name).

Seat	Held by
Chair	The Chief Technology Officer and Data Protection Officer
Legal and compliance	The owner of this framework
Engineering	The leader accountable for AI engineering
Security and privacy	The leader accountable for information security and privacy
Product	A product leader representing AI capabilities

Seats may be held by delegates where the roster records the delegation. The committee is quorate when the chair or their delegate and at least three other seats are present.

‍

Mandate.

Ratify and maintain this framework and the overarching AI policy.
Own the AI risk register and the Statement of Applicability.
Approve new AI use cases, higher-risk classifications, and material changes to model posture or agent authority before general availability.
Review AI incidents and monitoring metrics at least quarterly.
Report AI risk to the board at least annually.

Authority. The committee holds explicit veto over the release of any in-scope AI capability. The committee is not the accountable owner of any capability; accountability rests with the named individual under Section 4.

Cadence. The committee meets at least quarterly and on demand for incidents or urgent approvals. Decisions and residual-risk acceptances are minuted.

‍

6. AI system inventory and risk classification

Zluri maintains an AI System Register as the authoritative, living record of its AI capabilities. Because capabilities change, the register lives outside this framework and is updated continuously; this framework defines how the register works, not its current contents.

For each capability the register records its purpose, accountable owner, provider or deployer role, the data it uses and where that data comes from, tooling, human-oversight requirements, its risk classification, and references to its impact assessment and guardrail baseline. The register template is in Annex A.

Risk classification. Each capability is classified using the EU AI Act risk tiers, because that is the most demanding tiering Zluri is subject to and it maps cleanly onto control depth:

Risk tier	Description
Prohibited	Zluri operates none.
High-risk	AI systems that fall within applicable regulatory high-risk classifications are subject to enhanced governance, regulatory review, and additional control requirements before deployment. If a capability comes to materially determine a person's access, employment, or other significant outcome autonomously, it may fall here and would then require the enhanced regulatory regime. Any such change is reclassified immediately and escalated to the committee before release.
Limited-risk (transparency)	Capabilities that converse with people or generate content, which trigger user-disclosure duties (Section 8.8).
Minimal-risk	Capabilities such as matching, scoring, and retrieval that do not interact directly with people.

Agentic capabilities receive enhanced controls (Section 9) regardless of their external tier, because they can take action.

‍

7. AI risk management

Zluri manages AI risk through a defined and repeatable process.

Stage	How it works
Criteria	AI risks are scored on likelihood and on impact, where impact spans safety, fairness, privacy, security, financial and access consequences, and societal effect.
Assessment	Risks are identified at design, at the review gate, and on material change, drawing on the impact assessment (Section 10), AI-specific threat modeling (Section 8.4), and production monitoring (Section 13).
Treatment	Each risk is mitigated, transferred, avoided, or accepted. Residual risk may only be accepted through documented approval by the AI Governance Committee or another delegated approval authority established by management.
Record	The controls selected to treat AI risk are recorded against recognized control references in the Statement of Applicability, with the justification for each inclusion. This record, not this framework's structure, is where the standards mapping is operationalized. The AI risk register is reviewed at least quarterly and on material change.

‍

8. Operating dimensions and controls

Zluri's principles are put into practice through eight operating dimensions. Each dimension states a commitment and the controls that enforce it, with the standards that justify those controls noted in brackets.

8.1 Fairness. Zluri evaluates AI output for biased or inequitable behavior. Accuracy is measured against governed evaluation data with a defined threshold per capability. Fairness measures appropriate to the use, such as differences in outcome across comparable groups, are defined and gated before release, with the primary measure for access-affecting decisions being equality of opportunity. Where privacy controls limit testing on protected attributes, Zluri applies segment analysis on permitted dimensions and mandatory human review of consequential decisions. (Informed by ISO/IEC 42001 impact-assessment controls, NIST AI RMF MEASURE, and EU AI Act data-governance expectations.)

8.2 Explainability. Retrieval-based capabilities surface the evidence used to produce an output, and the model configuration behind an output is recorded. Each capability has a model or system card documenting intended use, inputs, limitations, evaluation results, and known failure modes. (Informed by ISO/IEC 42001 lifecycle-documentation controls and NIST AI RMF MEASURE and MAP.)

8.3 Privacy and security. AI capabilities inherit Zluri's platform security and privacy controls, including personal-data tokenization and vaulting, customer-controlled encryption, encryption in transit and at rest, strong authentication and privileged access management, and tenant isolation. Customer data is not used to train foundation models. These controls are inherited from Zluri's ISO/IEC 27001 information security and ISO/IEC 27701 privacy management systems and are not restated here. This commitment applies to the customer Input and Output defined in the Zluri AI Terms (together, “AI Content”), and is stated there without the “in a way that could impact other customers” qualifier, so the two documents read consistently. (Informed by ISO/IEC 27001, ISO/IEC 27701, and ISO/IEC 42001 data controls.)

8.4 Safety. A baseline guardrail configuration is defined and, on the schedule set by the AI Governance Committee, applied across AI capabilities, covering content filtering on input and output, prompt-attack protection, detection and redaction of sensitive data including secrets, denied topics, and thresholds for contextual grounding and answer relevance. The specific baseline settings are maintained as a guardrail configuration standard referenced by the register. The safety dimension addresses the full set of recognized large-language-model application risks, including prompt injection, sensitive-information disclosure, supply-chain risk, data and model poisoning, improper output handling, excessive agency, system-prompt leakage, weaknesses in vector and embedding stores, misinformation, and unbounded consumption (including cost and usage anomaly alerting to guard against denial-of-wallet). (Informed by OWASP Top 10 for LLM Applications and ISO/IEC 42001 verification controls.)

8.5 Controllability. Model versions are pinned, capabilities are released through feature flags and staged rollout, service-level rate limits apply, and rollback is available. Autonomous agents operate within an explicitly bounded action set, and human oversight is required for decisions that materially affect a person. (Informed by NIST AI RMF MANAGE and GOVERN, and ISO/IEC 42001 deployment and use controls.)

8.6 Veracity and robustness. A minimum accuracy threshold is defined per capability and enforced as a release gate that rejects regressions. Retrieval and generation are evaluated separately. Adversarial and robustness testing, including prompt-injection and jailbreak red-teaming, is a gate criterion. A capability that cannot meet its accuracy or robustness bar is not shipped.The customer-facing accuracy commitment is stated in the Zluri AI Terms; the per-capability thresholds here are how that commitment is met and evidenced. (Informed by NIST AI RMF MEASURE and ISO/IEC 42001 verification and validation controls.)

8.7 Governance. This dimension is implemented through the governance structure (Sections 4 and 5), the AI System Register (Section 6), the risk management process (Section 7), and the Statement of Applicability. (Informed by ISO/IEC 42001 management-system requirements and NIST AI RMF GOVERN.)

8.8 Transparency. Zluri publishes its responsible-AI position, maintains internal model documentation, and, where a capability interacts with a person or generates content on their behalf, discloses that AI is in use and labels AI-assisted output where disclosure is expected. Applicable transparency obligations are monitored and implemented in accordance with relevant legal and regulatory requirements.(Informed by EU AI Act transparency obligations and NIST AI RMF GOVERN and MAP.)

‍

9. Human oversight and agentic AI

Autonomous and agentic capabilities can take action, so they carry enhanced controls regardless of their regulatory tier:

Control	Requirement
Bounded autonomy	An agent operates only within an explicitly scoped set of permitted actions.
Tiered human oversight	Each class of agent action is mapped to the required human control, ranging from notification, to prior approval, to prohibition. Privileged or irreversible actions require human approval.
Traceability	Every agent action is logged with its inputs, decision path, and outcome.
Alignment testing	Goal alignment and refusal of unsafe actions are evaluation criteria at the review gate.

(Informed by the NIST AI RMF generative-AI risk guidance and the OWASP excessive-agency risk.)

‍

10. AI system impact assessment

Zluri completes an impact assessment before general availability of any new AI or agent capability, and on material change such as a new model posture, an expanded agent action set, a new data category, or a change in the provider or deployer role.

The assessment evaluates impact on individuals and groups across fairness, privacy, safety, security, and financial and access consequences, and considers societal and environmental effect. It records residual risk and is signed by the accountable owner and the security and privacy seat. For capabilities processing personal data, the AI impact assessment is conducted jointly with the data protection impact assessment, producing one integrated record that also satisfies applicable data-protection obligations. Assessments are retained as controlled documents and reviewed each annual cycle. The template is in Annex B.

‍

11. AI development and release controls

Every AI capability passes through the Responsible AI review gate before release. Production deployment shall not proceed until the gate confirms:

No.	Gate check	Reference
1	A completed impact assessment	Section 10; Annex B
2	An updated AI System Register entry	Section 6; Annex A
3	The fairness threshold is met	Section 8.1
4	Adversarial and robustness tests are passed	Section 8.6
5	The accuracy threshold is met with no regression	Section 8.6
6	Foundation-model supplier attestations are current	Section 12

Sign-off is recorded with the reviewer, date, and evidence links. Changes to privileged or irreversible agent actions additionally require chair approval.

‍

12. Third-party and foundation-model risk

Zluri concentrates its AI supply chain onto a small number of assessed, contracted, and certified providers, deployed through its AI governance process; the current providers and regions are recorded in the AI System Register. This concentration is itself a risk control.

Allocation of responsibility. A shared-responsibility matrix (Annex D) records what Zluri, the cloud provider, and the model provider are each responsible for.

Supplier due diligence. Zluri relies on and verifies cloud and model-provider attestations, including the relevant service certifications, the commitment that prompts and outputs are not used to train models, and the inference-time isolation between model-provider tenants. Attestation currency is checked at least annually and at each gate.

Customer obligations. Customer-facing responsibilities, including AI disclosure and the sub-processor disclosure in the Data Processing Agreement, are maintained by Legal, and reflect the current AI supply chain recorded in the Register.

MCP external-client boundary. AI interfaces exposed to external clients. Zluri exposes certain data through AI interfaces, for example an MCP server, to which external customer-chosen AI clients may connect. These interfaces are read-only, and access is authenticated and scoped to the requesting tenant, so a client can reach only its own tenant's data. Tool outputs minimize and redact sensitive data. Zluri operates no model at this boundary: once data reaches the connecting client, that client's model governs it, including any training or retention behavior, which is outside Zluri's control and is disclosed to customers. Introducing any write or state-changing capability is a material change (Section 10) requiring re-assessment, review-gate approval, and the Section 9 agentic controls before release.

Financial-sector customers. Where a customer is subject to operational-resilience regulation, Zluri supports the customer's register of ICT providers, agrees the required contractual clauses, cooperates on incident reporting, and maintains an exit path.

‍

13. Incident management and monitoring

AI incident taxonomy. Harmful or unsafe output; prompt-injection or jailbreak compromise; agent overreach or unsafe action; accuracy or drift breach; data leakage; model or provider outage; and cost or consumption anomaly.

Reporting timeline. An AI incident is reported internally immediately and no later than 24 hours to the security and privacy seat and the accountable owner. This single internal clock supersedes any conflicting service-level target. Downstream obligations run in parallel:

Obligation	Notified to	Timing
Personal-data breach	Relevant supervisory authority, coordinated through dpo@zluri.com	Within the statutory window
Customer notification	Customers	As per the executed Data Processing Agreement
Regulatory serious-incident and operational-resilience reporting	Relevant regulators	Where applicable

Monitoring. Production dashboards track accuracy, latency, cost and consumption, and drift. A threshold breach raises an incident and feeds the risk register.

‍

14. Data governance for AI

Although Zluri does not train foundation models, it curates evaluation data and runs retrieval over customer data. For that data Zluri records provenance and lineage, defines quality and acceptance criteria, documents the labeling process and reviewers, reviews evaluation data for representativeness relevant to its fairness measures, and applies access control, tenant isolation, and freshness controls to embedding and retrieval stores. (Informed by ISO/IEC 42001 data controls and OWASP vector and embedding risk.)

‍

15. Competence, awareness, and acceptable use

Staff who build, operate, maintain or make decisions with AI complete AI-literacy and responsible-AI training appropriate to their role, and completion is recorded and retained as evidence. All staff are made aware of this framework, the overarching AI policy, and the Acceptable Use of AI policy, which governs staff use of AI tools including approved third-party AI services and out-of-scope internal tooling. Use of foundation models over live customer data for debugging or incident investigation is a governed activity: it is access-controlled, logged for audit purposes, logged, restricted to authorized staff, and subject to the same confidentiality, no-training and data-handling obligations, and it is not exempt from this framework. (Informed by EU AI Act AI-literacy obligations and ISO/IEC 42001 competence controls.)

‍

16. Assurance, review, and continuous improvement

Zluri audits its AI management system against this framework and the Statement of Applicability at least annually. The AI Governance Committee conducts a management review at least annually covering audit results, incidents, metrics, risk status, and regulatory change. Nonconformities are logged, root-caused, and closed, and this framework, the Statement of Applicability, and the risk register are updated accordingly. This framework is reviewed at least annually and out of cycle on material change to model posture, agent authority, or applicable regulation.

‍

Annex A: AI System Register template

The register is maintained as a living record outside this framework. The fields recorded per capability are shown below, with an illustrative example for format only (not a scoped commitment).

Field	Illustrative example
Capability name and purpose	Conversational assistance capability
Accountable owner (named)	[name]
Provider or deployer role	Deployer and provider
Data categories and provenance	Customer configuration and usage data
Tooling and integrations	—
Human-oversight requirements	—
EU AI Act risk tier	Limited-risk (transparency)
Impact assessment reference	[ref]
Guardrail baseline reference	[ref]
Model posture reference (cloud provider, model provider, and deployment region)	—
Last reviewed date	[date]

‍

Annex B: AI system impact assessment template

No.	Section	What to record
1	Capability details	Capability name, accountable owner, and provider or deployer role
2	Purpose	Purpose and intended use; out-of-scope use
3	Data	Data categories, provenance, and personal-data determination (link the joint data protection impact assessment)
4	Affected parties	Affected individuals and groups
5	Impact analysis	Fairness, privacy, safety, security, financial and access consequences, societal and environmental effect
6	Applicable controls	Map to Section 8 and the Statement of Applicability
7	Residual risk	Residual risk and treatment decision
8	Risk tier	Risk tier and any transparency obligations
9	Sign-off	Accountable owner and security and privacy seat, with date
10	Review date	Date of the next review

‍

Annex C: Model and system card template

Each model or system card records the following:

Section	Content
Intended use	Intended use and users
Scope	In-scope and out-of-scope tasks
Model posture	Underlying model posture (cloud provider and model provider)
Inputs and outputs	Inputs and outputs
Evaluation	Evaluation datasets and results, including accuracy and fairness measures
Limitations	Known limitations and failure modes
Guardrails	Guardrail configuration reference
Human oversight	Human-oversight requirements
Last updated	Last-updated date

‍

Annex D: Shared-responsibility matrix

Area	Cloud provider	Model provider	Zluri
Infrastructure and model-service security	Owns		Configures
Foundation-model safety, alignment, model documentation		Owns	Consumes and verifies
Model isolation and no-training commitment	Owns and attests	Supports	Relies on and verifies
Guardrail configuration			Owns
Prompts, retrieval data, application logic			Owns
Human oversight and impact assessment			Owns
Customer disclosure and sub-processor disclosure			Owns

‍

Annex E: Standards basis and control mapping

This annex records the recognized standards that inform each operating dimension. It is reference material; the framework body is organized around Zluri's principles, not these standards. The formal, control-level record is the Statement of Applicability maintained by the AI Governance Committee, and the AI controls inherit from Zluri's existing ISO/IEC 27001 and ISO/IEC 27701 certifications rather than duplicating them.

Dimension	ISO/IEC 42001 basis	NIST AI RMF	EU AI Act	OWASP LLM
Fairness	Impact assessment; data quality	MEASURE; MAP	Data governance	Data and model poisoning
Explainability	Lifecycle documentation	MEASURE; MAP	Information to users	Misinformation
Privacy and Security	Data controls (with 27001 and 27701 inheritance)	MANAGE; MEASURE	Data governance	Sensitive-information disclosure; system-prompt leakage
Safety	Verification and validation	MANAGE; MEASURE	Accuracy and robustness	Prompt injection; improper output; embedding weaknesses; unbounded consumption
Controllability	Deployment and use	MANAGE; GOVERN	Human oversight	Excessive agency
Veracity and Robustness	Verification and validation; logging	MEASURE	Accuracy and robustness	Prompt injection; misinformation
Governance	Management-system requirements	GOVERN	Risk management (if high-risk)	Supply chain
Transparency	Information for interested parties	GOVERN; MAP	Transparency obligations	Misinformation

‍

Definitions

For the purposes of this framework the following terms apply. Terms are ordered alphabetically.

Term	Definition
Accountable owner	The single named individual answerable for a given AI capability across its lifecycle. Recorded and dated in the AI System Register. Cannot be a committee or a generic title (Section 4, Tier 3).
Agentic AI	An AI system that can plan and take actions toward a goal with reduced step-by-step human instruction, including invoking tools or changing state. Agentic capabilities carry the enhanced controls in Section 9 regardless of regulatory tier, because they can act and not only generate.
AI capability	A specific Zluri feature or function delivered by an AI system, for example catalog and data intelligence, identity-risk scoring, conversational assistance, or agentic automation. The AI capability is the unit that is registered, classified, owned, and gated under this framework. The current set is maintained in the AI System Register, not fixed in this document.
AI Content	As defined in the Zluri AI Terms, the customer Input to and Output from Zluri AI, taken together. Zluri does not use AI Content to train foundation models (Section 8.3).
AI Governance Committee	Zluri's top-level body for responsible AI, defined by seat and mandate in Section 5, with current membership held in a separate controlled roster.
AI system	A machine-based system that, for explicit or implicit objectives, infers from the input it receives how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. In this framework, an AI system is any deployed configuration of models, data, tooling, and logic that delivers an AI capability.
AI system impact assessment	The documented evaluation of an AI capability's impact on individuals, groups, society, and the environment, completed before general availability and on material change, and conducted jointly with the data protection impact assessment where personal data is processed (Section 10, Annex B).
AI System Register	The authoritative, living record of Zluri's AI capabilities and their attributes, maintained outside this framework and updated continuously (Section 6, Annex A).
Foundation model	A general-purpose model trained on broad data at scale and adaptable to many downstream tasks, which Zluri consumes from an approved external model provider rather than trains. Zluri does not train or fine-tune foundation models (Section 12).
Generative AI	A class of AI system that produces new content such as text, code, or structured output in response to a prompt, typically built on a foundation model. Generative capabilities are subject to the transparency duties in Section 8.8 and the safety controls in Section 8.4.
Guardrail baseline	The defined minimum configuration of automated safety controls, covering content filtering, prompt-attack protection, sensitive-data detection and redaction, denied topics, and thresholds for contextual grounding and answer relevance, applied across AI capabilities and maintained as a referenced configuration standard (Section 8.4).
Human oversight	Human review, approval, or intervention applied to an AI capability so that a person can understand, correct, override, or halt its behavior. The required level ranges from notification, to prior approval, to prohibition (Section 9), and is mandatory for decisions that materially affect a person (Sections 2.3 and 8.5).
Material change	A change to an AI capability significant enough to require re-assessment before release, including a new model posture, an expanded or altered agent action set, a new category of data, a change in provider or deployer role, a move to a higher risk tier, or the introduction of write or state-changing capability at an external interface. A material change triggers a new impact assessment (Section 10) and review-gate approval (Section 11).
Personal data	Information relating to an identified or identifiable natural person, as defined by applicable data protection law including the GDPR and the India Digital Personal Data Protection Act. Handled under Zluri's ISO/IEC 27701 privacy information management system (Section 8.3).
Provider and deployer	Role designations for an AI capability. Zluri is a provider when it develops or supplies an AI system, and a deployer when it operates an AI system under its own authority. The role is recorded in the register and determines the applicable regulatory obligations (Section 6, Annex A).
Residual risk	The risk that remains after controls are applied. Accepted only through documented approval by the AI Governance Committee or a delegated approval authority (Section 7).
Responsible AI review gate	The operational control through which every AI capability must pass before release, confirming the checks listed in Section 11.
Retrieval-augmented generation (RAG)	A technique that supplies an AI system with relevant retrieved content at inference time, using an embedding and retrieval store, to ground its output. Subject to the data-governance and embedding-store controls in Sections 8.4 and 14.
Statement of Applicability	The controlled record of which controls apply to which AI systems and why, mapped to the recognized standards in Section 3, maintained by the AI Governance Committee (Sections 3 and 7).