Responsible AI

NOVA™ is self-assessed against NATO’s six Principles of Responsible Use of AI.

NOVA is developed and operated in line with NATO’s Principles of Responsible Use of AI. It is self-assessed against all six principles, which is the mechanism NATO currently prescribes while its Data and Artificial Intelligence Review Board finalises a formal Responsible AI certification standard.

The six principles, and how NOVA meets them

Lawfulness. Each nation authors under its own doctrine and law and keeps full data sovereignty. In place
Responsibility & accountability. A human review workflow governs every output; actions are audited; nothing publishes without approval. In place
Explainability & traceability. Every output is grounded in doctrine, cited, and traceable to source. In place
Reliability. Deterministic where possible, grounded to doctrine, structurally validated, with a no-fabrication rule. In place
Governability. Human review can amend or reject any output; generation can be paused; nothing auto-actions. In place
Bias mitigation. Outputs are constrained to authoritative allied doctrine; formal bias assessment is on the roadmap. Partial / roadmap

Status: Self-assessed · certification pending. NOVA awaits NATO’s Responsible AI certification standard; NATO sets the standard, while certification and enforcement rest with individual nations or an appointed third party. NOVA does not claim NATO endorsement, affiliation or certification.

Governance documents

These are the five governance papers that sit behind the self-assessment above. Each one opens here on the page. You can widen it to full screen or copy the whole text.

Model CardNOVA-MC-001 · What NOVA is, what it is built on, what it is for, and where its limits are.
AI Ethical Risk AssessmentNOVA-AERA-001 · The ethical risks identified in NOVA, how likely each is, and what is done about it.
Bias and Harm Mitigation AnalysisNOVA-BHMA-001 · Where bias or harm could enter the outputs, and the controls that hold it back.
Statement of AI Ethics AssuranceNOVA-SAEA-001 · The formal statement of how NOVA is held to the six principles above.
Security Risk AssessmentNOVA-SRA-001 · The security risks to the platform and the data it holds, and the measures in place.

NOVA™ Ltd · Responsible AI · Self-assessed against NATO’s Principles of Responsible Use of AI (NATO AI Strategy, 2021). Certification pending the NATO Responsible AI certification standard.

Model CardNOVA-MC-001

⚠️ ARTIFICIAL INTELLIGENCE DISCLOSURE This document was produced with AI assistance via NOVA™ (Anthropic Claude Sonnet / Pinecone RAG). All content has been reviewed and signed off by the named responsible person. Human approval constitutes acceptance of accuracy and doctrinal compliance.

NOVA™ MODEL CARD

NOVA™ — Agentic Allied Defence Training Platform

Document Ref: NOVA-MC-001 | 8 March 2026 | Will Kennedy-Long FLPI, FITOL

JSP 936 Reference: Cover Note (Nov 2024) — Model Card is a required Minimum Viable Product for AI-enabled Defence systems. JSP 936 §63 — appropriate mechanisms must permit stakeholder understanding of AI development and operation.

1. Model Details

Platform Name	NOVA™ — Agentic Allied Defence Training Platform
Version	v1.0 (March 2026)
Developed by	Will Kennedy-Long FLPI, FITOL
AI Model (LLM)	Claude Sonnet 4.6 — Anthropic PBC (claude-sonnet-4-6)
Retrieval System	Pinecone Vector Database — 4 doctrine indexes: nova-uk-doctrine, nova-us-doctrine, nova-nato-doctrine, nova-asd-doctrine
Embedding Model	OpenAI text-embedding-3-small (used for Pinecone ingestion)
Doctrine Corpus	134 doctrine files, 207,000+ lines. UK (JSP 822, DTSM 1–5), US (TRADOC 350-70), NATO (Bi-SC 75-7), ASD/AIA (S6000T)
Classification	OFFICIAL only. Not suitable for OFFICIAL-SENSITIVE, SECRET or above.

2. Intended Use

2.1 Primary Use Cases

Automated generation of DSAT-compliant training needs analysis documentation (JSP 822 / DTSM 1–5)
AI-assisted extraction of training data from Statement of Requirement, RolePS, Scoping, TGA, SoTG and FTS source documents
Doctrinal compliance checking of training design documents against JSP 822, DTSM, TRADOC 350-70, NATO Bi-SC and S6000T
Generation of SCORM 1.2 e-learning packages from DSAT Training Objectives
Cross-doctrinal synonym resolution across UK, US, NATO and ASD/AIA frameworks

2.2 Out-of-Scope Use

WARNING — NOVA must NOT be used for: Kinetic targeting or weapons selection | Command and control decision support | Intelligence analysis or threat assessment | Processing of SECRET or above classified material | Any application outside the bounded domain of training design documentation

2.3 Operational Design Domain (ODD)

NOVA is designed and tested to operate within the following bounded context:

ODD Dimension	Definition
Domain	Defence training design and DSAT lifecycle documentation
Users	Trained TDOs and training design specialists with DSAT knowledge
Documents	SoR, RolePS, Scoping, TGA, SoTG, FTS, TNA activities (21 activities across 4 DSAT phases)
Language	English only
Classification	OFFICIAL only
Frameworks	JSP 822, DTSM 1–5, TRADOC 350-70, NATO Bi-SC 75-7, ASD/AIA S6000T

3. Performance

3.1 Reported Metrics

Metric	Result	Basis
Doctrinal Compliance Rate	96%+	Testing against EW Course 2303C documents
DSAT Cycle Time Reduction	4,400 hrs → ~3 hrs (99.9%)	Comparison of manual vs NOVA-assisted TNA
Cost Reduction (est.)	£386K → £50K (87%)	Contractor day rate estimates
Synonym Coverage	1,519+ terms, 9,647 synonyms	178 cross-doctrinal categories

3.2 Known Limitations

Claude Sonnet may generate plausible-sounding but doctrinally incorrect content if doctrine retrieval fails. Human review mitigates this.
RAG retrieval quality depends on doctrine index completeness. Doctrine files are verified but may not reflect very recent amendments.
NOVA does not process audio, video or image content within source documents.
Performance degrades outside the defined ODD. NOVA should not be used for roles, frameworks or document types not covered in the doctrine index.
English language only. No support for non-English source documents.
Pinecone retrieval is dependent on US-hosted service. Downtime affects AI generation capability (graceful degradation to non-AI mode).

4. Data and Privacy

Data Category	Handling
User-uploaded documents	Parsed client-side (PizZip/fast-xml-parser CDN) or transiently via Railway backend. Not persistently stored beyond user session.
AI prompts / completions	Sent to Anthropic API. Per Anthropic API Terms of Service, prompts are not used to train models. No persistent storage by Anthropic for API customers.
Activity data	Stored in Cloudflare D1/KV under user account. User-controlled. Exportable at any time.
Doctrine content	Stored in Pinecone vector DB. Doctrine files only. No user PII.
PII	Minimal. User email/name for account only. Not passed to AI services.

5. Ethical Considerations

This Model Card should be read alongside the NOVA AI Ethical Risk Assessment (NOVA-AERA-001), Bias and Harm Mitigation Analysis (NOVA-BHMA-001) and Security Risk Assessment (NOVA-SRA-001).

NOVA implements the five MOD AI Ethical Principles (ASR 2022) through: mandatory human sign-off at every lifecycle stage; doctrine-grounded generation (RAG); transparent criteria scoring with clause references; AI disclosure on all exports; and a bounded ODD preventing use outside training design.

6. Document Control

Document Owner	Will Kennedy-Long FLPI, FITOL
Version	v1.0
Date	8 March 2026
Review Frequency	Annual minimum, or on material change to AI components
Next Review	March 2027

AI Ethical Risk AssessmentNOVA-AERA-001

AI ETHICAL RISK ASSESSMENT

NOVA™ — Agentic Allied Defence Training Platform

Prepared by: Will Kennedy-Long FLPI, FITOL | Date: 8 March 2026 | Version: v1.0

Document Reference	NOVA-AERA-001
JSP Reference	JSP 936 V1.1 §86 — AI Ethical Risk Assessment and Management
ASR Reference	Ambitious, Safe, Responsible (ASR) 2022 — Five Ethical Principles
Classification	OFFICIAL
Next Review Date	March 2027 (annual minimum)
Responsible Person	Will Kennedy-Long FLPI, FITOL

1. Introduction and Scope

This AI Ethical Risk Assessment (AERA) has been prepared in accordance with JSP 936 V1.1 (Dependable AI in Defence) §86, which mandates that an AI ethical risk assessment addressing the five MOD AI Ethical Principles must be conducted at the outset of any project and at any point where material changes to scope suggest changes to the overall risk profile.

NOVA™ is a training design automation platform that automates the UK Defence Systems Approach to Training (DSAT) lifecycle. It provides AI-assisted generation of training needs analysis, design, delivery and assurance documentation in compliance with JSP 822, DTSM 1–5, TRADOC 350-70, NATO Bi-SC 75-7 and ASD/AIA S6000T.

NOVA is intended for use by UK MOD Training Development Officers (TDOs), training design specialists, and allied defence training authorities. Classification scope is OFFICIAL only.

Operational Design Domain (ODD): NOVA operates exclusively within the bounded domain of defence training design documentation. It does not operate in kinetic systems, command and control, weapons selection, intelligence analysis, or any domain with direct physical effects. All outputs are training documents subject to mandatory human review and sign-off before use.

2. System Description

2.1 AI Components

Component	Provider	Function	Data Residency
Claude Sonnet (LLM)	Anthropic PBC (USA)	Document generation, extraction, compliance checking	US-based API. No persistent storage of prompts per Anthropic API ToS.
Pinecone Vector DB	Pinecone Inc (USA)	RAG retrieval of doctrine content (134 doctrine files)	US-based. Doctrine content only — no user PII stored.
Cloudflare Pages / D1 / KV	Cloudflare Inc (USA/EU)	Frontend hosting, user data storage, session management	EU-based data centres where available. User activity data stored in D1/KV.
Railway FastAPI Backend	Railway Corp (USA)	Analysis agent, document parsing orchestration	US-based. Processes document content transiently during extraction.

2.2 Human Control Mechanisms

NOVA incorporates mandatory human oversight at every stage of the DSAT lifecycle:

Every AI-generated output requires a human TDO to review and explicitly "Mark Complete" before it is accepted or cascaded to downstream documents.
No document is generated, stored as final, or exported without positive human action.
The criteria compliance system surfaces doctrinal references so operators can verify AI reasoning against source doctrine.
All generated DOCX exports carry a mandatory AI Disclosure statement on the cover and footer per JSP 936 §40.
Users can edit, override, or reject any AI-generated content at any point in the workflow.

3. Assessment Against MOD AI Ethical Principles

The following assessment addresses each of the five MOD AI Ethical Principles as defined in the Ambitious, Safe, Responsible (ASR) policy (2022) and elaborated in JSP 936 V1.1 §49.

3.1 Human-Centricity

Principle: The impact of AI-enabled systems on humans must be assessed and considered, for a full range of effects both positive and negative across the entire system lifecycle.

Stakeholder Group	Impact Assessment
TDOs / Training Designers	Positive: Reduction in administrative burden (est. 4,400 hours to ~3 hours per TNA cycle). Risk: Over-reliance on AI outputs without adequate doctrinal knowledge. Mitigation: Criteria system requires human judgement; Activity Guide provides doctrinal grounding.
Trainees (end beneficiaries)	Positive: More consistently DSAT-compliant training programmes. Risk: Errors in AI-generated Training Objectives could affect training quality. Mitigation: Human sign-off at TNA, Design, Delivery and Assurance stages.
MOD Institution	Positive: Significant cost reduction, doctrinal consistency across TLBs. Risk: Concentration of training design tooling in single supplier. Mitigation: All data exportable; no lock-in architecture.
General Public	No direct interaction. Indirect benefit through improved military training quality and readiness.

Assessment	Risk Level	Residual Risk
Human-Centricity	Low — human control embedded throughout	Low — mitigations in place

3.2 Responsibility

Principle: Human responsibility for AI-enabled systems must be clearly established, ensuring accountability for their outcomes, with clearly defined means by which human control is exercised throughout their lifecycles.

Responsibility within NOVA is clearly established at multiple levels:

Platform Level: Will Kennedy-Long FLPI, FITOL bears responsibility as sole developer and data controller for NOVA.
Organisational Level: Each MOD TLB using NOVA must appoint a Responsible AI Senior Officer (RAISO) per JSP 936 §100. NOVA’s supplier documentation supports this requirement.
Operational Level: The TDO who marks an activity complete accepts doctrinal responsibility for that output. The cascade architecture ensures each decision is traceable to a named user action.
No AI output enters the training design record without a positively identified human approver. This eliminates the accountability gap described in JSP 936 §58.

Assessment	Risk Level	Residual Risk
Responsibility	Low — clear chain of accountability	Low

3.3 Understanding

Principle: AI-enabled systems and their outputs must be appropriately understood by relevant individuals, with mechanisms to enable this understanding made an explicit part of system design.

NOVA addresses the Understanding principle through the following design features:

Every criteria check displays the specific JSP/DTSM clause reference, allowing operators to verify AI reasoning against source doctrine directly.
The NOVA Activity Guide explains what each activity does, why it is required, and what the AI generates — supporting operator calibration of trust.
AI-generated content is visually distinguished from human-entered content throughout the interface.
The Compliance Check function provides field-by-field transparency on what is and is not complete.
NOVA’s Model Card (NOVA-MC-001) documents system capabilities, limitations and intended use in detail.

Gap noted: A formal NOVA User Training Package is in development (NOVA-UTP-001) to address JSP 936 §130 — training users to understand system behaviour, performance and limitations.

Assessment	Risk Level	Residual Risk
Understanding	Moderate — pending user training package	Low–Moderate — UTP in development

3.4 Bias and Harm Mitigation

Principle: Those responsible for AI-enabled systems must proactively mitigate the risk of unexpected or unintended biases or harms from these systems.

See also: NOVA Bias and Harm Mitigation Analysis (NOVA-BHMA-001) for full detail. Summary:

NOVA generates training design documents, not decisions about individuals. Direct discriminatory harm to persons is therefore low risk.
Potential bias vectors: Claude Sonnet may reflect biases present in training data. Pinecone retrieval is bounded to 134 verified doctrine files, reducing open-web bias contamination.
Doctrine-grounded generation (RAG architecture) constrains outputs to approved doctrinal content rather than general model inference, substantially reducing free-form bias risk.
Role-based outputs (Training Objectives, KSAs, DIF analysis) are reviewed by TDOs with subject matter expertise before acceptance.
Anthropic publishes and maintains a Constitutional AI approach and model cards documenting known limitations of Claude Sonnet.

Assessment	Risk Level	Residual Risk
Bias & Harm Mitigation	Low–Moderate — RAG bounds generation	Low — human review required for all outputs

3.5 Reliability

Principle: AI-enabled systems must be demonstrably reliable, robust and secure.

Dimension	NOVA Assessment
Reliable	NOVA operates within a defined ODD (DSAT training documentation). Performance is bounded and testable against known doctrine. 96%+ doctrinal compliance reported in testing against verified EW course documents.
Robust	Fallback mechanisms prevent null outputs. All document builders are resilient to missing data and generate from available inputs. Error handling prevents cascade failures from propagating.
Secure	See NOVA Security Risk Assessment (NOVA-SRA-001). OFFICIAL classification only. Data flows documented. No SECRET or above data should be processed through NOVA.
Resilient	Service Worker implementation ensures client-side resilience. Cloudflare Pages provides 99.9%+ uptime SLA. Railway backend is independently restartable.

Assessment	Risk Level	Residual Risk
Reliability	Moderate — depends on third-party API availability	Low–Moderate — documented fallbacks in place

4. Overall Risk Rating

Per JSP 936 V1.1 Table 1, the overall ethical risk rating for NOVA is assessed as follows:

Dimension	Impact	Likelihood	Residual Risk Rating
Human-Centricity	Low	Low	Minor
Responsibility	Low	Low	Minor
Understanding	Moderate	Low	Moderate (UTP pending)
Bias & Harm Mitigation	Low	Low	Minor
Reliability	Moderate	Low	Moderate
OVERALL RATING	Moderate	Low	MODERATE — TLB-Level Oversight

JSP 936 Table 1 Referral Level: MODERATE — TLB-Level Oversight, with delegation to business level processes as appropriate. No referral to JROC/IAC or Ministers required at this risk level.

5. Mitigations and Actions

Ref	Mitigation / Action	Status	Target Date
M-01	AI Disclosure on all generated documents (JSP 936 §40)	COMPLETE	March 2026
M-02	Model Card published (NOVA-MC-001)	COMPLETE	March 2026
M-03	Security Risk Assessment published (NOVA-SRA-001)	COMPLETE	March 2026
M-04	Bias and Harm Mitigation Analysis published (NOVA-BHMA-001)	COMPLETE	March 2026
M-05	User Training Package (NOVA-UTP-001)	IN DEVELOPMENT	Q2 2026
M-06	Annual review of this AERA (next: March 2027)	SCHEDULED	March 2027
M-07	Monitor Anthropic API updates for model behaviour changes	ONGOING	Continuous

6. Sign-Off

Name	Will Kennedy-Long FLPI, FITOL
Role	Founder and Developer, NOVA™ Agentic Allied Defence Training Platform
Date	8 March 2026
Signature	(Signed electronically — document owner acceptance constitutes sign-off)
Next Review	March 2027

Bias and Harm Mitigation AnalysisNOVA-BHMA-001

BIAS AND HARM MITIGATION ANALYSIS

NOVA™ — Agentic Allied Defence Training Platform

Document Ref: NOVA-BHMA-001 | 8 March 2026 | Will Kennedy-Long FLPI, FITOL

JSP 936 Reference: JSP 936 V1.1 §69 — An analysis of data, AI learning algorithms and models must be made for unwanted bias that may lead to unintentional harms. JSP 936 §70 — Where harms may arise, monitoring and mitigation strategies must be developed.

1. Scope and Purpose

This document analyses potential bias and harm vectors in NOVA™ across its AI components (Claude Sonnet LLM, Pinecone RAG) and the NOVA platform design. It identifies mitigations in place and residual risks requiring ongoing monitoring.

NOVA operates in the bounded domain of defence training design. It does not make decisions about individuals, determine employment, assess performance, or recommend disciplinary action. This substantially constrains the harm surface compared to AI systems with direct human impact.

2. Potential Bias Vectors

2.1 LLM Training Data Bias (Claude Sonnet)

Claude Sonnet (Anthropic) is trained on large-scale internet and curated text corpora. Potential biases include:

Bias Vector	NOVA-Specific Risk	Mitigation
Western / English-language bias in training data	Low — NOVA operates in English within Allied frameworks that are already English-language	ODD restricted to English. NATO/allied doctrine in English.
Gender bias in role descriptions	Low-Moderate — LLM may generate gender-stereotyped role descriptions or training objectives	TDO review required before sign-off. RolePS source documents define roles; AI extracts rather than invents.
Temporal bias (stale doctrine knowledge)	Moderate — Claude’s training cutoff may not reflect recent JSP/DTSM amendments	Pinecone RAG retrieves from verified current doctrine files. Claude’s general knowledge is secondary to RAG outputs.
Hallucination (confident incorrect output)	Moderate — LLM may generate plausible but incorrect doctrinal references	Every criteria check shows JSP/DTSM clause reference. Human TDO verifies. RAG constrains generation to sourced content.

2.2 RAG Retrieval Bias (Pinecone)

Bias Vector	NOVA-Specific Risk	Mitigation
Index over-representation of one doctrinal framework	Low — four separate doctrine indexes (UK, US, NATO, ASD/AIA). Each queried by framework.	Separate Pinecone indexes per framework prevent cross-contamination.
Stale doctrine in index	Moderate — index may not reflect latest JSP/DTSM amendments	Doctrine files reviewed and updated by owner. Version-controlled.
Retrieval failure (no relevant chunks returned)	Low — LLM falls back to general knowledge, increasing hallucination risk	Fallback prompts instruct Claude to flag uncertainty rather than fabricate.

2.3 Harm Analysis

Harm Category	Rating	Analysis
Physical harm	Very Low	NOVA generates training documentation. It does not control physical systems, weapons or autonomous platforms. Training design errors could theoretically affect training quality, but not cause direct physical harm.
Discriminatory harm to individuals	Low	NOVA does not assess, rank or make decisions about individual service personnel. Role descriptions are based on MOD-defined RolePS documents.
Doctrinal non-compliance harm	Moderate	If NOVA generates non-compliant training objectives that are not caught by human review, downstream training could fail JSP 822 standards. Mitigated by criteria scoring and mandatory TDO sign-off.
Data security harm	Moderate	User-uploaded documents passed to Anthropic API and Railway backend. Mitigated by OFFICIAL-only classification scope and Anthropic API data handling commitments. See NOVA-SRA-001.
Over-reliance / automation bias	Moderate	TDOs may accept AI outputs without adequate review if the platform appears authoritative. Mitigated by criteria transparency, clause references, and pending User Training Package.

3. Mitigations in Place

3.1 Design Mitigations

RAG Architecture: Generation is grounded in 134 verified doctrine files rather than open-ended model inference. This is the primary bias control mechanism.
Mandatory Human Sign-Off: Every activity requires a human TDO to review and mark complete. No output is finalised without positive human action.
Criteria Transparency: Every compliance check shows the JSP/DTSM clause reference, enabling operators to verify reasoning against source doctrine.
AI Disclosure: All generated documents carry mandatory AI disclosure per JSP 936 §40, ensuring recipients know AI was used.
Bounded ODD: NOVA is explicitly scoped to training design documentation and refuses to operate outside this domain.

3.2 Upstream Provider Mitigations (Anthropic)

Anthropic publishes and maintains a Constitutional AI (CAI) approach designed to reduce harmful outputs from Claude models.
Anthropic’s Acceptable Use Policy prohibits uses that cause harm. NOVA’s use case (training documentation) is clearly within acceptable use.
Anthropic’s API Terms of Service confirm that API inputs are not used to train Claude models, protecting sensitive defence content.
Anthropic publishes model cards and safety documentation for Claude Sonnet detailing known limitations and bias assessments.

3.3 Monitoring

Monitoring Activity	Frequency	Owner
Review Anthropic model card updates	Quarterly	Will Kennedy-Long FLPI, FITOL
Review Pinecone doctrine index for stale content	On JSP/DTSM amendment	Will Kennedy-Long FLPI, FITOL
User feedback review for bias indicators	Monthly	Will Kennedy-Long FLPI, FITOL
Annual review of this BHMA document	Annual	Will Kennedy-Long FLPI, FITOL

4. Overall Assessment

Conclusion: NOVA•s primary harm mitigation is its RAG-grounded, doctrine-bounded architecture combined with mandatory human review at every stage. The residual bias and harm risk is LOW to MODERATE, with the Moderate element attributable to third-party API data flows and automation bias risk — both of which are addressed through documented mitigations and the pending User Training Package.

5. Document Control

Document Owner	Will Kennedy-Long FLPI, FITOL
Version	v1.0
Date	8 March 2026
Next Review	March 2027

Statement of AI Ethics AssuranceNOVA-SAEA-001

STATEMENT OF AI ETHICS ASSURANCE

NOVA™ — Agentic Allied Defence Training Platform

Document Ref: NOVA-SAEA-001 | Annual Statement | 8 March 2026

JSP 936 Reference: JSP 936 V1.1 §99 — TLB Executive Boards must provide Statements of AI Ethical Assurance to 2PUS on an annual basis, underpinned by auditable evidence. This document is the supplier-level statement supporting that obligation for any MOD TLB using NOVA™.

1. Statement

I, Will Kennedy-Long FLPI, FITOL, as the developer and responsible person for NOVA™, hereby confirm the following in respect of the NOVA™ Agentic Allied Defence Training Platform for the period ending 8 March 2026:

	Assurance Statement	JSP 936 Reference	Evidence Reference
✓	NOVA has been assessed against all five MOD AI Ethical Principles (Human-Centricity, Responsibility, Understanding, Bias & Harm Mitigation, Reliability)	JSP 936 §49–50	NOVA-AERA-001
✓	An AI Ethical Risk Assessment has been completed and the overall residual risk is rated MODERATE (TLB-Level Oversight)	JSP 936 §86–90	NOVA-AERA-001
✓	All AI-generated documents carry mandatory AI Disclosure statements on cover and footer per JSP 936 §40	JSP 936 §40	NOVA v1.0 platform
✓	A Model Card has been published documenting AI components, ODD, performance, limitations and data handling	JSP 936 Cover Note	NOVA-MC-001
✓	A Security Risk Assessment has been completed covering all data flows, third-party services and classification constraints	JSP 936 §195–197	NOVA-SRA-001
✓	A Bias and Harm Mitigation Analysis has been completed and documented	JSP 936 §69–70	NOVA-BHMA-001
✓	Human control is embedded at every stage: no output is finalised without positive human review and "Mark Complete" action	JSP 936 §56–58	NOVA v1.0 platform
✓	NOVA is classified OFFICIAL only. A classification warning is displayed throughout the platform and in all documentation.	JSP 936 §195	NOVA-SRA-001
⚠️	User Training Package (NOVA-UTP-001) is in development. Interim mitigation: Activity Guide and criteria system provide operator guidance.	JSP 936 §130	NOVA-UTP-001 (pending)

2. Commitments

In support of any MOD TLB Executive Board’s Statement of AI Ethical Assurance to 2PUS, I confirm the following ongoing commitments:

This Statement will be reviewed and renewed annually, or on material change to NOVA’s AI components, scope or risk profile.
All supporting evidence documents (NOVA-AERA-001, NOVA-MC-001, NOVA-SRA-001, NOVA-BHMA-001) will be made available to the contracting TLB’s RAISO on request.
Any material change to NOVA’s AI components (e.g. change of LLM provider or model, change of vector database, change of data residency) will be notified to contracting organisations within 30 days.
NOVA will monitor Anthropic’s published model cards and safety documentation and incorporate relevant updates into NOVA’s bias and harm analysis.
The User Training Package (NOVA-UTP-001) will be completed by Q2 2026 and provided to all contracting organisations.

3. Limitations of This Statement

This statement covers NOVA™ as a platform and tool. It does not constitute assurance for any specific MOD project or programme that uses NOVA as its tooling. Each MOD TLB remains responsible for:

Appointing a RAISO and ensuring their organisation’s AI governance is in place per JSP 936 §100–108.
Conducting project-level AI ethical risk assessments for specific training programmes designed using NOVA.
Ensuring that TDOs using NOVA are suitably qualified and experienced for the DSAT activities they are conducting.
Ensuring that all outputs generated by NOVA are reviewed and approved by a subject matter expert before use.

4. Sign-Off

Name	Will Kennedy-Long FLPI, FITOL
Role	Founder and Responsible Person, NOVA™ Agentic Allied Defence Training Platform
Date	8 March 2026
Signature	(Signed electronically)
Valid Until	March 2027 (subject to earlier review on material change)

Security Risk AssessmentNOVA-SRA-001

SECURITY RISK ASSESSMENT

NOVA™ — Agentic Allied Defence Training Platform

Document Ref: NOVA-SRA-001 | 8 March 2026 | Will Kennedy-Long FLPI, FITOL

JSP 936 Reference: JSP 936 V1.1 §195–197 — A Secure by Design approach is required. Security analyses must include conventional risks and adversarial interference with AI performance. Unique security risks associated with AI development and behaviour must be analysed from the earliest practicable opportunity.

⛔ CLASSIFICATION BOUNDARY — CRITICAL: NOVA is approved for OFFICIAL classification only. Under no circumstances should OFFICIAL-SENSITIVE, SECRET, or TOP SECRET material be uploaded to or processed by NOVA. This is a hard operational constraint and must be communicated to all users.

1. Architecture and Data Flow Overview

NOVA™ uses a distributed architecture across the following components. All data flows are documented below for security assessment purposes.

Component	Provider / Location	Data Processed	Data Residency
Frontend (Pages/D1/KV)	Cloudflare Inc. (US/EU)	User session, activity data, generated documents	Cloudflare EU data centres where available. D1/KV data stored per Cloudflare data localisation settings.
Analysis Backend	Railway Corp (US)	Document text during parsing. Transient — not persisted.	US-based Railway servers. Data transient in memory during processing only.
LLM API	Anthropic PBC (US)	Prompts containing document extracts and user-entered field content	US-based Anthropic infrastructure. API inputs not retained for training per Anthropic API ToS.
Vector Database	Pinecone Inc. (US)	Doctrine content vectors only. No user data in Pinecone.	US-based. Doctrine content only — no user PII or project data.

Data Residency Note: NOVA processes OFFICIAL data through US-based third-party services (Anthropic, Railway, Pinecone). MOD organisations must satisfy themselves that this is acceptable under their own data handling policies before use. Cloud security standards applicable: Anthropic and Cloudflare hold SOC 2 Type II certification.

2. Security Risk Register

ID	Risk	Likelihood	Impact	Mitigation
SR-01	User uploads OFFICIAL-SENSITIVE or SECRET material to NOVA	Low–Moderate	High	Hard classification limit stated throughout platform, in all documentation, and in onboarding. User Training Package (pending) to reinforce. Contractual obligations on MOD customers.
SR-02	Anthropic API data interception in transit	Very Low	Moderate	TLS 1.3 encryption in transit. Anthropic API requires HTTPS. No plaintext transmission.
SR-03	Cloudflare D1/KV data breach	Very Low	Moderate	Cloudflare SOC 2 Type II certified. Data isolated per account. Row-level security on D1.
SR-04	Prompt injection attack via malicious document upload	Low	Low–Moderate	Document content is parsed and structured before LLM submission. Structured prompts reduce injection surface. Claude Sonnet has instruction hierarchy protection.
SR-05	Data poisoning of Pinecone doctrine index	Very Low	Moderate	Pinecone write access controlled by API key held by developer only. Doctrine files are version-controlled. No user can modify doctrine index.
SR-06	Railway backend compromise / data exfiltration	Low	Moderate	Railway processes document content transiently only — no persistent storage of user data. Railway uses isolated container environments per deployment.
SR-07	Account takeover / unauthorised access	Low	Moderate	Cloudflare Access authentication. Accounts approved individually by the platform owner through the Admin Panel. Session tokens with expiry.
SR-08	Third-party API service outage (Anthropic / Pinecone)	Moderate	Low	NOVA degrades gracefully — form fields remain accessible and saveable without AI features. Activity data not lost on API failure.
SR-09	Model behaviour change on Anthropic API update	Moderate	Low–Moderate	Model version pinned in API calls (claude-sonnet-4-6). Anthropic notifies of deprecation. Monitored quarterly.

3. Security Controls Summary

Control Area	Control	Standard / Reference
Data in Transit	TLS 1.3 on all connections (Cloudflare, Anthropic, Railway, Pinecone)	NCSC Cloud Security Principles
Data at Rest	Cloudflare D1/KV encryption at rest.	Cloudflare SOC 2
Access Control	Account-based access. API keys stored as environment variables, not in code.	Secure by Design
Classification Enforcement	OFFICIAL-only hard limit. Warning displayed in platform and all documentation.	JSP 936 §195
Third-Party Assurance	Anthropic SOC 2 / Cloudflare SOC 2 / Railway SOC 2 / Pinecone SOC 2	Supplier security certifications
Incident Response	Data breach notification within 72 hours per UK GDPR Article 33. Contact: Will Kennedy-Long FLPI, FITOL	UK GDPR Art. 33

4. Recommendations for MOD Customers

NOVA should be accessed from MOD-managed devices with current endpoint protection.
MOD organisations should not process OFFICIAL-SENSITIVE project data through NOVA without specific approval from their RAISO.
MOD IT security teams should be notified of NOVA’s data flows (particularly Anthropic API US routing) and clearance obtained per local security policy before use.
User accounts should use government email addresses where possible.
MOD customers should review Anthropic’s Enterprise data handling commitments if processing sensitive OFFICIAL content.

5. Overall Security Risk Rating

Overall Security Risk: LOW to MODERATE. The primary risk is user error (uploading above-OFFICIAL material). Technical controls are strong across all service providers. Data flows are to US-based services — MOD customers must satisfy themselves this is acceptable under their data handling policy. No SECRET-capable architecture is in place or planned.

6. Document Control

Document Owner	Will Kennedy-Long FLPI, FITOL
Version	v1.0
Date	8 March 2026
Next Review	March 2027 or on material change to architecture