Joint Service Publication 936 — MOD Policy for AI — distilled to 5 fillable sections.
Owner: CSOAI LTD (16939677) · Status: v0.1 GENERATOR LIVE · Last update: 10 Jul 2026
5
Sections
48
Questions
12
Artifact Types
100%
UK Sovereign
1. The 5 JSP 936 Sections
JSP 936 is the MOD's binding policy for trustworthy AI. DEFONEOS collapses the 80-page document into 5 fillable sections + 48 questions. Each answer generates a SIGIL-signed artifact the buyer can file at audit.
#
Section
What MOD Audits
DEFONEOS Generator
Artifact Type
1
Dependability
Is the AI reliable, repeatable, safe under load?
12 questions on data lineage, model card, test coverage
Dependability-Statement-v1.0.pdf
2
Risk Assessment
What can fail? How badly? What's the mitigation?
10 questions on harm, bias, adversarial, single-points-of-failure
Risk-Register-v1.0.xlsx
3
Governance
Who decides? Who can override? Who audits?
8 questions on human-in-loop, escalation, audit trail
Governance-Charter-v1.0.md
4
Lifecycle
How is it built, deployed, retired, replaced?
10 questions on CI/CD, retraining, sunset criteria
Lifecycle-Plan-v1.0.md
5
Assurance
How is it tested? What's the ongoing monitoring?
8 questions on red-team, monitoring, KPIs
Assurance-Plan-v1.0.md
2. Section 1 — Dependability
12 Questions
Q1.1 — What data sources train the model? (UK-only / EU-only / Global)
Q1.2 — Is the training data versioned + hash-pinned?
Q1.3 — Does the model card exist + SIGIL-signed?
Q1.4 — Test coverage: precision / recall / F1 on held-out test set?
Q1.5 — What is the latency SLO under p99 load?
Q1.6 — Single-point-of-failure analysis: how many region failures before service degrades?
Q1.7 — Reproducibility: same seed + same data = same output bit-for-bit?
Q1.8 — Adversarial input testing: any robustness tests passed?
Q1.9 — Calibration: does model confidence correlate with accuracy?
Q1.10 — Drift detection: how is input/output drift monitored?
Q1.11 — Backwards compatibility: does v2 break v1 contracts?
Q1.12 — Disaster recovery: RTO + RPO documented?
3. Section 2 — Risk Assessment
10 Questions + 4×5 Risk Matrix
Q2.1 — Worst-case harm if model fails (catastrophic / major / moderate / minor / negligible)
Q2.2 — Probability of failure (rare / unlikely / possible / likely / almost certain)
Q2.3 — Affected population size (1 user / 100 / 10k / 100k / 1M+)
Q2.4 — Demographic bias: any groups disproportionately affected?
Q2.5 — Adversarial surface: prompt injection / model theft / data poisoning
Q2.6 — Cascade risk: does failure of this model break other models?
Q2.7 — Single-vendor dependency: what's the exit cost?
Q2.8 — Insider threat: who can rewrite the model / training data?
Q2.9 — Supply chain: any non-sovereign dependencies in the stack?
Q2.10 — Irreversibility: can the harm be undone?
4. Section 3 — Governance
8 Questions
Q3.1 — Named human accountable for the AI (not "the team")
Q3.2 — Escalation path: who gets paged at 02:00?
Q3.3 — Override authority: who can shut it down, and how fast?
Q3.4 — Audit trail: every decision SIGIL-signed and exportable
Q3.5 — Stakeholder consultation: who's been briefed + when?
Q3.6 — Change-control: how are model / data updates approved?
Q5.8 — Continuous improvement: feed-back loop from incidents to training
7. The 12 Artifacts This Generator Produces
#
Artifact
Section
Format
SIGIL?
1
Dependability-Statement-v1.0
§1
PDF
✅
2
Risk-Register-v1.0
§2
XLSX
✅
3
Governance-Charter-v1.0
§3
MD
✅
4
Lifecycle-Plan-v1.0
§4
MD
✅
5
Assurance-Plan-v1.0
§5
MD
✅
6
System-Card-v1.0
all
HTML
✅
7
Data-Lineage-Diagram-v1.0
§1
SVG
✅
8
Threat-Model-v1.0
§2
MD
✅
9
Escalation-Matrix-v1.0
§3
PDF
✅
10
CI-CD-Pipeline-Diagram-v1.0
§4
SVG
✅
11
Monitoring-Dashboard-v1.0
§5
HTML
✅
12
JSP-936-Compliance-Receipt-v1.0
all
JSON
✅
8. The Honest Limitation
JSP 936 is policy, not law. The DEFONEOS generator produces the artifacts + the SIGIL chain an auditor needs, but a human owner (Nick) must attest each section. JEEVES can fill, sign, file, and surface — but the named accountable person is human. Always.