The Responsible AI Readiness Framework

You cannot mitigate a harm you have never named.

This dimension is first on purpose. Every other dimension is a control, and a control with no stated risk behind it is a guess about what mattered. NIST puts MAP before MEASURE for the same reason, and the AI Act puts the risk management system in Article 9 — ahead of the requirements it governs.

• A documented risk assessment names who could be affected, how, and how likely — kept current as the system changes, not written once at kickoff.
• Intended purpose and out-of-scope uses are written down, and behaviour outside them is defined: refuse, escalate or degrade, never undefined.
• A human can review, override or halt a consequential decision, and that intervention is recorded.
• The people relying on the output are told what the system can and cannot do, and how confident it is — in the interface, not in a contract annex.
• Where the system has a legal or similarly significant effect on a person, the lawful basis and the route to human review exist and are testable.

Operationalises AI Act Art. 9 (risk management system) · Art. 13 (information to deployers) · Art. 14 (human oversight) · Art. 27 (fundamental rights impact assessment) · Art. 50 (transparency obligations) · GDPR Art. 22 (automated individual decision-making) · Art. 35 (data protection impact assessment) · NIST AI RMF MAP · GOVERN · ISO/IEC 23894 · 42005

An AI decision is only as defensible as the data behind it.

• Schema contracts on every ingestion boundary (types, ranges, required fields), enforced — not assumed.
• Input validation rejects or quarantines bad records instead of silently passing them downstream.
• End-to-end lineage: any model output can be traced back to the source data and transformations that produced it.
• Train/serve feature parity is enforced and proven by test — no training-serving skew.
• Sensitive fields are classified and handled at the data layer, and special-category data is separated from ordinary personal data.
• Datasets are examined for bias relevant to the intended purpose, and what was found is recorded.

Operationalises AI Act Art. 10 (data and data governance, incl. 10(2)(f)–(g) on bias examination) · GDPR Art. 5 (accuracy, minimisation, purpose limitation) · Art. 9 (special categories) · NIST AI RMF MAP · MEASURE · ISO/IEC 42001

The system must refuse to produce unsafe or ungrounded output.

The heaviest weight in the framework, because it is the only dimension whose failure is visible to a user in the moment it happens.

• Input guardrails: prompt-injection, jailbreak, and out-of-scope request handling.
• Output guardrails: PII redaction, toxicity and relevance checks, and response grounding against an approved source.
• An evaluation harness scores outputs against a labelled set before any release — measuring both what it blocks and what it wrongly blocks.
• The adversarial set is versioned, grows when a new attack is found, and failing it blocks the release.
• LLM judgement is scoped to where it adds value over deterministic logic — and bounded everywhere else.

Operationalises AI Act Art. 15 (accuracy, robustness and cybersecurity, incl. 15(5) on manipulation) · Art. 9 (risk mitigation measures) · GDPR Art. 32 (security of processing) · NIST AI RMF MEASURE · MANAGE · AI 600-1 · ISO/IEC 23894

You cannot be accountable for a system you cannot see.

• Structured logging of inputs, outputs and decisions — with PII handled — for every inference, retained long enough to answer a question about a past decision.
• Dashboards for latency, error rate, cost and volume: live, not on request.
• Data drift and model/output-quality drift are monitored, with thresholds.
• Alerting fires on degradation before a user reports it, and routes to a named owner.
• Each decision is traceable end to end: request → features → model → guardrails → output.
• Real-world performance is collected and reviewed after release, not only before it.

Operationalises AI Act Art. 12 (record-keeping) · Art. 19 (automatically generated logs) · Art. 72 (post-market monitoring) · Art. 73 (serious incident reporting) · GDPR Art. 5(2) (accountability) · Art. 33 (breach notification) · NIST AI RMF MEASURE · MANAGE · ISO/IEC 42001

Accountability must be provable on demand, not promised in a PDF.

• Access control and data grants defined declaratively and version-controlled.
• Audit trail: who, what and when for data access, model changes and deployments — queryable.
• Technical documentation generated from the system and kept in sync with the code, covering what Annex IV asks for.
• Model and dataset cards: intended use, limitations, known risks, evaluation results — versioned.
• A model-promotion gate enforces quality and safety thresholds before anything reaches production.
• Registration and record-keeping obligations are tracked as work items with owners, not discovered at audit.

Operationalises AI Act Art. 11 + Annex IV (technical documentation — simplified form available to SMEs under Reg. 2026/1744) · Art. 17 (quality management system) · Art. 18 (documentation keeping) · Art. 49 (registration) · Art. 71 (EU database) · GDPR Art. 25 (data protection by design and by default) · Art. 30 (records of processing) · NIST AI RMF GOVERN · ISO/IEC 42001

Behaviour you can’t reproduce, you can’t trust.

• Core logic separated from cloud and framework execution, covered by a credential-free local test suite.
• Tests are CI-gated: nothing merges or ships on red.
• All infrastructure is code — no console-click deployments.
• Versions are pinned and builds are reproducible across machines.
• Behaviour-preserving refactors are verified by test, not assumed.

Operationalises AI Act Art. 15 (accuracy and robustness — the evidence behind the claim) · Art. 17 (quality management system: verification and testing procedures) · GDPR Art. 25 (by design) · Art. 32 (security of processing) · NIST AI RMF MEASURE · ISO/IEC 42001

A trustworthy system recovers without heroics.

• Idempotency: re-running a job produces the same result, not duplicates or corruption.
• Checkpointing and exactly-once semantics where state matters.
• Automated remediation for known failure classes — bounded, and fail-closed.
• Isolated state and minimised blast radius, so failures do not cascade.
• Guarded destructive operations to prevent accidental teardown.

Operationalises AI Act Art. 15(4) (resilience regarding errors, faults and inconsistencies; fail-safe plans) · GDPR Art. 32(1)(c) (ability to restore availability and access in a timely manner) · NIST AI RMF MANAGE · ISO/IEC 23894