AI Governance and Responsible AI ================================ **Why Smarter.** Most discussion of AI governance stays at the level of principles, because students have no system on which to try them. In Smarter, governance is concrete, and every control is a short YAML manifest that a policy, ethics, law or business student can read, change and apply without writing any code: - **Guardrails that inspect every message.** Smarter ships with `eighteen built-in guardrails `__, for personal data, secrets, prompt injection, jailbreaks, moderation, self-harm, off-topic requests, regulated advice, invented citations and privacy requests. Each one states its stage (the user's message, the model's reply, or both), its strategy (keywords, regular expressions, detectors, semantic similarity, a moderation model or an LLM judge), its action (log, flag, redact, transform, block, or escalate to a person) and its mode: enforce, or monitor, which records what it would have done. The `pipeline `__ that runs them is software, not an instruction that a model may ignore. - **Budgets that stop spending.** A :doc:`Budget ` limits dollars or tokens per hour, day, week or month, for an account, a user, an application, a provider or a self-hosted model, and blocks or warns at the limit. The `seven example budgets `__ include a monthly allowance for each student and a daily limit on one provider. - **Access that follows ownership.** Every resource belongs to an :doc:`account and a user `, and every query is `filtered by owner `__. Provider keys and passwords are encrypted :doc:`Secrets `, never shown to users or models. - **An audit trail that traces each answer.** Every guardrail intervention is a `guardrail event `__ with its severity, confidence, rationale and review status; every prompt, tool call and plugin selection is recorded with the person and account that asked; and the :doc:`journal ` records every change to a resource. - **Data that stays in the network.** With :doc:`LLMHost `, open models run on the organization's own cluster, so a class can compare the governance of a commercial model with that of a self-hosted one. Students can see how AI is controlled in practice, change those controls themselves, and find the parts of a policy that no software can enforce. **What the course covers.** How AI systems fail, from invented citations to runaway costs; the NIST AI Risk Management Framework, ISO/IEC 42001, the OECD AI Principles and the EU AI Act; mapping risks for a real application; why a rule in a system prompt is a request while a guardrail is a control; the guardrail toolbox, and measuring controls by their false positives and false negatives; budgets and cost accounting as controls; access, secrets, personal data and data residency; audit, retention and the privacy of the audit trail itself; fairness and explainability; human oversight and incident response; and vendor and model risk. **Look at.** - `guardrail/data/guardrails/ `__, the built-in guardrails. - `account/data/example-manifests/budgets/ `__, the example budgets. - The `customer_support `__ LLMClient, a governed application with four input guardrails. Each course plan lists every manifest and file that it uses under *Smarter in the Source Code*. **Read.** :doc:`/smarter-resources/smarter-guardrail`, :doc:`/smarter-resources/smarter-budget`, :doc:`/smarter-resources/smarter-prompt`, :doc:`/smarter-platform/cost-accounting`, :doc:`/smarter-platform/smarter-journal`, :doc:`/smarter-platform/security`. **Exercises.** - Write an acceptable use policy for AI in a fictional organization, then implement as much of it as possible as guardrail and budget manifests. Discuss which parts of the policy cannot be enforced by software, and why. - Compare a rule written in a system prompt with the same rule enforced by a guardrail. Try to get around each one, and explain the difference. - Run a new guardrail in monitor mode, read the events it records, and decide whether it is ready to enforce. - Pick one answer from the journal and trace it back through the guardrail checks, tool calls and model that produced it. Discuss what an auditor or a regulator would want to see. **Course plans.** Three complete, ready-to-teach plans, each with a schedule, graded assignments and rubrics, two capstone projects, recommended texts, further study and technical prerequisites. See :doc:`course-formats` for what the formats have in common. .. toctree:: :maxdepth: 1 Classroom bootcamp: 8 hours Two-week online course: 6 hours Four-week online course: 12 hours See the :doc:`Smarter Instructor Guide ` to install Smarter and set up a class.