Design a campaign workflow with a visible evidence trail
Review and limits. The course content has completed cross-model editorial review. Domain weights, timings and the pass standard remain provisional: no completed job-task analysis or empirical standard-setting study is claimed.
This blueprint extends the shared CAI-3 framework. CAI-2 is recommended, with no admission prerequisite. The common Agent Fundamentals module bridges the gap, and the challenge-only exam path remains available. Course completion does not itself earn a credential.
The whole-task spine is a Clearleaf campaign-workflow dossier. The learner specifies source retrieval and claim checks, builds a prompt chain with human approval boundaries, tests observed assistant behavior and audits operating evidence. Publication and spend authority stay with people. All deliverables are text or Markdown in the AI workspace; no upload, external account, live integration or code-execution environment is required.
Framework alignment
The framework's five domains retain equal 20% weight: D1 workflow architecture; D2 bounded tools and integration; D3 data boundaries and grounded verification; D4 human authority and recovery; D5 testing and operating oversight. The three exam tasks remain equally weighted, each with the same five equal criteria and exact ids in the shared framework. This blueprint does not change an exam definition, timer, cut score or validity claim. Practice uses the four existing exercise types. Module checkpoints target 0.8 mastery with feedback and retakes; practice results do not authorize a credential.
| Module | Workflow contribution | Main domains | Checkpoint added to the dossier | | ------------------------------------- | ------------------------------------------------------ | -------------- | -------------------------------------------------------------------------- | | 1 Agent Fundamentals | Bound capabilities, human authority, data and evidence | D1–D5 | Shared foundation: separate observations, assumptions and proposed actions | | 2 Research and evidence chains | Decompose a brief into sources, claims and handoffs | D1, D2, D3, D4 | Workflow map, field map and source/claim register | | 3 Asset chains and approvals | Draft, validate and bind both human approvals | D1–D5 | Prompts/schemas, actual trials, approval table, change and retest | | 4 Scheduling state and recovery | Separate a schedule proposal from an authorized action | D1, D2, D4, D5 | State/dependency table, action-key contract and recovery plan | | 5 Measurement and operating oversight | Audit quality and reproduce campaign measures | D2, D3, D4, D5 | Metrics memo, event-grounded findings, monitoring and restart rules | | 6 Marketing workflow capstone | Integrate design, build/test and audit evidence | D1–D5 | Complete versioned dossier and operating handover |
Each role module includes a branching scenario, a sandbox prompt-chain trial, a critique and an artifact checkpoint with its own capstone-practice rubric. Three retrieval questions include two earlier concepts. Four independent binary conditions define each criterion, so levels 0–4 mean zero through four conditions demonstrated. Fixed counts and named evidence replace vague completeness judgments. Candidate-authored workspace decisions support process criteria; an assistant's statement cannot establish that the candidate made a choice or that an external tool ran.
Fictional sources and authority
Clearleaf's product sheet v3 lists a starter set costing 22 units with two glass bottles and four tablets, and a refill pack costing 6 units containing ten tablets. Laundry products are unavailable. Other performance, environmental, health and competitor claims have no supplied approval. The course requires no outside product or market research. An absent claim or source remains unavailable.
Operator Ellis prepares internal work; brand owner Avery approves claims; marketing owner Priya approves publication or spend; Rowan owns integration and Jules owns quality. All people, products, cases and channel data are fictional. Both human approvals bind the exact asset/payload version. A change to claims, channel or spend invalidates the relevant approval. Approval expires after 30 minutes; pending, denied and expired states cannot authorize an action.
Written tools read supplied versioned sources, draft an asset, validate claims, propose a schedule and apply an exactly approved schedule. None is callable in this course workspace. A proposal is not publication. A tool success receipt, rather than a conversational completion claim, establishes a completed action. Unknown outcomes go to reconciliation before another write; explicit temporary failures have at most two attempts under a stable action key.
Only supplied aggregate channel data and fictional segment labels are used. Allowed trace fields are campaign id, asset id/version, source id/version, state, approval ids and receipt id. Individual contact and browsing records do not enter prompts or logs. Redacted training traces stay in the restricted campaign-quality store for 14 days, then are deleted.
Practice cases and evidence
Each sandbox provides normal, missing-source and unavailable-tool cases. The learner states expectations, records actual assistant outputs and observations, checks source support, changes a specific weakness and retests. Written tool behavior is a clearly labeled manual check. No exercise claims to publish, change spend or call an outside service.
Critiques give exactly four ordinary operational defects with policy/event ids, plus successful control events. Findings must identify the observed event, the applicable rule, affected scope, containment, correction, owner and safe restart evidence. A draft accepted by a colleague does not establish claim support or publication authority. Monitoring pairs useful asset throughput with unsupported-claim and missing-approval counts, including denominators and review windows. One observed unauthorized action can justify a scoped pause.
The matched-window example uses channel A spend 1,200 units, 60 leads and 12 qualified leads; channel B spend 900 units, 30 leads and 15 qualified leads. The qualified-lead definition is shared. Cost per qualified lead is 100 and 60 units respectively. These observations do not establish causality, profit, statistical certainty or an automatic spend decision. Missing inputs must be reported rather than imputed without a stated basis.
The final practice dossier separates workflow design; written configuration with actual trials, fixes and retests; and oversight audit with containment and operating controls. Different practice cases support those sections.
How results are decided
- A person reviews every CAI-3 result, pass or fail, and makes the decision. The AI score is only a starting point for that reviewer. No result is released automatically.
- Award rule (provisional): an overall mean of at least 2.5 of 4 across the three equally weighted exam tasks, with every task at least 2.0. There is no separate minimum for any single criterion, including recorded AI collaboration.
- The reviewer's decision is final and may differ from the AI score. You can still appeal any result.