Design a workflow another operator can use
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. Agent Fundamentals supplies the common bridge, and the challenge-only path remains available. Course completion does not itself earn a credential.
The whole-task spine develops a versioned internal-policy register, grounded request-steps response chain, proposed request-state change and people-workflow operating dossier. All work stays text or Markdown in the AI workspace. Tools are written contracts and manual dry-runs; actual assistant responses are separately observed. No external account, upload or live integration is required.
Framework alignment
The 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 families remain equally weighted with their shared criterion ids. These are course-practice artifacts, not new exam definitions, timers, cut scores or empirical validity claims.
| Module | Main domains | Dossier contribution | | -------------------------------------------- | -------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | 1 Agent Fundamentals | D1–D5 | Shared foundation: bounds, authority, source evidence, observed trials and recovery | | 2 Design policy evidence boundaries | D1, D2, D3, D4 | workflow, policy/decision-source register and minimum-field map. No statutory or individual decision is inferred | | 3 Build policy response chains | D1–D5 | prompts, schemas, actual assistant trials, revision/retest and authority state table; no entitlement or exception decision by the assistant | | 4 Integrate request state and human handoffs | D1, D2, D4, D5 | transition/dependency table and recovery/human handoff playbook. No personal information or actual personnel-record write | | 5 Oversee people workflow quality | D2, D3, D4, D5 | quality register, four-defect audit and monitoring plan | | 6 Assemble and test the people workflow | D1–D5 | Integrate design, build/test and oversight audit sections plus operating handover spanning five domains. H601 policy communication from POL-v4/CHANGE-v1, H602 approved decision version absent, H603 propose_request_update unavailable. Keep exceptions and consequential decisions with Mara/Dev, text-only practice |
Every role module has four lessons: a scored branching scenario, sandbox chain trials, an event-grounded critique and a dossier artifact checkpoint. Checkpoints contain three retrieval questions, with two revisiting earlier concepts, and a module-final capstone-practice rubric. Practice targets 0.8 mastery with corrective feedback and retakes. Each criterion has four independent binary conditions; levels 0–4 equal the number met. Practice scores do not authorize credentials.
Fictional operating context
Larkfield Home Services has about 420 staff across three fictional US states. POL-v4 is a fictional internal leave-request policy; CHANGE-v1 is an approved request-steps brief. Neither is a statement of law or statutory entitlement. A real leave or accommodation request can carry statutory rights that vary by jurisdiction and do not depend on internal balances, so a real workflow routes such requests to qualified HR or legal review and never uses a balance check to refuse them. A balance check and human approval are required; missing policy/balance data stays unresolved. Mara approves exact communications and request-record proposals; Dev decides policy exceptions and changes. Bea owns integration repair and Noor owns quality review. QH-v1 has ten reviews: seven grounded/correctly routed, two wrong-version drafts, one missed handover, in mutually exclusive categories. All ten form the supplied rate denominator. No individual or fairness inference follows. Shared trace fields are request id, policy id/version, category, state, proposal/approval/receipt ids; redacted training traces stay 14 days in the restricted people-quality store, then are deleted. Names, medical/payroll/benefit details and individual survey answers stay out.
Every exact approval binds the current payload and relevant source versions, expires after 30 minutes and is invalidated by relevant changes. Pending, denied or expired approvals cannot authorize an action. The immutable proposal retains its expected state/version, and the application must compare it atomically again before writing. Only a successful receipt tied to the original action and exact payload establishes completion; matching current state and an assistant claim are insufficient. Explicit temporary failures allow at most two total attempts with the same action key and integration-enforced deduplication. Unknown outcomes block new writes pending receipt/state reconciliation.
Cases, evidence and feedback
Each sandbox supplies independent normal, missing-input and unavailable-tool fixtures. The learner records expectations before each actual assistant trial, checks its response against source/policy, gives a verdict, makes one specific revision and records the relevant retest. Written integration and manual arithmetic checks remain explicitly labeled. Worked prose contains fictional examples; those examples are not a learner's observed practice evidence.
Each critique supplies a separate restricted audit view with exactly four operational defects, policy/event ids and successful controls. The learner cites observation, identifies scoped impact and uncertainty, proposes corrections, retains minimal redacted evidence and names the authority and retest needed for restart. The records concern ordinary source, approval, completion and trace-handling failures. No outside access or live action is required. Restricted approval payloads and shared quality traces have different allowed fields.
Measures state numerator, denominator, window, threshold, owner and response. Useful output is paired with unsupported-output, missed-handover, invalid-input or unapproved-action signals as applicable. One forbidden consequential action can pause its affected path; throughput alone cannot justify restart.
The final practice dossier has three separately identifiable sections: workflow design; configuration plus observed trials, fix and retest; oversight audit plus operating handover. Different cases support design, trials and audit.
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.