Level 400 credential route (draft): AiGovOps Foundation Open Mastery in AI Engineering and Assurance (free)
Draft route for review. This page describes Level 400 project marking and proposes a course credential route. It does not award or promise a credential. See the credential specification.
Anything marked Proposal is not decided yet. Bob Rapp and Ken Johnston, the review board co-chairs, decide it unless the page names Bob alone.
What you must do
Read the lessons and complete each module's quiz and workbook. For objectives 400.P-a and 400.P-b, plan, simulate and verify an adaptation, exercise the release gate, rehearse a rollback and make a monitoring plan. Assemble an evidence bundle so a stranger can verify it offline and write a bounded statement.
Proposal, needs the review board co-chairs’ (Bob and Ken) decision: make the quizzes and workbooks practice only, not a credential requirement. The current credential specification calls for completing them.
The rubric: how your project is marked
Every row must be true.
| Check | Meets when |
|---|---|
| Comparison | Three rows use the same held-out set, each with an interval; the two closest have a paired difference; at least one cohort row names its count and interval; the release rule and exact cohort guard are dated before the run |
| Release gate | Names an approver who neither built the model nor holds the Level 300 card; shows a denied run with a receipt naming the missing evidence |
| Rollback | A dry run restores prior model, data, code and configuration versions and records what was checked |
| Monitoring | Builds on the Level 300 hold card; names input shift, late-label performance, a cohort row, versions and a threshold that pages the named holder |
| Bundle | Contains the chain, policy and hash, crosswalk, and a VERIFY.md with file list and receipt schema; has no prompts, outputs or personal data; people appear only as roles and references, and carried Level 300 items have provenance labels |
| Verification | A stranger checks offline using a separately received key and checkpoint and signs one bounded sentence, with name and date |
A worked example
Priya's fictional Big Widget comparison uses the sandbox's held-out results. A fine-tune that loses or a release that stays held can still pass. A missing uncertainty interval or a gate that never said no cannot.
How the project is marked
Two Level 400 practitioners or a host mark the project. They check each row and mark prepared results, paper exercises and actual runs for what they are. The sandbox adaptation is planned, simulated with prepared answers and verified by scoring; it is not real model training. Hosting a Thursday call is optional and is not a condition of the mark.
If it is "not yet": revise and resubmit
Supply missing evidence or narrow an overbroad claim. Have a stranger verify the bundle that is actually being marked. No separate credential resubmission or appeals process is established here.
What the certificate would say, and what it would not
The proposed claim is limited to a person marking your practice against this rubric. It would not rate a system, approve a release, or confer a degree, professional certification or vendor certification.
The certificate would carry the lines "Free and open. Every source cited. Human-reviewed." and "Earned, never bought."