Teach around artifacts.
Assess through evidence.
For facilitators, FDE leads and independent learners: prepare samples, review code and failures, assess six dimensions, and plan revisions.
Four preparations before starting
- Choose a business thread and sandbox or authorized pilot.
- Assign domain, engineering and delivery roles explicitly.
- Capture initial self-assessment and a baseline task artifact.
- Run the kit and prepare offline failures without real customer data.
Revisit the decision
Use the prior artifact to explain change.
Method and counterexample
Show one example and one failure.
Practice and peer review
Use new samples; identify evidence gaps.
Review and failure challenge
Inspect code, outcomes and traces.
Revision commitment
Assign revision, ownership and recheck.
Another person must be able to inspect the work
Use the synthetic retail sandbox or adapt it to prospecting, marketing or knowledge service. Real projects need separate authorization.
Required evidence pack
- W1–W11 artifacts, code, example config and known limits.
- Normal, insufficient-evidence, access-denial, duplicate and recovery evidence.
- Actual model participation and traces; simulations labeled.
- Three-layer acceptance, assumptions, unmet items and asset layers.
Suggested defense · 30 minutes
- 3 min: problem, baseline and selection rationale.
- 4 min: semantic and architectural decisions and boundaries.
- 5 min: normal task and a failure.
- 3 min: impact, limits, reuse and contribution.
- 15 min: committee questions and new constraints.
Six dimensions. Separate critical blockers.
Suggested rubric: total ≥75, each 20-point dimension ≥12, each 10-point dimension ≥6, and no critical blockers for internal completion review.
Verifiable job, value, scope and ownership.
Reconciled data, reproducible metrics, identity safeguards.
Contracts, idempotency, recovery and reproduction.
Normal, edge, adversarial and separate safety gates.
Three-layer evidence, limits, costs and responsibilities.
Decision records, assets, transfer and source rights.
| Level | Evidence | Review action |
|---|---|---|
| Not established | Only concepts or screenshots; cannot explain inputs and outcomes. | Build a minimal verifiable example. |
| Emerging | Happy path works; failures and ownership incomplete. | Assign counterexamples and reassess. |
| Meets standard | Complete evidence, clear bounds, reproducible by others. | Defend and test on new samples. |
| Transferable | Explains second-scenario adaptations and limits. | Capture templates and domain differences. |
What should the weekly review ask?
FDE role and project qualification
High value, but sample access is unauthorized. What next?
Discovery and business modeling
How should “sales thinks leads are poor” be recorded?
Metrics and delivery scope
Does higher adoption prove AI caused revenue growth?
Data profiling and tenant adaptation
HTTP 200, but tax treatment in revenue is unknown. Accept?
Ontology and the advisory agent
Does a working MCP connection authorize all customer data?
Data security and deployment design
Does private deployment automatically ensure compliance?
Agent evaluation and scope change
Aggregate score improves but an unauthorized write appears. Ship?
Action agents and operational integration
After an execution timeout, what comes first?
Production engineering and launch review
Are total token counts enough to explain task failures?
Attribution and project acceptance
Treatment gains 8 points; comparison gains 5. What is DiD?
Retrospective and product feedback
Are renamed customer-A strategy thresholds a public product asset?
Capstone defense and competency review
All offline simulator tests pass. Has real model capability been validated?