W12 · Capstone defense and competency review

FIELD JOURNAL / 04Acceptance and reuse

See the structure, then implement and verify with the notes.

This page retains the English overview and workshop. The expanded practitioner lecture notes are currently available in Chinese. Read the complete Chinese lecture →

12.1

Trace each capstone claim to its evidence

Trace a completion claim to versions, actual traces, tests and inputs. Screenshots illustrate; they do not replace reproducibility.

THIS WEEK IN THE FIELD

What situation are you walking into?

The committee wants more than a successful demo: what if data is wrong, why not rules, what did you build, and what happens if business impact is inconclusive?

DELIVERABLE / Capstone pack + defense materials + growth record

01 / Hand the project to another person

Document setup, installation, data, run commands, tests and limitations for a fresh environment. Provide example config without secrets and synthetic data without customer records. Demonstrate normal work, missing evidence, denied access and recovery. Video supplements reproducible evidence.

Label the portfolio as a training or sandbox project, not a real customer production deployment.

02 / Defend why, not only what

Use fifteen minutes for problem, baseline, decisions, boundaries, demonstration, results, limits and reuse, followed by fifteen minutes of challenge. Link each material claim to evidence. A technology list cannot replace a rationale; more agents do not automatically improve a system.

For unknowns, explain the evidence gap and validation plan rather than inventing an answer.

03 / Assess competence, not certificate theater

Assess diagnosis, semantics, engineering, evaluation and safety, value evidence, and collaboration and reuse. Treat secrets, fabricated evidence, unauthorized actions, non-runnable work or mislabeled simulation as blockers. Site self-checks are learning feedback, not accreditation or employment guarantees.

Engineering completion needs model-driven task traces; the offline kit proves only its specific engineering mechanisms.

04 / Make growth and portfolio claims verifiable

Repeat the initial self-assessment and support changes with artifacts. Document context, your contribution, decisions, evidence, results and failures. Practice coding, system design, decomposition and customer communication without presenting a universal interview sequence.

Create a thirty-day plan: one capability gap, one new domain and one reproducible validation.

WORKED EXAMPLE · SYNTHETIC

Put the method to work

Credible portfolio wording: implemented identity mapping, metric contracts and controlled tools on synthetic retail data with scope, duplicate and recovery tests; planning is simulated and model-driven traces remain to be added. This is stronger than fabricating a customer growth story.

Practice this week

Submit prior deliverables, final code and an evidence index.
Demonstrate normal work, failures and a new constraint.
Complete reassessment, portfolio description and a thirty-day plan.

Acceptance checks

Reproducible in a fresh environment, with honest scope.
Decisions trace to evidence.
Failures can be diagnosed, stopped or recovered.

Download editable worksheet Open the labs

FIELD WORKSHOP / W12

120 minutes: defend delivery under an unseen constraint

A defense is more than a recording. Reviewers introduce late data, a timeout, reduced permissions or a changed definition. Explain the impact before demonstrating reproducible evidence and a safe stop.

0–30 min: freeze the portfolio version and start from a clean environment. Link each capability claim to a file, test, trace or review.
30–90 min: run two 30-minute defenses: problem (5), working slice (10), failure (10), value (5). Rotate builder and reviewer roles.
90–120 min: reassess on a comparable unseen task. Record personal contribution, collaboration needs and the next real validation.

Artifact excerpt · synthetic example

Portfolio index
01-fit/: value, boundary, unknowns
02-data/: synthetic/authorized data, contracts, recalculation
03-system/: tools, model integration, versions
04-evals/: failures, regression, actual traces or simulation labels
05-ops/: runbook, costs, stop and recovery
06-value/: acceptance, attribution limits, next action
07-reuse/: rights, reusable assets, transfer tests

Instructor review

Unimplemented work can be acknowledged; ambiguous implementation claims cannot. Review role-appropriate evidence. Reading checkmarks are not assessment or external accreditation.

An unseen field challenge ↗

Answer: what business judgment changed, how does the system bound harmful action, and what lets the next owner continue operating it?

Download offline practice kit ↓

12.2

State what the portfolio actually demonstrates

A design, deterministic simulation, model integration and production operation require different evidence. Label the level actually reached.

ONE-MINUTE CHECK

All offline simulator tests pass. Has real model capability been validated?

I have read this module (not an assessment result)

My study notes · stored in this browser only

Local storage is not encrypted; do not store sensitive content on shared devices.

References and further reading

Anthropic ↗ · Demystifying evals for AI agents2026-01-09 · Anthropic Anthropic ↗ · Building effective agents2024-12-19 · Anthropic

Read alongside the assignment and record design constraints and tests. Verify specifications against your deployment version.