Assessment · Exit diagnostic

The same six judgments, harder

Twelve scenarios matched one for one against the entry diagnostic: same six skills, different situations, and every one of them a step further in. Take it after the capstone, and compare.

Time: 30–40 minTake it after the capstone Jump to your route

How the routing works here

Identical rule to the entry diagnostic: six skills in phase order, two questions each, and you are routed to the earliest skill where you did not get both right. The difference is what the route now means. At entry it said “start reading here”. At exit it says “this is the phase whose judgment did not survive a harder version of the same question”, and the remedy is usually practice rather than re-reading — the code lab for Phases I–III, the decision-change simulations for IV.

The honest limits from the entry diagnostic all still apply, and one more joins them: you have now seen this instrument once. Recognizing the shape of these questions is a skill this test cannot distinguish from design judgment. Treat a clean sweep as necessary rather than sufficient, and let the capstone rubric carry the weight.

1 · Find a structural problem

A reporting entry point that has taken on two options.

def build_report(rows: list[Row], options: dict) -> bytes:
    if options["format"] == "pdf":
        renderer = PdfRenderer(options["page_size"])
    else:
        renderer = CsvRenderer(options["delimiter"])
    if options.get("compress"):
        return gzip.compress(renderer.render(rows))
    return renderer.render(rows)

Over eighteen months utils.py has reached 400 lines and appears in 62 of the last 80 pull requests.

2 · Protect a domain invariant

An account object that enforces its own rule.

class Wallet:
    def __init__(self, balance: Money) -> None:
        self._balance = balance

    @property
    def balance(self) -> Money:
        return self._balance

    def withdraw(self, amount: Money) -> None:
        if amount > self._balance:
            raise InsufficientFunds()
        self._balance -= amount

The rule “a customer may have at most three open returns” is enforced inside the ReturnRequest aggregate, by counting the customer’s open returns through a repository query.

3 · Identify a leaking boundary

A repository method that streams results.

class SqlOrderRepository:
    def find_open(self, customer_id: str):
        with self._session.begin():
            for row in self._session.stream(select(OrderRow)):
                yield Order.from_row(row)

The domain publishes LabelRequested(order_id, carrier_response_code=202).

4 · Compare an architectural trade-off

Team A owns pricing. Team B needs a new price component. Either B sends a pull request to A’s module and waits for A to review and release, or A exposes a plugin point that B configures.

A report takes 4.2 s, of which about 90% is one N+1 query. The proposal is a materialized view that takes it to 0.3 s and adds a refresh job. Three analysts run the report about twelve times a day.

5 · Turn a vague requirement into a quality-attribute scenario

A driver reads: “the eligibility endpoint responds in under 300 ms.”

Two drivers conflict: a decision returns within 800 ms at p95, and no return is accepted without a fraud check, which takes 2 s at p99.9.

6 · Replan an initiative after evidence changes

Four weeks into the pilot: median resolution is 3.1 h against a target under 4, and support contacts are down 31% against a target of 25%. The manual review queue is averaging 68 a day against a guardrail of 50, and fraud loss is up 0.4 points.

The order team says its wave-schedule API will be six weeks late. Your delivery map has a fallback — a 15-minute-stale dispatch export — with a trigger date of week 3. It is week 5.

Your route

Answer the twelve questions above and your route appears here.

Entry against exit

The scores alone will not tell you much — twelve questions is a small sample and the two sets are not calibrated against each other. What is worth writing down is the second and third columns. Fill this in from memory of taking the entry diagnostic and from what you have just done; if you cannot remember, that is itself a finding about how far back it was.

The fourth column is the one that matters. A diagnostic score is a claim about a Tuesday afternoon; a superseded ADR with its predecessor intact is evidence. If a row’s evidence column is empty, that skill has been read about and not yet practiced, whatever the questions said.

After the exit diagnostic

There is nothing left to score. What is left is the practice that makes any of it durable:

• Use the rubric on somebody else’s work. Scoring a packet you did not write is the fastest calibration available, and the peer-review protocol in the rubric is built for exactly one session.
• Re-run the decision-change simulations with a different first commitment. They have no predetermined answer, so they stay useful after the questions stop being new: eight simulations.
• Take one artifact into work this month. One driver written as a scenario, one ADR with a revisit trigger, one import rule with a test. The playbook holds the blanks.