0Pricing
Claude Architect · レッスン

シナリオ問題の採点方法

100～1000に換算され、720点で合格し、推測による減点はありません

「シナリオ問題の採点方法」はCoddyKit上の無料Claude Architectレッスンです。 これはレッスン1/4です。 下記で完全なレッスンを無料で読むことができます。その後、ブラウザ内の組み込みコードエディタと24時間対応のAIチューターでハンズオン演習できます。 これはClaude Architect学習パスの一部であり、ウェブとCoddyKitアプリ全体で進捗が同期されます。 Claude Architectコースには全4レッスンが含まれています。

このレッスンの一部はまだ翻訳されておらず、英語で表示されています。

What "Scoring" Really Measures

Every question on the Claude Certified Architect exam is a scenario-based multiple-choice item: 4 options, exactly 1 correct. Scoring is the process that turns your pattern of right and wrong answers into a single number the certification body can pass or fail you on.

This lesson is about that machinery: how raw answers become a scaled score, what the 720 threshold means, and why the no-guessing-penalty rule changes how you should behave on every single item.

Raw Score vs. Scaled Score

Your raw score is simply how many items you answered correctly. The exam does not report that number to you. Instead it converts your raw performance onto a fixed scaled range of 100 to 1000.

Scaling exists so that different exam forms — which may draw different scenarios of slightly different difficulty — remain comparable. A 720 on an easier form and a 720 on a harder form represent the same demonstrated competence. You never compute the scale yourself; you just need to clear the line.

# Conceptual model: raw correctness -> fixed scaled band
SCALE_MIN, SCALE_MAX = 100, 1000
PASS = 720

# Two forms, different raw difficulty, same scaled bar
for form, scaled in [("easy form", 725), ("hard form", 718)]:
    print(form, "->", "PASS" if scaled >= PASS else "FAIL")

The 720 Pass Line

The pass mark is 720 on the 100-1000 scale — not 72%. Because the score is scaled rather than a raw percentage, you cannot reliably translate "720" into "answered N of M correctly."

The practical consequence: aim for a comfortable margin above 720. Treating 720 as your target leaves you one or two unlucky items away from failing. Architects who pass consistently are operating well clear of the line, not flirting with it.

PASS = 720

def result(scaled_score: int) -> str:
    margin = scaled_score - PASS
    status = "PASS" if margin >= 0 else "FAIL"
    return f"{status} (margin {margin:+d})"

print(result(745))  # comfortable
print(result(721))  # razor-thin
print(result(715))  # just under

Each Item Is Scored Independently

Items are scored independently: each question contributes on its own. There is no partial credit within a 4-option item — you either selected the single correct option or you didn't.

Crucially, 4 of the 8 scenarios are shown on any given sitting, and every question on those four scenarios is scored. None of them is a throwaway. There is no "only the hardest count" rule and no dropped-lowest-item mechanic to rely on.

No Penalty for Guessing

This is the single most important scoring rule for your behavior: a wrong answer and a blank both score zero. Wrong answers are not subtracted from your raw score.

That makes leaving an item blank strictly inferior to answering it. A blank guarantees zero points; any answer — even an uninformed one — has a non-zero chance of being correct. The rule is simple and absolute: answer every question.

# Why a blank is never optimal under no-penalty scoring
blank_ev = 0.0                 # blank: guaranteed zero
blind_guess_ev = 1 / 4         # 4 options, 1 correct
print("Blank EV:", blank_ev)
print("Blind guess EV:", blind_guess_ev)
print("Always answer:", blind_guess_ev > blank_ev)

Elimination Multiplies Your Odds

Because there's no penalty, every option you can confidently rule out raises the expected value of guessing. A blind guess on four options is worth 1/4; eliminate one and it's 1/3; eliminate two and you're at 1/2.

On this exam the easiest options to eliminate are the anti-patterns — parsing model text for words like "done" to end a loop, using iteration caps as the primary stop mechanism, enforcing critical business rules with prompts instead of hooks, or requiring schema fields that may be absent. Spot one and you've usually found a wrong answer.

# Expected value rises as you eliminate distractors
def guess_ev(remaining_options: int) -> float:
    return 1 / remaining_options

for remaining in (4, 3, 2):
    print(f"{remaining} options left -> EV {guess_ev(remaining):.2f}")

Scoring Reaches Across All 5 Domains

The scored items span the five weighted domains, and the weights describe how the exam is composed:

  • D1 Agent Architecture & Orchestration — 27%
  • D2 Tool Design & MCP — 18%
  • D3 Claude Code Config & Workflows — 20%
  • D4 Prompt Engineering & Structured Output — 20%
  • D5 Context Management & Reliability — 15%

D1 carries the heaviest weight, so points there move your scaled score the most. But a weak domain you skip is points permanently lost — there's no domain you can safely ignore.

DOMAIN_WEIGHTS = {
    "D1 Agent Architecture & Orchestration": 27,
    "D2 Tool Design & MCP": 18,
    "D3 Claude Code Config & Workflows": 20,
    "D4 Prompt Engineering & Structured Output": 20,
    "D5 Context Management & Reliability": 15,
}
assert sum(DOMAIN_WEIGHTS.values()) == 100
print("Heaviest:", max(DOMAIN_WEIGHTS, key=DOMAIN_WEIGHTS.get))

Only the Single Best Option Scores

Many scenario items have more than one technically valid-looking option. Scoring credits only the single best answer for the stated context — the one that reflects sound Claude architecture for that exact situation.

That's why surface plausibility isn't enough. "It could work" loses to "it's the correct production practice here." The exam is testing the applied judgment of an architect with 6+ months of production Claude experience, not your ability to recognize a defensible-sounding option.

How a Scenario Question Tests You

A scored item drops you into a real situation and asks what to do next. The correct option maps to a fact-sheet principle; the distractors map to anti-patterns. Example shape: an agentic loop returns stop_reason == "tool_use" — what's the right control flow?

The credited answer terminates on stop_reason, runs tools, appends results to the full history, and repeats until end_turn. Distractors would parse the assistant's text for "done" or hard-stop on an iteration cap. Recognizing the principle behind the scenario is what earns the point.

# The credited control flow on a scored loop item
while True:
    resp = client.messages.create(
        model=MODEL, max_tokens=1024, messages=history, tools=tools
    )
    if resp.stop_reason == "tool_use":
        history.append({"role": "assistant", "content": resp.content})
        history.append(run_tools(resp))   # append results, continue
        continue
    break  # end_turn / max_tokens / stop_sequence -> stop on stop_reason

Pacing Protects Your Score

Since blanks score zero and every item is independent, a slow pace is a scoring risk: time spent agonizing over one item is time stolen from items you could answer correctly. Don't let a single hard scenario starve the rest of the exam.

A disciplined loop: read the scenario, eliminate the anti-patterns, commit to the strongest remaining option, and flag-and-move if you're stuck. You can revisit flagged items, but you must never run out of time with answers left blank — that's points you forfeited for free.

# A simple per-item triage you can run in your head
def triage(seconds_spent, eliminated):
    if eliminated >= 3:
        return "answer now (1 option left)"
    if seconds_spent > 90:
        return "commit best guess, flag, move on"
    return "keep eliminating anti-patterns"

print(triage(120, 1))  # commit best guess, flag, move on

Turning the Scoring Rules Into a Plan

The scoring model rewards a specific strategy:

  • Answer 100% of items — no penalty means a blank is pure forfeited expected value.
  • Eliminate anti-patterns first to push each guess from 1/4 toward 1/2.
  • Pick the single best option for the context, not merely a workable one.
  • Spread prep across all 5 domains, weighting D1 (27%) most, since you can't choose which 4 of 8 scenarios appear.
  • Build margin above 720 — treat 720 as the floor, not the goal.

Quick Check: Scoring Mechanics

Apply what you've learned about how scenario questions are scored.

Recap: How Scoring Works

Key takeaways on how scenario questions are scored:

  • Scaled, not raw: performance is mapped onto a fixed 100-1000 scale so forms stay comparable.
  • Pass = 720 on that scale (not 72%) — aim for margin above it.
  • Independent items: 4 of 8 scenarios shown, every question scored, no partial credit, single best option only.
  • No guessing penalty: blanks and wrong answers both score zero, so answer everything.
  • Eliminate anti-patterns to raise each guess from 1/4 toward 1/2.
  • Cover all 5 domains (D1 27%, D2 18%, D3 20%, D4 20%, D5 15%) and pace yourself so nothing is left blank.

よくある質問

「シナリオ問題の採点方法」レッスンは無料ですか？

はい。「シナリオ問題の採点方法」の完全なテキストはこのウェブで無料で読めます。インタラクティブに演習し（組み込みコードエディタと24時間対応のAIチューター）、Claude Architectコースの残りをアンロックするには、CoddyKit PROにアップグレードしてください。 Claude Architectコースには全4レッスンが含まれています。

「シナリオ問題の採点方法」で何を学びますか？

100～1000に換算され、720点で合格し、推測による減点はありません ブラウザで直接実行するハンズオンコードでClaude Architectを演習し、24時間対応のAIチューターがレッスンを進める中での質問に答えます。

Claude Architectを始めるのに経験は必要ですか？

事前経験は必要ありません。CoddyKitのClaude Architectは初級者から上級者向けに構成されているため、ここから始めるか最初から始めて、自分のペースで進むことができます。 これはレッスン1/4です。

「シナリオ問題の採点方法」レッスンにはどのくらい時間がかかりますか？

ほとんどのCoddyKitレッスンは約5～10分かかります。各レッスンはコンパクトでインタラクティブなので、着実に進歩し、ウェブとアプリ全体で正確に前回の場所から再開できます。

このClaude Architectレッスンでコードを書いて実行できますか？

はい。すべてのClaude Architectレッスンに組み込みコードエディタが含まれているため、ブラウザでリアルコードを書いて実行し、即座のAIフィードバックを取得できます。ローカル設定は不要です。

このコースのすべてのレッスン

  1. シナリオ問題の採点方法
  2. シナリオプロンプトの読み方
  3. 誤答を排除する
  4. 模擬試験の完全解説
← Claude Architectに戻る