Week 12: Quiz (10 questions, 8/10 to pass)

Answer from memory first, then check the key. Each question names the section or cell it tests, so a miss tells you exactly what to re-read.

Questions

1. (MCQ) A model and a harness are not the same thing. Which statement captures the division of labor correctly? (see Concepts, "What a harness is")

   • A) The harness makes the model smarter by adding more parameters.
   • B) The model supplies reasoning; the harness supplies agency (loop, tools, context, verifiers).
   • C) The harness is just an editor; the model does all the work.
   • D) The model supplies agency; the harness supplies reasoning.

2. (MCQ) You are working in OpenCode. Which file does it read automatically as persistent project instructions? (see Concepts, "Rules files")

   • A) CLAUDE.md
   • B) opencode.json
   • C) AGENTS.md
   • D) .cursor/rules/*.mdc

3. (Short answer) In the Week-12 SPEC.md, the single most important line is the refused tradeoff. What does it do that the rest of the spec does not? (see Concepts, "Worked example 1")

4. (MCQ) Which item belongs in the "stays out" column of the managed-context table? (see Concepts, "Managed context")

   • A) The failing test output
   • B) SPEC.md
   • C) The .env file with API keys
   • D) The specific source file being edited

5. (Short answer) The notebook's final cell (cell §5) prints one number. What is it called, and exactly what does it count? (see notebook cell §1 and §5)

6. (MCQ) A spec is the definition of done the loop closes against, which means it is primarily… (see Concepts, "Planning vs execution")

   • A) documentation for future human readers
   • B) a prompt to make the model more polite
   • C) an executable statement of what "correct" means, checked by the verifier
   • D) a replacement for writing tests

7. (Short answer, with numbers) A harness run uses 40,000 input tokens and 10,000 output tokens at $3.00/1M input and $15.00/1M output. Show the arithmetic and give the total cost in dollars. (see Concepts, "Worked example 2")

8. (MCQ) A blast-radius rule calibrates autonomy… (see Concepts, "Verifiers, SPEC.md, and blast radius")

   • A) once per session, at login
   • B) per action, based on how much damage the action can do if wrong
   • C) based on how much you like the model
   • D) only for database writes, never for file reads

9. (Short answer) Name any two numeric columns of the comparison worksheet and what each one records. (see notebook cell §3)

10. (MCQ) In OpenRouter, the :free model variant is best used for… (see Concepts, "OpenRouter and cost control")

    • A) production throughput with a service-level guarantee
    • B) testing plumbing (auth, config, prompts) before paying for real throughput
    • C) fine-tuning your own model
    • D) replacing the harness's permission system

Answer key

1. B. The harness adds the loop, tools, context management, and verifiers; the model only reasons. That is why two harnesses on the same model differ, they differ on those four.

2. C. OpenCode auto-loads AGENTS.md (project + global). CLAUDE.md is Claude Code; .cursor/rules/*.mdc is Cursor; opencode.json is config, not instructions.

3. It forbids the agent from resolving ambiguity by guessing. The rest of the spec says what to build; the refused tradeoff is the one place you say, in advance, where the agent is not allowed to guess, here, it must never invent an ETA for a missing/NaN shipment.

4. C. Secrets (.env, keys) stay out of the window, anything pasted in can be logged, summarized, or repeated. Spec, failing test output, and the specific source file go in.

5. The harness-readiness score, an integer 0 to 3. It is sum(1 for n in ('claude','opencode','dsh') if shutil.which(n)), the count of the three coding harnesses installed and on your PATH.

6. C. The spec is the executable definition of done the verifier checks against, not human documentation, and not a substitute for writing the tests (the spec names them).

7. $0.27. Input: 40,000 × $3.00 / 1,000,000 = $0.12. Output: 10,000 × $15.00 / 1,000,000 = $0.15. Total = $0.12 + $0.15 = $0.27.

8. B. The leash is per action: reading a scratch file is nearly free to get wrong; writing production data is not, the same agent gets different leashes on the two.

9. Any two of plan_quality_1to5 (how close the plan was to your spec), tokens_used, cost_usd, tests_green_first_try (whether tests passed first try), diff_files_changed, code_quality_1to5.

10. B. :free models are zero-cost but rate-limited and lower-availability, use them to prove auth/config/prompts work, then swap to a paid slug for real work.