Lesson 9 — C2 Governance: Your Rules on Top of the Code

Learning objective

Bloom's level: Analyze. Identify which of your own and your training committee's AI-use rules applies to a given real task, and correctly follow it — without being told which rule fires.

Plain language: Same as knowing which parts of the code book apply to this job, and which extra shop rules sit on top of it — figure out which AI rule applies before you start, not after.

Standard: AIHC.1.C2 — "Find and follow the applicable AI-use policy before using AI on consequential work; comply with stage-appropriate restrictions (including 'no AI here') without supervision." Plain words: the rules you follow before you use it.

Your real artifacts today: your actual free-tier claude.ai account settings, and the training committee's real apprentice records (names, quiz scores) you're responsible for.

SSD cycle 1 — two layers of rules

• SAY: Anthropic's own Usage Policy is the baseline code book: "a framework for how Claude should and shouldn't be used, providing clear guidance for everyone who uses Anthropic's products" (doc-set §9, source S11). On top of that baseline, you get to set your own personal rules — and, as a training-committee member, help set the committee's. Real term: AI-use policy ↔ trade version: the code book (NEC) plus the shop's own posted safety rules layered on top of it.
• SEE: A two-layer stack, described: bottom layer, "Anthropic Usage Policy (S11)" — three named categories (cybersecurity misuse, political-process abuse, high-risk consumer-facing uses); top layer, blank, labeled "Frank's / the committee's own rules — to be filled in."
• DO: Read the high-risk category (legal, financial, employment) from S11 and write, in your own words, which of your regular tasks fall inside it (apprentice-certification material, per last lesson's reasoning) and which sit outside it (news-claim brainstorming, wording help).

SSD cycle 2 — what never gets pasted in, period

• SAY: Anthropic's own guidance is explicit, and it holds regardless of any setting you choose: never paste in financial identifiers (SSNs, card numbers, bank accounts), health or medical records, passwords or login credentials, or confidential business/personal documents (doc-set §8, source S16). Real term: sensitive data ↔ trade version: the stuff you'd never write on a job card that gets passed around the site — someone's medical restriction, the safe combination, a real SSN.
• SEE: A checklist mapping S16's four categories against the actual data types your training committee handles: apprentice names + test scores (personal data), any wage or benefit figures (financial-adjacent).
• DO: Write your own committee's "never paste this" list — at least four items, each one grounded in S16 and named using your committee's actual record types.

SSD cycle 3 — setting your own dials

• SAY: Free, Pro, and Max users are asked to choose whether their chats may be used to improve Claude, via a setting under name → Settings → Privacy → "Help Improve Claude" (doc-set §8, sources S7, S8). Turning it off is forward-looking, not retroactive — Anthropic's own words: "Your data will still be included in model training that has already started and in models that have already been trained" (doc-set §8, S7). Separately, incognito chats are "temporary conversations that aren't saved to your chat history or to Claude's memory" (doc-set §8, S14) and are never used for training regardless of your Help Improve Claude setting (S6) — though they're still retained 30 days by default and can't be reopened once closed. Real term for the toggle: model-training setting ↔ trade version: whether your corrected work gets folded into the next class's training materials, or kept out of the file. Real term for incognito: ↔ trade version: clocking out — no shift record in your personal history, though payroll still keeps something behind the scenes for a stretch.
• SEE: A described view of the actual Settings → Privacy screen with the Help Improve Claude toggle, next to a blank line for his own written reason.
• DO: Go into your real free-tier account settings. Set the Help Improve Claude toggle one way, and write your one-line reason. Then write your own personal rule for when you'll use incognito (for example: "any time I'm drafting something with a real apprentice's name in it").

Independent at-bat (embedded scenario judgment — unscaffolded)

You're about to ask Claude to help you grade an apprentice's actual quiz — a sheet with the apprentice's real name and score written on it. Apply both layers from this lesson (S11's employment-adjacent category, and S16's flat never-paste rule) without being told which one fires first, and write the fix you'd make before pasting anything in.

Exit ticket (Bloom's climb: Remember → Understand → Apply → Analyze → Analyze, objective level)

1. (Remember) Name the two layers of rules from cycle 1.
2. (Understand) Why is "never paste this" (S16) a flat rule, rather than something you decide case by case the way you lane other tasks?
3. (Apply) Your committee gets a new member who wants to try Claude for lesson planning. What's the first rule you'd hand them?
4. (Analyze) In your at-bat's grading scenario, which rule fired first — and would the answer change if it were a practice quiz with no name attached?
5. (Analyze — objective level) Walk through one full real task from this week and show which policy layer(s) applied, and how you knew.

Graded against S11, S16, and your own written committee rules — never against whether a choice merely felt cautious enough.

Ledger write

ledger_append:
  learner_id: L4-PUB-FRANK
  lesson_id: L4-lesson09-c2-governance-20260719
  standard: AIHC.1.C2
  exit_ticket: {score: "", bloom_reached: analyze}
  auto_mastery: "0.0 ->  (projected ~0.45 — first lesson on this anchor; the shop-rules-on-top-of-the-code parallel landed fast given his committee role)"
  self_score: ""
  calibration_gap: " — not computed until a real self-score exists"
  journal_prompt: >
    What's one shop rule you'd want written down for the next training-committee
    member who tries this, so they're not guessing the way you were?
  structure_used: "PBL with an embedded SBL judgment call in the at-bat"
  referents_used: [electrical-trade shop-rules-on-top-of-the-code culture]
  next_lesson_seed: "C3 Measurement — bid a real quiz-writing job before starting it, then check the bid against what actually happened; his committee's 'never paste this' list stays in force as a background constraint."

Rubric self-audit (R1–R13)

Escalation: all load-bearing indicators pass → auto-ship (per playbook §5 policy).