[straff2002]

Plan JG is a draft, docs only. It's the regulations question tester and apprentice drive-time practice for Field Assist, built on Plan IW's goal model. It was revised the same day to a reviewed question bank, because a model's questions aren't trusted as an answer key. The owner then answered all twelve open questions, and those answers are now decisions in the plan.

What it plans

• One question-bank format, shipped only inside a vault. VaultManifest gains practice_dir, with one file per standard document and edition, plus an optional syllabus (topic names and clause ranges, no text). The contract is Contracts/question-bank.md, with invented fixtures and a portable contract test.
  • The vault author builds banks into the vaults they sell. P0 writes a guide for them (docs/field-assist-question-bank-guide.md) and a checker that uses the app's own validator.
  • An organisation builds its own bank in Avenkin Office. That's the office half, an FX-series plan still to be drafted.
• Approved by publication, and trusted only from a signed source: an EG pack, an archive from a catalogue-listed publisher, or an office assignment signed by the organisation's publisher.
  • An unsigned bank is dormant: it's kept but never asked.
  • An archive with a bad signature installs nothing.
• A bank never carries the book's text. It holds the clause, a short question and answer, and at most a digest of the quote; the quote is read from the organisation's own loaded copy.
  • Licence terms are the publisher's agreement with the vault author or the organisation. The app enforces that the organisation's own copy is loaded.
  • Items stay dormant until the matching edition is loaded.
  • When a newer edition becomes the default, the old items go to needs re-review. They're still asked when the technician names that edition ("quiz me on the 2018 edition").
• Your own questions, the one exception to "the phone never generates a question it can ask". A technician may write questions, or have the phone draft them from their own copy through QuestionAnswerabilityGate.
  • Each draft must be confirmed before it can be asked.
  • Own questions are labelled "your own questions" and graded "against your own answer".
  • They never leave the phone.
  • The plan proposes the same, private only, for organisation technicians.
• Grading uses fixed rules against the approved answer, its variants and its tolerance. An "unsure" or "not heard" answer never counts as a miss.
• Phases:
  • P0: the contract, the validator, the author guide, and asking from a bank.
  • P1: a results feed to the office.
    • It's owned by the organisation, off by default, and disclosed to the technician before the first send.
    • Only rule-graded bank results are sent.
    • Suggested employee-notice wording is shown when the organisation's profile turns it on.
    • The same PR adds a section to docs/eu-ai-act-review-2026-10.md recording the owner's decision: transcription plus validation against a human-approved key is not an AI system.
  • P2: questions from the day's jobs and CarPlay drive time.
    • It's voice only and counted against IW's daily limit.
    • In passenger mode, which is only ever set by the technician saying so, the caps relax and the phone's screen can show the questions.
  • P3: learning goals as IW goals, following the syllabus when there is one.
  • P4: anonymous question proposals to the office, on FP's candidate pattern. A test pins the payload so it carries no identity.
  • P5: your own questions.
  • P6: free-form term answers.

For the owner

The plan records 11 owner decisions and 18 proposed design decisions, including the tier split: practice, drive time, goals and own questions in every tier; organisation banks, results and proposals with team and enterprise. Three open questions remain:

• whether organisation technicians may keep private own questions (proposed: yes);
• the passenger caps (proposed: 15 a drive, 30 a day);
• an adviser's confirmation of the Manitoba note beside the notice wording.

🤖 Generated with Claude Code