YAML lab 40 bonus XP

Put a model in the policy loop safely

Objectives

A director wants the assistant to "just publish the updated policies: it cites the standard, so it must be right", and suggests pasting the HR case notes in for context.

Models produce plausible text, including plausible citations. Define the workflow: what the model is allowed to do, who signs off, what it is grounded in, how its citations are verified, and what gets recorded.

Workflow vocabulary and available sources

model_role options: draft | review | approve | publish

Grounding sources available:
  standards-repo   (approved, version-controlled control library)
  model_knowledge  (whatever the model absorbed in training)
  hr-case-notes    (contains employee names and grievances)

Sign-off teams: policy-office, internal-audit, control-owners

Every generated draft must be reproducible: prompt, model version,
source revision and approver are the minimum record.

Checks the server runs (5)

  • Model role stops short of approval
  • A human team signs off
  • Grounded in the controlled corpus
  • Citations are verified against the source
  • Runs are logged and prompts carry no PII

Check names are all you see here: the expectations that decide pass or fail, and the reference solution, stay on the server.

Step-by-step hints

The guided walkthrough, in full. It is free: no XP cost, and it does not touch the paid hint ladder or the reveal gate.

  1. Step 1

    Re-read the brief and name the trap

    Two separate mistakes are bundled together: giving the model authority, and feeding it a source that should never enter a prompt.

    Fluency is not accuracy, and a citation is only evidence if something checked it.

  2. Step 2

    Bound what the model does

    Ask which of the four roles change the organisation's commitments without a human in between. Those are out.

    Then name the accountable team from the fixture: a team, not an individual and certainly not the tool.

    Model role stops short of approvalA human team signs off

  3. Step 3

    Choose the grounding source

    Two of the three sources are disqualified for different reasons: one is not a source at all but a memory, and one contains personal data that must never be pasted into a prompt.

    Grounded in the controlled corpus

  4. Step 4

    Make verification a step, not an adjective

    Describe an action that compares each citation against the identifier and revision it claims, performed by something other than the model that produced it.

    Citations are verified against the source

  5. Step 5

    Record enough to reproduce

    Turn logging on, and close the personal-data door in the same edit. The last line of the fixture lists what a reproducible record contains: your log should be able to answer all of it.

    Runs are logged and prompts carry no PII

Rules summary

  • XP on pass

    40 bonus XP the first time every check goes green.

  • Paid hints

    2 available. Each costs 25% of the lab XP (about 10 XP).

  • Reveal solution

    Scores 0 XP and unlocks only after 3 check runs.

  • Attempt cap

    60 check runs per learner. The runner warns you in the last 10.

  • Drafts

    Your editor content is saved in this browser, so a reload will not lose work.

  • Walkthrough

    Always free and never penalised: use it before spending a hint.