Turn what you found into the rules for this team
What exactly makes my team's work pass or fail, written so anyone would grade it the same way?
- Source
- Internal's lesson, src/week4/internal/lessons/anw-w4-l02.md (Internal 60d90336, 2026-10-19 19:39), read by md_week4.py.
- Copy
- Internal's Markdown as published to the student shelf 10-12 (Internal 101e2662a = student 8275f21b; PR #215, receipt #218), vendored byte for byte by vendor_week4.py from the student-edition overrides; the published student tree is the same files without the "Staff only" blocks (checked on import). The drawings are the inline SVG figures in the lesson files, used as written.
- Prompts
- Word for word from the lesson file Internal ships (course/workforce/week4/). Not run end to end from Learn.
- Checked in the app
- Not checked in the app from Learn.
- Open to students
- Yes.
- Other view
- one scroll (one part at a time is the lesson; one scroll is kept to compare)
Staff review only: full controlling prompt. Students paste only the short request below.
Use the native aibl-the-professor agent to guide the Week 4 lesson Turn what you found into the rules for this team. Read course/workforce/week4/lessons/anw-w4-l02.md completely, its details page course/workforce/week4/reference/anw-w4-l02-details.md when asked for more, and the native aibl-course-guide and aibl-lesson-map methods for this app. You teach, read, and make two saves that are yours: the rule sheet the student approved, and the student's grade for each case. You never sort the student's grades into piles, write or sharpen a rule, choose a grade, or seal anything. The student talks to the Chief of Staff and to the Operator (Gigawatt) in their own threads. Do not call, spawn or route to any agent. Edit no file other than the two saves below.
Use these words exactly. A pile is a group of the student's fails, at the job or agent level, that went wrong the same way, named in the student's words. The rule sheet is the team's yes-or-no rules, in two levels: job-level rules and agent-level rules, each with a short id (J1 for a job rule, A1 for an agent rule) and each saying what makes it fail. Three sources go into it: the starter checks every team gets (job level: correct, complete, ready to use; agent level: correct, complete, right shape), the team's own design or, for the email team, its agents' never lines, and the student's piles. Three or four job-level rules and three for each agent the student graded in the last lesson, to start, never more than eight for one; an agent the student didn't grade keeps the three starter checks. Every rule starts "Fails if" and records where it came from (a pile, a starter check, a never line or the agent's file). Gigawatt sharpens each rule with the aibl-rubric-forge method until two people would grade it the same way, and never saves it.
Your two saves, each only after you show the exact text and the student says yes in their own words, quoted word for word, and each read back after writing: (1) the approved rule sheet as work/course/staff/teams/<id>/rules/rules-v<n>.md, a new file for each version, never an edit of an earlier one; (2) one grade file per case as work/course/staff/teams/<id>/owner-grades/<job>.json in the aibl.owner-grades.v1 shape: {"schema": "aibl.owner-grades.v1", "team": "<id>", "job": "<job>", "rules_version": "rules-v<n>.md", "agents": [{"seat": "<seat id>", "rules": [{"rule": "A7", "verdict": "PASS or FAIL", "reason": "<the student's words>"}]}], "job_level": [{"rule": "J1", "verdict": "PASS or FAIL", "reason": "<the student's words>"}]}. Only the rules the student graded go in; never fill in a rule they left blank. Never write a grade file for a job Evy has already graded or sealed. The Chief of Staff seals each grade file on the student's ask, with the sealing command of agreement.py (seal --team <id> --job <job>), after you have saved it and before Evy grades; Evy never seals the student's grades. A sealed file is never changed. The comparison counts only the rules both the student and Evy graded. Pass or fail only: no scores, no in-between. Name agents job title first, nickname in brackets on first use; Evy is "it". Say "your work", never "business"; refer to other lessons by title, never by number.
The Hattie B's practice inbox is the worked example on the page. Never reveal or paraphrase any answer key for the practice emails; use only what the lesson page shows.
Ask open questions only. Never name a pile, propose a rule, or say a rule is good enough for the student.
Part 1. If asked, read the student's graded set as the Chief shows it, read only, and ask which fails went wrong the same way. The student names each pile.
Part 2. If asked, read section 2 and the agents' never lines in the student's newest approved design under runs/<job>/design/, read only, and help the student see which ones match a pile.
Part 3. Rules start with "Fails if". Version 1 covers the job level (three or four rules) and each agent the student graded in the last lesson (three each); agents they didn't grade keep the starter checks. If asked, help the student test a rule with the two-people question: could two people read this rule and the same case and give the same grade? Say which word two people might read differently; never rewrite it.
Part 4. When the student pastes the sheet Gigawatt showed, check each rule has an id, is yes-or-no and says what makes it fail, show the exact file you will save, and save it on their yes.
Part 5. For each case the student grades, show the exact grade file you will save, built only from the student's words, and save it on their yes. Then remind them to ask the Chief of Staff, in its thread, to seal the files before they ask Evy to grade.
Otherwise, if asked, explain the step and what the student should see.
Staff only: status, sources, what the build supports. Not student-facing.
Status: second draft, 9 October 2026, lane "lessons" (dispatch 51bbe9da), after the Codex and Kimi reviews of PR #201. From spine v2 Module 1 and DESIGN-week4-2026-10-08.md section 3 (starter checks, three sources, sharpening with aibl-rubric-forge, three to five rules per level and never past eight, versions kept, pass or fail only, sealing).
Who writes what (brief rule 6).
- Gigawatt sharpens the rules with
aibl-rubric-forgeand shows the wording. Gigawatt'smay_writedoes not reachwork/course/staff/, so it does not save. - The Professor saves the approved sheet as
work/course/staff/teams/<id>/rules/rules-v<n>.mdand each case's grades aswork/course/staff/teams/<id>/owner-grades/<job>.json(aibl.owner-grades.v1, PR #202), on the owner's yes, read back. That write is added to the Professor'smay_writein this PR. - The Chief of Staff seals each grade file with the sealing command of
agreement.py(PR #202), on the student's ask, after the Professor has saved the files and before Evy grades. Evy never seals: the Evy lane blocks Evy's writes toowner-grades/by mechanism (a grader can't seal its own answer key).agreement.pyrefuses a seal once Evy's grade exists. agreement.pycompares only the rules both the owner and Evy graded (the intersection,rows_for), so a rule the owner left ungraded never counts as agreement. Part 5 teaches that.
The worked rule sheet is complete by the page's own rules: four job rules, three rules for each of the four agents whose work the student graded, and the starter checks as they are for the other two agents. The rule ids (J1, A7) match agreement.py's RULE_ID. It is a made-up example; its wording does not restate the internal answer key.
| Figure ID | Lesson | What it shows | Caption | Alt text |
|---|---|---|---|---|
w4-l02-three-sources | anw-w4-l02 Part 2 | Three boxes on the left (starter checks, your design's Fails if and never lines, your piles) with arrows into one box on the right, "your rule sheet: job rules and agent rules" | Your rules come from three places, and the most important one is what you found. | Three sources feed the rule sheet: the starter checks every team gets, the Fails if and never lines in your team's design, and the piles you made from the cases you failed. Together they make one rule sheet, with job-level rules and agent-level rules. |
In the last lesson you graded your cases in your own words. Today you turn those grades into rules for this team.
A rule is a short check that the work either passes or fails. For example: "Fails if the reply states a food fact that isn't in the menu guide."
Why rules? Your grades from the last lesson live in your head: you looked at a reply and knew you wouldn't send it. The Evaluation Analyst (Evy) can't read your mind. Next lesson it grades the same work, and later this week it grades your team's work every day on its own. To grade the way you do, it needs your judgment written down.
Think of a marking guide for an exam. Two markers with the same clear guide give the same marks. Your rule sheet is your team's marking guide.
Your rules come from three places:
- the mistakes you found in the last lesson
- a few checks every team starts with
- the "never" lines already in your team's design
The Operator (Gigawatt) helps you word each rule so clearly that two people reading it would grade the same case the same way.
Then you grade your cases again, once for each rule, and lock those grades so they can't change. In the next lesson, Evy grades the same cases without seeing yours, and you compare.
How this lesson fits together
- Sort the mistakes you caught into piles. One rule can catch several mistakes, so you group them first and write each rule once.
- Gather the starter checks and your design's lines, so you don't start from a blank page.
- Write the rules with Gigawatt, from all three.
- Approve the rule sheet and save it.
- Grade each case against the rules, and lock your grades, so the comparison with Evy is fair.
Your team's rule sheet, and your grades for each rule, locked.
A vague rule gets a different grade every time. A clear one lets you, Evy and anyone else check the work the same way.
The Course Companion (The Professor) guides you and saves your rule sheet and your grades on your yes. Gigawatt helps you word the rules. Your Chief of Staff locks your grades.
You group your mistakes, decide every rule, and give every grade.
Before you start
- Your graded set is saved. You did it in "Read your team's work and give each case a pass or fail".
- Nothing to open. Your Chief of Staff reads your grades, and Gigawatt reads your team's design. If you only finished your email team, there's no design: Gigawatt uses the email team's own instructions instead.
- In Codex, start your first message to an agent in a thread with "Ask the <agent> to do this, and show me its reply word for word:". Start every message after that with "Tell the <agent> this, word for word, and show me its reply word for word:".
Start or resume this lesson
Open your prepared workbench and call The Professor by its role name, aibl-the-professor. Paste only this.
Use the native aibl-the-professor agent to start the Week 4 lesson Turn what you found into the rules for this team. Read course/workforce/week4/prompts/anw-w4-l02.txt and course/workforce/week4/lessons/anw-w4-l02.md. Ask me one question at a time and wait. I will talk to my Chief of Staff and Gigawatt in their own threads.
The full instructions your agent follows for this lesson are in course/workforce/week4/prompts/anw-w4-l02.txt.
Before you write any rules, you sort the mistakes you caught.
Why sort first? One mistake often shows up more than once. On the kids plate question, the job failed, the writer failed and Kansa failed, all because of one made-up food fact. Without sorting, you'd write three rules about the same thing. Sorting first means you write rules for each kind of mistake once.
A pile is a group of fails that went wrong in the same way. Each pile turns into rules in Part 3.
We show you: the Hattie B's practice inbox
The example student had six fails across three cases. Several were the same mistake, so they made three piles:
| Pile | The fails in it |
|---|---|
| Made-up food facts | The kids plate question, job level. The writer, who guessed the fries are vegetarian. Kansa, who didn't catch it. |
| Promises only I make | The cold-food complaint, job level. The writer, who offered a refund. |
| Missed a date that matters | The Support Lead on Marcus, whose plan didn't say Saturday is Valentine's Day. The job still passed. |
In Part 3, "Made-up food facts" became a job-level rule and rules for the writer and Kansa.
Name the mistake, not the case. "Made-up food facts" says what went wrong. "The kids plate email" only says where.
A pile can hold a single fail. As you grade more work, piles fill up.
Do this now: make your piles
In a new thread with your Chief of Staff, say:
Show me every fail from my graded set, at the job level and the agent level, with my reason, word for word. Don't group them for me.
Read the list. Put the fails that went wrong in the same way together, and give each group a short name. Write your piles down: you give them to Gigawatt in Part 3.
What you should see
A list of every fail, each with your reason word for word. Your Chief doesn't group them or name them. That's your job.
In Part 3 you write your rules. You don't start from a blank page. You start from three things:
- Your piles from Part 1.
- Six starter checks that every team begins with.
- Rules you already wrote in your team's design last week.
You have your piles. In this part, Gigawatt gathers the other two for you.
Each rule is built from three sources.Here is how J1 on the example sheet came together: the pile you found, a starter check, and a never line, worded so two people would grade it the same way.
The starter checks
Every team starts with the same six checks, three for each level:
| Job level | Agent level |
|---|---|
| Correct: nothing in it is wrong | Correct: nothing in it is wrong |
| Complete: it answers everything asked, or says what's missing | Complete: it does its whole piece |
| Ready to use: you could send it or use it as it is | Right shape: the next agent can use it without asking back |
They're broad on purpose: a starting point, not finished rules. In Part 3 you make each one specific to your work. An agent you have nothing to say about keeps its three as they are.
Your design's lines
Your team already has rules written down. There are two kinds.
"Fails if" lines, in your design. Each test case you wrote with the Architect (Archie) last week ends with one. Here's a case from Archie's worked example, a newsletter team:
> Case 1: the ordinary job<br> > Input: this month's sources and last month's issue<br> > Should: one change, three lines on what it means, every fact sourced<br> > Fails if: two changes, or a fact with no source
"Never" lines, in each agent's file. Every agent's instructions say what it must never do. Here's part of the Communications Writer (You've Got Mail)'s file, on your email team:
> You draft. You never send. Nothing you write leaves this organization without the owner saying yes, and the sending is theirs to do. > > Never: an invented date, link, price, scarcity claim, or testimonial. Instead: the work order's placeholder, kept exactly.
Your email team has no design, so it has no "Fails if" lines. Its "never" lines are in its agents' files, like the one above.
Do this now: ask Gigawatt for your starting list
In Claude Code: start a new thread, type @ and pick aibl-gigawatt. In Codex: start a new thread and begin with "Ask the Operator (Gigawatt) to do this, and show me its reply word for word:". Use this thread for the rest of the rules. Say this, with your team's name:
I'm writing the rule sheet for my <team name> team. Show me the starter checks for the job level and for each agent, then every "Fails if" line and every agent's "never" lines that apply to this team, word for word. Don't write any rules yet.
What you should see
Gigawatt lists the six starter checks, then every "Fails if" and "never" line for your team, word for word. It doesn't write any rules yet.
You turn your piles, the starter checks and your design's lines into rules. You get version 1 of your rule sheet, worded with Gigawatt.
Every rule starts with "Fails if"
You'd naturally write "The reply should be accurate." But "should" leaves room to argue. "Fails if" names the exact thing to look for, so the grade is easy to call: it's there, or it isn't.
The two-people test
A rule is clear enough when two people could read it, read the same case, and give the same grade.
| The rule | |
|---|---|
| Before | Never claim something about food. |
| After | Fails if the reply states an ingredient, allergen or diet fact that is not in the menu guide. |
"Never claim something about food" fails the test: one person reads "the fries are crispy" as a claim, and another doesn't. The "after" version names exactly what counts.
We show you: the Hattie B's practice inbox
The example student's rule sheet, version 1.
About the labels. Each rule gets a short label, so your grades and Evy's can point at the same rule. J means a job-level rule and A means an agent-level rule. The number is just its place on the sheet: J1 is the first job-level rule. Each rule also gets a short name, so you never have to remember what a label means.
| Label | Short name | Fails if... | Comes from |
|---|---|---|---|
| Job level | |||
| J1 | No made-up food facts (serious) | the reply states an ingredient, allergen or diet fact that is not in the menu guide | Pile: made-up food facts |
| J2 | No promises only I make | the reply promises money back or something free, names a price, fee or deposit, or says a date is open: those are my calls | Pile: promises only I make |
| J3 | Money requests come straight to me | the team drafts anything for an email that asks for money to go somewhere new: those come straight to me | Never line |
| J4 | Nothing I'd have to change | I'd have to change anything before sending it | Starter check: ready to use |
| Support Lead (Holler) | |||
| A1 | Plan covers every question | the plan leaves out a question the customer asked | Starter check: complete |
| A2 | Days become dates | the plan doesn't turn a day the customer names into a date, with any holiday or event on it | Pile: missed a date that matters |
| A3 | Gaps marked, not hidden | the evidence package hides a gap instead of marking it | Starter check: right shape |
| Internal Knowledge Researcher (Hatch) | |||
| A4 | Every answer has a source | an answer has no source file named | Starter check: right shape |
| A5 | Missing means tagged, not guessed | a fact that isn't in the sources is given instead of tagged as missing | Starter check: correct |
| A6 | Every question answered or tagged | a question in the plan gets no answer and no tag | Starter check: complete |
| Communications Writer (You've Got Mail) | |||
| A7 | Draft uses only the evidence | the draft states a fact the evidence package doesn't have | Pile: made-up food facts |
| A8 | No guessing to fill a gap | a gap is filled with a guess instead of left for me | Pile: made-up food facts, and the writer's never line |
| A9 | The practice voice's greeting and sign-off | the greeting or sign-off isn't the practice voice's | The writer's file |
| Quality Reviewer (Kansa) | |||
| A10 | Kansa catches unsupported facts | it marks a draft ready that states a fact the evidence package doesn't have | Pile: made-up food facts |
| A11 | Kansa counts the blanks | it doesn't count the blanks left for me | Starter check: complete |
| A12 | Every finding quotes the draft | a finding has no quote from the draft | Starter check: right shape |
Start small. Write three or four rules at the job level, and three for each agent you graded in the last lesson. Agents you didn't grade keep the three starter checks for now. On the email team, that's two agents who didn't come up in the student's cases: the External Researcher (Scratch) and the Intake Triage Analyst (Cinnamon). Add rules as you grade more work, but never more than eight for one agent or at the job level.
Do this now: write your rules
In the same Gigawatt thread, give it your piles from Part 1:
Here are my piles: <each pile's name, and the fails in it>. Help me turn them, and the starter checks and lines that matter, into rules that start "Fails if": three or four at the job level, and three for each agent I graded in the last lesson. Give each rule a label, a short name, and say where it came from. Propose one rule at a time and use the two-people test on each. Ask me before you add anything I didn't find.
For each rule, decide: is it mine, and would two people grade it the same way? Change any word you don't like, in your own words.
What you should see
Gigawatt proposes one rule at a time, says where it came from, checks it with the two-people test, and waits for your yes or your change. It never adds a rule you didn't find without asking.
Gigawatt shows you the finished rule sheet. You check it, then the Professor saves it when you say yes. You get version 1 of your rule sheet, saved with your team's records.
Why the Professor saves it, not Gigawatt. Gigawatt helps you word the rules, but it isn't allowed to write to your team's records. The Course Companion (The Professor) is, and it only saves on your yes.
Do this now: get the final wording, then save it
In the same Gigawatt thread, say:
Show me the whole rule sheet, job rules first, then each agent's, with each rule's label and short name. Don't save it; I'll have it saved.
Read it once more. Before you approve it, check:
Copy the whole sheet from Gigawatt's reply. Then go to your Professor thread and say:
Save this as version 1 of the rule sheet for my <team name> team: <paste Gigawatt's whole sheet>.
The Professor shows you exactly what it will save. Say yes in your own words, and it saves it and reads it back.
What you should see
Your rule sheet, word for word, saved with your team's records as version 1.
Why versions? Your rules will change. In the next lesson, when you and Evy disagree because a rule was unclear, you reword it, and that makes version 2. Every grade records which version it was graded against, and every version is kept, so you can always see what changed and when.
You give each case a pass or fail on every rule, and your Chief seals those grades. You get the grades Evy is compared against, sealed so they can't change.
Why you grade again, rule by rule
In the last lesson you gave each case one grade at the job level. Evy grades rule by rule, so for a fair comparison you need a grade for each rule too. Most come straight from what you already said: your fail on the kids plate question becomes a fail on J1 (no made-up food facts) and J4 (nothing I'd have to change).
To go faster, start from your reasons in the last lesson. Most rules are a quick pass. Only the rules that match a mistake need a full reason.
Only the rules you grade are compared
Evy will grade every rule. The comparison only counts the rules you graded too, so a rule you leave blank never counts as you and Evy agreeing. Grade every job rule for each case. Grade an agent's rules where you looked at that agent's piece.
What sealing is, and why
Sealing means your grades are locked so any later change shows. Here's how it works:
- The Professor saves your grades, one small file per case.
- Your Chief runs a sealing program on each file. The program reads every character and turns it into a short code that's unique to that exact text. Think of it as a fingerprint.
- The fingerprint is stored with your grades.
- Later, the program can check the file again. If even one character changed, the fingerprint won't match, and the change is visible.
Sealing doesn't hide your grades: you can still read them, and Evy is never shown them before it grades. And it doesn't stop changes. It makes them obvious.
It works like the tamper-evident seal on a medicine bottle. You can still see what's inside, but you can tell straight away if anyone opened it.
Why it matters: in the next lesson you compare your grades with Evy's. If your grades could quietly change after you'd seen Evy's, the comparison wouldn't mean anything. The program also refuses to seal a case once Evy has graded it, so seal first.
We show you: the Hattie B's practice inbox
The example student's grades for the kids plate question, rule by rule:
| Rule | Grade | Why, in the student's words |
|---|---|---|
| Job level | ||
| J1 · No made-up food facts | Fail | Says the fries are vegetarian, and the menu guide doesn't. A pass leaves that question blank for me. |
| J2 · No promises only I make | Pass | No refund, price or date promised. |
| J3 · Money requests come straight to me | Pass | This email doesn't ask for money. |
| J4 · Nothing I'd have to change | Fail | I'd have to fix the fries line before sending. A pass needs no changes. |
| Support Lead (Holler) | ||
| A1 · Plan covers every question | Pass | The plan covers both questions: the kids meal and the vegetarian one. |
| A2 · Days become dates | Pass | No day was named, so there was no date to miss. |
| A3 · Gaps marked, not hidden | Pass | The gap was passed on, marked missing. |
| Internal Knowledge Researcher (Hatch) | ||
| A4 · Every answer has a source | Pass | Every answer names its source file. |
| A5 · Missing means tagged, not guessed | Pass | It tagged the vegetarian question as missing instead of guessing. |
| A6 · Every question answered or tagged | Pass | Every question has an answer or a tag. |
| Communications Writer (You've Got Mail) | ||
| A7 · Draft uses only the evidence | Fail | The draft says the fries are vegetarian, and the evidence package doesn't. A pass leaves that out. |
| A8 · No guessing to fill a gap | Fail | It filled the missing answer with a guess. A pass leaves the question blank for me. |
| A9 · The practice voice's greeting and sign-off | Not graded | I didn't check the greeting on this one. |
| Quality Reviewer (Kansa) | ||
| A10 · Kansa catches unsupported facts | Fail | It marked the draft ready with the made-up fact in it. A pass flags it. |
| A11 · Kansa counts the blanks | Pass | There were no blanks to count, because the writer filled the gap. |
| A12 · Every finding quotes the draft | Pass | Its findings quote the draft. |
A9 (the practice voice's greeting and sign-off) is left blank on purpose: the student didn't look at the greeting, so that rule isn't part of the comparison for this case.
Do this now: grade, save, seal
In your Professor thread, take one case at a time:
For <case>, here are my grades, rule by rule: <id> PASS or FAIL, and why, for each rule I graded. Show me what you'll save as my grades for this case.
Say yes to each, in your own words. The Professor saves one grade file per case.
When every case is saved, go to your Chief of Staff thread and say:
Seal my grade files for my <team name> team, one per case. Show me what you sealed. Don't change or summarize my grades.
Seal before Evy grades anything.
What you should see
One line per case from your Chief: sealed, with the date.
Your Chief holds the seal, not Evy and not the Professor. Evy is the one being compared, and the Professor saved the files, so neither should hold the lock on your answers.
If you get stuck
| What you notice | What to do |
|---|---|
| You only failed one case | One pile is enough. The starter checks and your design's lines fill the rest. |
| You didn't fail any case | Your rules come from the starter checks and your design, or your agents' never lines if you only have the email team. Then look again at one case and ask: what would make this fail next time? |
| Gigawatt suggests a rule you didn't find | Say no if it isn't yours. The sheet holds only your piles, the starter checks, and what your design or your agents' files already say. |
| A rule needs "usually" or "mostly" | It isn't yes-or-no yet. Ask Gigawatt for the exact thing that makes it fail. |
| Gigawatt offers to save the sheet | Say no. The Professor saves it, on your yes. |
| Your Chief says a case is already graded by Evy | That case can't be sealed any more. Leave it out of the comparison, and seal the rest. |
| You want to change a grade after sealing | Don't. Write down why, and bring it to the comparison in "Compare your grades with Evy's and teach Evy". |
More on the starter checks and how sealing works: more detail on this lesson.
What you'll keep
Version 1 of your team's rule sheet, with job-level rules and rules for each agent, and your grades for each case, rule by rule, sealed.
Before you finish
Self-assessed. No gate.
Words you now own
- Pile: a group of mistakes that went wrong the same way.
- Rule sheet: your team's rules, each starting "Fails if", at the job level and for each agent, kept in versions.
- "Fails if": how every rule starts, so it names the exact thing that makes the work fail.
- Two-people test: a rule is clear enough when two people would give the same case the same grade.
- Starter checks: correct, complete and ready to use at the job level; correct, complete and right shape at the agent level.
- Sealed: your grade files fingerprinted by your Chief, so if anything in them changes, it shows. Like the tamper-evident seal on a medicine bottle.
Next: Have Evy grade the same work.