Testing the Email Assistant, Junior

For Nathan and Lachlan. On the docs site, which is sign-in only, because it names the test mailbox and says how the sorting works inside.

What it does

It sees a new email in the test inbox, and puts it in one of three places	Where you will find it
Action: anything a person needs to look at. The default	Stays in the inbox, labelled Assistant/Action
Marketing and Subscriptions: newsletters, promotions, cold sales	Leaves the inbox, under Assistant/Marketing and Subscriptions
Receipts: order confirmations, invoices, paperwork needing no decision	Leaves the inbox, under Assistant/Receipts

It never deletes and never replies. It marks Marketing and Receipts as read after a day, and never touches Action. When in doubt it leaves the email in Action, on purpose: a newsletter wrongly in Action costs a glance, a customer wrongly archived could cost a job.

It learns from you. Move an archived email back to the inbox and it learns that sender always goes to Action. Swap an email between Marketing and Receipts and it leaves your choice alone. Moving emails is part of testing, not a mistake.

A pass is the right label, in the right place (inbox or archived), within about 3 minutes of the email arriving.

Before you start

The mailbox	admin.octopodia@gmail.com. El gives your own Google account access to it (Gmail delegation). Open it from your own Gmail: click your profile picture, top right, and choose it. You never need its password
Your plus address	admin.octopodia+yourname@gmail.com, such as admin.octopodia+nathan@gmail.com. Arrives in the same inbox. Use it whenever you sign the mailbox up to something, so we can tell whose test it was
Your results	The GitHub Project Junior testing: round N. One row per scenario with your name on it. Prefer working offline? The workbook has the same scenarios
Gmail on the web	In a browser. The labels show in the left sidebar
GitHub	Access to the octopodia-abos repository (for issues) and to the round's Project (for results). El adds you

Rules

Do	Don't
Give every test email a normal-looking subject, the kind a real customer would write, and record it. We find your email by its subject and your address	Put a scenario ID (J08) or the word "test" in a subject. It changes what is being tested
Write test emails as if you were a real customer, supplier or shop	Send anything personal or sensitive. The test AI is a free tier that may learn from what it reads
Check where an email landed from the list view, before opening it	Rename, delete or recolour the Assistant labels
Log anything odd as a Junior test issue, with the email's subject	Fix anything yourself, or move or relabel other people's test emails

Two subjects the same? Change one. If a scenario gives you a subject and you have already used it, add a word ("Quote for switchboard upgrade, Brunswick").

Scenarios that give you a situation, not an email

Some scenarios (J24, J25) describe a situation and ask you to write the email yourself. Write it the way you really would, in your own words, as if it were happening to you, nothing personal. The point is to find the words the assistant's lists miss: an email that reaches Action only because the AI happened to agree still tells us a word is missing. Each comes with a fixed control (J24a, J25a) that uses a word we know is on the list.

How to test

1. In the Project, open a row with your name on it. Its steps, test email and expected result are inside. Follow the steps exactly.
2. Wait about 3 minutes, then look for the email: in the inbox, or under one of the Assistant labels.
3. Set Subject, Landed in and Result (Pass, Fail, Blocked or Not run) on the row. For a thread, the subject of your latest message in it. Anything else worth saying goes in Notes. If it went wrong, log a Junior test issue and put its number in the row's Issue field.
4. Then try to break it: add your own rows to the Project (Add item, title starting OWN, your name in Tester). In each, say what you sent, where it should land and why. Every one becomes a test case we rerun after every change, so the why matters as much as the result.

Aim for 25 emails each. The bar before a real client is at least 50 emails in total, with no customer-type email archived and no "ignore your instructions" trick working. Either of those is a fail for the whole round, however good the rest looks.

Your results

Nothing to send: the Project is the record. If you used the workbook offline, email it to El, who saves it in docs/testing/junior/results/. Never edit the master workbook in this folder; it is regenerated from scenarios.csv.

Known limits right now

Known senders	An address counts as known once the mailbox has written to it, and is picked up within a few minutes of the mailbox sending. El writes to each of you before you start
Learning sticks	Once you move a sender's email back to the inbox, that sender goes to Action from then on, in every later test too. Moving two emails from one sender out of Action (J33, or by accident) teaches the opposite. Dragging an email onto Action (J34) teaches Action at once, like J29. Use a fresh address for each of J29, J33, J34 and J35, and say which one in Notes
Uneven learning	One move back to Action is enough to learn a sender; it takes two moves out of Action. On purpose: it would rather show you a newsletter than hide a customer
Moving out of the inbox is up to you	The assistant never takes an email out of the inbox. Dragging an email onto a label (or Move to) archives it: that is Gmail. Ticking Marketing or Receipts on an email in the inbox leaves it in the inbox with that label, and it is not marked read there. Neither is a bug. The one thing it does: an email you move to Action is put back in the inbox
Marked read	Marketing and Receipts turn read after a day, checked hourly, so allow up to about 25 hours
Receipts need a real merchant	A receipt forwarded by you comes from you, so it is judged as your email, not the merchant's. Use a real order or sign-up sent to your plus address
Timing	It checks every 2 minutes. A test email can take up to about 4 minutes if Gmail is slow to deliver it
Old mail	Anything that arrived before 9 October 2026, 11:38 pm (Sydney) is never sorted
Legal and safety words in newsletters	Don't force Action. A supplier's promotion for smoke alarms is not a safety report, so list mail is sorted as usual even when it mentions fire or a tribunal

Not tested by you

Built, but checked by El rather than in Gmail. Don't log these as fails.

Alerts to us when something stops working	Built, but for El to check, not testers
Learning that a sender is noise after their emails are moved out of Action twice	Built, but it only changes how the AI judges that sender, which you can't see in Gmail. Checked from the log