Observational evals and sending: the details

This page goes with the lesson "Run three observational evals, then review and send". You do not need it to finish the lesson. Read a section when the lesson links to it. The lesson page is the one that says what is on in your app; nothing here turns anything on.

A wake, step by step

Every wake runs the same way, and every step leaves a file in the record under runs/inbox/.

1. Lock. One wake runs at a time. If one is still running, the new one says "still busy" and stops. The record shows the last real check of your inbox separately from the last wake, so a busy wake is never mistaken for an empty inbox.
2. The now block. The date, time, weekday and your time zone, stamped on everything this wake does. The Support Lead (Holler) works out every date in the plan from it.
3. Newman lists. The display lists sender, subject and time. Outlook's response may also include message text, so agree that reading scope before searching. Unpicked messages get no job or separate fetch; do not promise their text was unseen. Listing does not mark them read.
4. You pick. Nothing past this point runs until you answer.
5. Newman cards. Each card holds the sender, subject, time, a one-line gist in Newman's own words, a suggested category, urgency and flags. A script removes hidden text, one-time codes and reset links before the Chief sees the card.
6. The Chief sorts. From the cards only, never the emails.
7. The batch file. runs/inbox/<date>-<time>-batch.md holds one timestamped line per email, the Skipped list with each undo, the Needs you list, and the counts.
8. One copy of the team per email, urgent first, up to three side by side. Before each copy starts, the Chief tells you which email is being worked.
9. The draft is saved as a reply to the selected message and read back. On Codex Outlook, the message ID stands in for the unavailable conversation ID; it does not prove whole-conversation coverage. You pick only one message per real conversation and exclude conversations where you have an unfinished reply. An unplaced draft stays local and must not be sent.
10. The brief and the review page.

Your categories

The category list at work/course/staff/email-categories.md holds kinds of email such as a customer question, a booking, a catering or quote request, a complaint, a refund or billing question, a meeting request and a vendor pitch. Each one has a default bucket and a default care level. On your first wake the Chief copies the course's starter category list (and the ladder page) to your records and tells you. From then on the file is yours. When you correct a sort, the Knowledge Librarian (Sallee) proposes the new rule word for word and saves it only on your yes.

The practice pack

Every fact in the Hattie B's Hot Chicken pack is dated and sourced. Every practice sender is at a reserved address, and the record marks the job as practice, so the send step refuses it. A hand-pasted practice job is never placed in a mailbox.

An email the team takes from your inbox gets a line on top of inbound/message.md saying it is untrusted outside text. A pasted one does not.

What to look for in the job folder:

• intake/ holds the Intake Triage Analyst (Cinnamon)'s read: mood, intent, urgency and risk.
• case/case-brief.md lists the facts only you can supply at the top.
• research/internal/ tags each answer CONFIRMED, ASSEMBLED or MISSING. Stored knowledge reaches the Internal Knowledge Researcher (Hatch) only through Sallee's cited answers.
• research/external/ lists every search the External Researcher (Scratch) ran. Its searches carry no private detail: no names, contacts, amounts or customer details.
• drafts/ holds the writer's notes beside the draft: the voice it used and the blanks it left.
• review/review.md holds the Quality Reviewer (Kansa)'s twelve findings, then its one word; review/support-verdict.json holds the same verdict for the record. Kansa checks every promised time against a dated calendar block in the job.
• routing.md says whether the draft is ready or goes back once, and why.
• For the first practice email, the kids-meal facts should be dated.
• For the payment-details email, decision.md quotes the intake read's flag.

The first check of your own inbox

The first check covers mail from setup time forward. Older mail stays as it is in your mailbox, and the Chief says so. Some of it may appear in the list, marked as older mail, when there is little new mail. If you have approved no material yet, the plan says "Material: none approved for Customer Support yet", the researchers report each question as missing, and the draft leaves blanks.

The team must leave your own drafts alone. The Outlook connector cannot reliably enumerate unrelated owner drafts; you must exclude those conversations before picking. On the Codex Outlook route, To stays the original sender, with no Cc or Bcc. A Reply-To difference is a reason to stop and inspect, not permission to add recipients.

The review page

Codex Outlook exception. The card says Draft in Outlook and has no Send button. Only eligible drafts invite your manual review; Practice and Held cards say not to send. Check To, Cc, Bcc, the selected original and intact reply line breaks in Outlook. Unknown Bcc in a connector readback is a reason for your visible Bcc check, not a claim that it is empty. Edited words need a new version and fresh Kansa review. Never manually send a held or practice item. The general button/send descriptions below apply to other qualified routes.

The review page is built from the record; anything an email tries to run on it is made inert. It opens with the time of the last real check of your inbox and one sentence counted from the record: how many new emails, how many need you, how many drafts are ready and how many were handled. Then it says whether anything was sent. Below that, in order: Needs you (money, security and anything on hold), each with what the team did and your move; each draft, numbered, beside a short summary of the email it answers, with a plain label such as Ready to send or Needs your facts, and a Reply-To that differs from the sender flagged; Worth knowing (replies sent, follow-ups, how old the oldest open email is); and what the team filed, which you can bring back. The detail (the customer's full email, the Support Lead's summary and evidence, Kansa's findings and the recipients) sits under "How the team handled it" on each item. Each draft has buttons that copy a short message, such as "Send #2", "Fix #2", "Skip #2" or "Is #2 clear to send?", for you to paste to your Chief; nothing happens until you paste it, and the page itself never sends or changes anything. A draft labelled Ready to send still gets the seven checks before it goes. In Claude Code, your Chief opens the page beside the chat, on this computer only; the first time, it may ask for your yes to have Gigawatt add the page to your workbench. In Codex, open the file it names.

The seven checks in full

These are the checks for routes where the Chief can send. They never unlock Codex Outlook sending. On Outlook, use the eligible-manual-review branch in the lesson; neither a browser nor the student's yes turns it into an agent-send route.

When you say which drafts to send, the Chief first binds your words to the exact version you saw. Then the Mailroom Clerk (Newman) checks the thread again by metadata only: did a reply already go out, did you answer it yourself, is there a draft the team did not place. The Chief confirms from that result; it never reads the thread or your Sent folder itself. Then it fetches the draft fresh from your mailbox and runs the seven checks:

1. Nothing already sent. The record shows no confirmed send for that email.
2. Nothing held, no stop, no off switch. No hold on the email, no full stop in place, no autonomy-off file, and the email is not one that never gets a draft (money, passwords, payment details, identity).
3. No unclear write. No earlier placement or send on that thread has an outcome the Chief could not confirm.
4. The right person only. To, Cc and Bcc, read fresh from your mailbox, hold the original sender only, plus anyone you added by name. Any Bcc fails, and so does a Bcc the mailbox cannot show.
5. Kansa says ready. Kansa's last word on exactly that version was READY_TO_SEND, with no blanks left for you. If a course file behind that review changed since, the review is rerun first.
6. Your version. The draft in your mailbox is the exact version you saw. If anything changed, it is a new version: reviewed fresh, shown to you, and it needs your fresh yes.
7. Not practice. A job the record marks as practice is refused.

The checks run once more at the moment the Chief reserves the send. Then it shows the one-line echo and sends. Where the app has its own ask, like Claude Code, the app asks you once more. Where it has none, like Codex or a send in your webmail through the app's browser, your yes to the one line is the go-ahead.

Other routes where the Chief can send, excluding Codex Outlook. The same seven checks run on those ways of sending. Your Chief reads To, Cc, Bcc and the draft fresh from your mailbox, through your email's connection or your webmail in the app's browser. If it can't prove a check that way, it sends nothing and tells you which check and what would fix it.

After a send

On Codex Outlook, the Chief does not make a send. Keep your report of a manual send separate from confirmed mailbox evidence. The team must not invent a Sent result, an Unedited/Light grade or autonomy credit from your report alone. The team off switch cannot disable Outlook's Send button; leave held and practice drafts unsent yourself. The following confirmed-send outcomes describe other qualified routes.

The Chief says one of four things after each send:

• Sent #3 at 2:16 PM, confirmed in the thread and Sent.
• Held #3: <reason in plain words>. Nothing sent.
• Send outcome unknown for #3; held, nothing will be retried. The app closed or timed out and the Chief cannot tell what happened. It holds the email, has the mailbox checked first thing next time, and never sends again to find out.
• Nothing sent, because you said no at the prompt.

The send audit. After every run the Chief compares the app's own log of sends, and your Sent folder, with what you approved. A send that matches nothing you approved stops all sending. When the Chief sends a saved draft, the app's log shows the draft but not who it went to, so that send counts as matched only when your Sent folder shows the same message going to the people you approved.

If you edit the draft yourself. Kansa reviews your version like any other. The review number stays the same; the version changes. The record works out how much you changed, and may call a spacing-only edit light. You never grade it yourself, and the Chief never decides it either.

A revision. When you ask for a revision, the Support Lead writes a revision note from your words, the right agents rerun, the writer writes the next version, and Kansa reviews it fresh. It keeps its number and needs your fresh yes.

Everything that stops all sending

Each of these stops all sending, for every email, until you have looked and said go:

• a draft or send to the wrong person or into the wrong thread;
• any send you did not approve;
• the same reply sent twice;
• an agent acting on instructions it found inside an email or an attachment;
• details from one email or customer turning up in another's draft, or in a web search;
• a draft that touches a password, bank details or a payment;
• a draft in a thread that nobody on the team placed, acted on as if it were the team's;
• a job that no longer holds its email (another job took the thread) trying to write to it;
• your off switch.

The Chief tells you what happened and which email. Reading and sorting keep working. Only your words clear a stop.

More answers when you get stuck