ADR-0064 P0 gate: the hand-labelled open-loop set

Pre-registered 2026-09-25, before any classifier rule was written. This file is committed ahead of the mail module on branch mail/p0; the commit order in the PR is the evidence.

How the set was built

Both of Kannaka's mailboxes were read read-only (IMAP EXAMINE, BODY.PEEK only, no flag changes) on 2026-09-25 ~21:00Z:

The 30-day window (since 2026-08-26) covers each mailbox's whole life: Zoho opened 09-09 and ninja-portal on 09-16. Messages were grouped into threads per account by Message-ID / In-Reply-To / References. That gives 51 threads (49 zoho, 2 np). Each thread was then labelled by a person reading it: the headers, and the bodies where a header alone could not say. Bodies were read live and never stored.

Stalwart does not file SMTP submissions into Sent Items, so the np account has no Sent copy of the mail Kannaka sent from it. Those letters appear as inbound mail in the zoho INBOX (T42, T43).

What each label means (P0 semantics)

• needs_reply: a person wrote to us, a response from us is warranted, and none has been sent.
• waiting_on_them: our message is the latest one, and it asks a person something they have not answered.
• done: nothing is owed by either side in this thread. It was answered, acknowledged or purely informational, or it was resolved elsewhere and recorded with mail close.
• ignored: automated or bulk mail, our own tool or test traffic, agent-to-agent traffic inside ninja-portal.com, or unsolicited mail that is not correspondence.

A commitment stated in prose ("I owe you a reply", "I'll send a plan", "next table when the run lands") is P1 and does not change a P0 label. Such commitments are noted in the table.

The gate metric

• Gate (strict): exact agreement of kannaka mail status with the label below, over all 51 threads, with the four states compared as they are. Pass at ≥ 95% (at most 2 of 51 wrong).
• It is reported twice. Structural means before any close: the label column, with T40 as needs_reply. After closures means after the one pre-registered close on T40, where T40's label becomes done. No other close is pre-registered, and no other will be run for the gate.
• Also reported: the open-loop view (needs_reply / waiting_on_them / not-open), and every disagreement with its cause.
• Expected misses, declared before the rules exist: T19 (a script-sent key delivery from a person's address) and T20 (recruiter cold mail from an Outlook tenant). Neither carries any automation header (List-*, Auto-Submitted, Precedence, Feedback-ID, VERP Return-Path), so a header rule that caught them would be overfitting.
• The rules are designed with this set in view, and no held-out set exists (these two mailboxes are the whole corpus). The number is therefore in-sample. The honest out-of-sample check is to re-run this gate on the next 30 days.

The labelled set

Anchor = the latest message in the thread (account:folder:uid). Dates are UTC.

Totals (structural): needs_reply 3 (T31, T32, T40) · waiting_on_them 2 (T14, T48) · done 24 · ignored 22 (20 zoho + 2 np) = 51. After the T40 close: needs_reply 2 · waiting_on_them 2 · done 25 · ignored 22.

Measured result (2026-09-25 ~21:5xZ, appended after the run)

Run: kannaka mail sync --days 30, then kannaka mail status --all --json, compared to the table above by anchor membership. The np threads are matched by subject tag, because JMAP anchors are email ids rather than UIDs. Sync read zoho INBOX 56 + Sent 68 over IMAP, and np Inbox 2 + Sent Items 0 over JMAP: 126 refs, 51 threads, and every label matched exactly one thread with none left unlabelled.

Disagreements (both runs), exactly the two declared in advance:

• T19 "Your Kannaka Brain key": labelled ignored, status says needs_reply (their last message is unanswered). A script sent it as nick@spacechild.love with no automation header.
• T20 "Ubie US HealthTech Career Opportunities": labelled ignored, status says needs_reply. Recruiter cold mail from an Outlook tenant, with no automation header.

Both are false opens: the list shows two extra items to dismiss, and no real loop is hidden. Neither was fixed by a rule, since that would be fitting the rules to the test set. A close with a note is the P0 tool for them, and a sender-reputation or first-contact signal is a P1 question.

The three 09-25 cases:

1. Vincent (T48, "The Agentic Arena: which activity feeds msgs_per_bot?", ours 09-25 20:42Z): waiting_on_them, because our last message asks and is unanswered. ✓ The earlier postmortem thread (T39) is done; his "I owe you a reply" there is a P1 commitment.
2. Nick's "Help if possible" (T40, t:e4c81f1ddd): needs_reply → `kannaka mail close t:e4c81f1ddd --note "deployed 09-23 (kax-scada-desk-run2)" → done (closed: deployed 09-23 (kax-scada-desk-run2)). ✓ The closure is only a local row in <data_dir>/mail/closures.jsonl`, with covers_through 2026-09-23T03:24:45Z; a newer message in the thread would re-open it.
3. Nick's forward of Victor Cypher's offer (T38): done (FYI), because Nick's own words are "Fyi..". ✓ P1 adds "Victor owes a plan" as a commitment on them.

Read-only, checked against the authority. "Help if possible" was still unread on the server (\Recent, no \Seen) after two full-body syncs and two live thread fetches. Every fetch was a PEEK. refs.jsonl (118 KB) holds no body text: grep for a known body phrase and for the sk- key prefix returns 0.

Caveat, stated before the run and restated here: the number is in-sample. The re-check is this same procedure on the next 30 days of mail.