A caller's message always reaches the office

Spot-check sheet: seven real calls, old email next to new email (Oct 29).

Live proof, Willows test line, Oct 28 5:40 PM MT (conversation conv_voice_e3e53a2f-2c6f-4a9f-88ea-dcf9ca9e3792, notice tn_eb1fd8dc-358d-4a43-b248-232ee9631d03, delivered 20 s after hang-up, SendGrid MYHWCY7FRSybtHRqUZDu7g):

Subject: Marcus — general callback

Marcus wants to know what services the
property management team offers…

Callback number: (970) 310-3966

Subject: Pat R. — Security deposit refund

Pat Rivera (resident).
moved out of Unit 2B at the end of September
and hasn't received the security deposit
refund yet. Wants the office to follow up
on it.
Callback number: (970) 555-0142
Called Wed, Oct 28, 5:40 PM MDT

No call time on the old note, and 8 of 47 messages in 30 days dropped by the wording checker (latest: Camellia tub touch-up voicemail Oct 28 21:45Z, recovered by hand 21:54Z). On the new slip the lowercase second line is defect #23341 (fixed, in queue); the number appears once; nothing can be dropped by wording any more.

Replay numbers

• 42 real October calls (Western Slope, JP&Co): 0 lost, 0 sent that must not, 0 worse than current.
• Slip score 234/240.
• Sentences missing a subject: 17 of 39 → 0 of 40 after #23341.

Follow-ups (issues)

• #23271 — 30-day replay of answered-question skips
• #23272, #23307, #23327 — cleanups
• #23273 — keep a different callback number at call end
• #23274 — voice fast-lane transcripts
• #23305 — number inside the note could become the callback on the sweep-only path
• #23306 — ADR-0145 decision 9 amendment
• #23289 — "Hi Mountain" greeting (comma in ", Inc")
• #23337 — Clara says "I've left the team" before the caller finishes

Final after: real Willows call, Oct 29 17:33Z, email received 80 s later

Evidence: conversation conv_voice_5ba9e54f-0cb2-40d2-bbca-12d96cebb378, notice row tn_dd5025f6-9e65-48ea-beca-b62bc321b402, email received 2026-10-29T17:34:18Z in the Willows test inbox; production page card shows the full subject. Since-live check at real customers (Oct 28 23:39Z to Oct 29 03:30Z): 13 calls, 3 notes delivered, 9 correctly silent, 1 judgment call (caller withdrew), 0 dropped. Spot-check sheet: seven real calls, old next to new.

On the call Clara already writes the note. Then at call end a second model rewrites it into one sentence, a checker polices that sentence (banned words such as "she", one sentence only, 28 words, grounding), and when the rewrite fails three times the engine sends nothing. The backup sweep is disarmed on the way out, and the ops card is a model's guess from the transcript.

1. The note Clara writes on the call is the message. No model sits between capture and send. It goes under a new, minimal slip template (who · what · number once in digits · when; topic in the subject). The old "message taken" template is retired: it is what produced the Dora email ("Property: not identified yet", "The office was closed", "Dora called.", subject "Gomez called to ask…"). The capture prompt is fixed too: today it tells Clara to write "I'll use the number you're calling from" into the note (89% of Western Slope notes restate the number, often in words), contradicting the tool's own "never restate the number" rule.
2. The note is written for the office, once, by Clara on the call. Fede on Dora's email: "this sucks, full of fluff". Measured: the engine's own sentences are short and clean but have no time and no topic in the subject; the fluff comes from Clara's captured note and the old template. The slip, graded by the harness:

   What went out

   Subject: Dora Gomez — Gomez called to ask about
   qualifying for one of the properties…
   
   Property: not identified yet
   Dora called. The office was closed (after hours).
   Dora Gomez called to ask about qualifying for one
   of the properties. She has a couple of questions
   about qualification and would like to speak with
   someone over the phone. Callback number confirmed
   as the number she is calling from, ending in one
   two nine one.
   Can someone call Dora back? (970) 201-1291

   What it should say

   Subject: Dora Gomez — does she qualify?
   Call back (970) 201-1291
   
   Dora Gomez (new prospect, not on file) wants to
   know if she qualifies for one of your rentals
   and has a couple of questions.
   Please call her: (970) 201-1291.
   Left Oct 27, 5:07 PM (after hours).
   
   — and for a resident/applicant —
   Maxwell R. (applicant, unit 515, Camellia) wants
   an update on his application. Call back (…) …

   Rules: who (and who we know they are from our record: resident at which property and unit, applicant for which unit, prospect, vendor; Fede 2026-10-27: "if it's a tenant we want to know their property, unit etc"), what, number, when; the number once, in digits; no "called to", no restating the callback; no "Property: not identified yet" (on a company line Clara asks which property); nothing about our own process.
3. One record, one owner. The row is written when the message is captured, state pending. Its only end states are delivered (provider id on the row), held (a visible hold, on Slack and the conversation page) or declined (a named product rule, shown on the conversation page). One sweep retries anything not delivered, bounded, then holds it. A retry of the same event re-attempts a row that is not delivered. Call end never disarms the backup.
4. Honest surfaces. The mid-call tool, the Slack card and the conversation page read the row's state. "Note left for the office" can only render from a delivered row.
5. Wording is a grade, never a gate. Banned words, length and grounding become a harness score on the capture prompt, reported on a page. With no model in the send path, every delivery rule is a deterministic test that runs in CI.
6. Same rule for texts and emails. When Clara's reply to a text or email fails or is held, the person's own message still goes to the team (it is already extracted verbatim). Staff never see our error; they always see the message.

A caller with no dialable number gets a note anyway, with "number withheld". One of the two Oct 8 "no number" losses had said and confirmed her number on the call; the engine refused anyway. Two more silent branches from the second pass: a transfer that went to voicemail counts as "handed to a person" (4 of the 8 losses), and a backstop logs "nothing was lost" seconds before the call-end step refuses the note.

Extends the existing scripts/team-notify-replay (no new harness dir; registered on the Atlas). Fixtures: the 8 real lost messages, scrubbed; ~30 delivered ones that must stay delivered; one synthetic fixture per silent branch (16 found) with its fault injected: model throws, model returns banned words, function dies after the row write, SendGrid fails, thread write fails, no team email. The one existing test that expects zero notices when the model throws is flipped to "known gap". Mutation control: re-enable the gate, suite goes red. Runs on the subscription; the hermetic fault-injection subset runs in CI.

Then a real call on the Willows line, a message left, the email opened, the row and the page checked. Live everywhere on merge (fix; Fede 2026-09-30).

• Oct 8 and Oct 15 Western Slope callers: 12 to 19 days old; recommend no late note. Say the word if you want them sent.
• Situs: 3 silent calls, no caller message ever; one office email on Oct 20 came from our own test-fleet number and reached their real inbox with no record kept (a leak, fixture in the harness). ConAm: no voice calls; its engine switch was unset until Gera removed the switch on Oct 27 (#22894), so the engine now covers every company.
• Written forwards repeat: 4 notes in 63 minutes about one caller; the AppFolio robot's own email forwarded twice as a question. In the harness as duplicates fixtures.
• Both inspections (first pass: 16 silent branches, every lost row; second pass: every real note graded, each fluff phrase traced to file:line, numbers re-checked): attached to PR 1.

• Asked: one sentence plus the number. Shipped: up to 3 tries × (write + check) = 6 to 10 model calls, ~35 banned patterns, 613 lines.
• The session's own replays dropped real messages three runs in a row (1→3, 0→4, 4→0: #16229, #16532, #16562); each time one check was patched instead of the design questioned.
• No reviewer asked for any check; the bot warned the other way (digit ban burns retries on a reference number; "call back" ban rejects real asks), left as suggestions.
• Session 003's memory promised a "redrive" of dropped notes. Never built. The caller's words are still computed on failure, then thrown away.
• Unseen 23 days: failures went to an engineering alert only (paging #alerts from Oct 15); the Slack card said "note left for the office" either way; the test proved the rule, not the outcome.

Rule added: a wording check grades a message, never blocks it. A caller's message always reaches the office; "no email" only when there is no message. A session offering Fede A/B on a customer-facing drop must include "send it plainly" as an option.