Overnight ship — vendor calling, 2026-07-29
One-night summary of the vendor-comms push that started with the Willows PO live test on July 28. Everything below is shipped, measured, and receipted; links to the deep-dive pages at the bottom.
The headline
The redesigned vendor Clara is live in production on the real Willows line (+1 844 285 3526).
Against the 50% goal: the recognized cohort's 26% is
below it; across all 66 replayed calls the transfer rate is 65%, above it. The 50% target is a production metric
(GET /api/vendors/adoption) and has not been re-measured live against this build. The
accounting-disclosure count is for unprompted disclosures; the post-ask case is covered by the separate
23-case triage suite (23/23), not by the corpus replay.
What a vendor now experiences
- Greeting leads with the likely job: "Hi, is this Miracle Method of Denver? It's Clara at The Willows — are you calling about the tub and shower walls refinish at unit 204?"
- "Is it approved?" → "You're good to go — that PO covers it." Never a word about internal accounting state (the leak was found on a live call, fixed, and fenced at three layers — prompt, context render, and a deterministic grader).
- "Front desk?" → one helpful offer anchored to their job, then an immediate graceful transfer if they still want a person. Never a menu ladder.
- PO number stated on ask with unit and work description — payload inside a helpful answer, never a list dump.
Outbound: calls lead with the PO, or don't happen
- Live: when the PM approves external work at a purchase-order property, Clara opens the vendor call with the PO ("I've got PO 90760 for you — the tub resurface at unit 101…").
- Proven on prod: with no PO in hand, the call is refused — session requeued, and the PO-required notice was composed and durably logged (delivery blocked at the test-recipient guard, so no human was paged). The same scenario was proven twice, minutes apart, on the same vendor/property/turnover.
- The prompt-sync tool (#4786) erased a 1,012-character pre-#4782 drift between the repo and the live outbound agent, found in its dry run; the actual apply landed after #4782 merged and pushed the post-#4782 prompt — a 3,756-character correction (characters, not bytes). Live prompts are now verified against the repo after every sync.
Safety catches worth knowing
- A bot re-review caught a seed-data fix that would have silently re-armed automated dialing to a real person's cell (the demo vendor's contact is Fede's number). Real vendors are now excluded from bench keying by default; arming is an explicit typed opt-in.
- The turnover fabrication guard caught Clara inventing an "access error" explanation on a failed dispatch — the guard class works.
- Review-process hazard found: the recheck workflow can carry a green review verdict onto new commits without re-reviewing. Flagged for a repo-level fix.
Found in production, carded for next
- Vendor recognition misses — A&K Appliance has nine labeled calls on the same number (seven of them before the 2026-07-28 corpus refresh) and an open PO 667 on unit 210; on a separate call the same day, that line was told "I don't have you on file as a vendor with us yet." Upstream resolver work; no replay gate can test it (top next item).
- PM disambiguation deadlock — two open turnover threads with no labels make Clara's "latest walk or the earlier one?" question structurally unanswerable (card PWdEH5LX).
- One-offer-before-transfer fires inconsistently (one call transferred while holding PO 646 without offering).
- Post-transfer call recording is captured inconsistently; transfers can land in a voicemail menu with no recovery.
Open for Fede
- Review queue: #4798 (seed fences), #4785 (replay harness), #4776, #4759, #4756, merged ADR #4720 text.
- Two Phase-4 decisions: vacant-unit auto-book autonomy; occupied/vacant flag source (see the decision brief).
- The together-test: Clara calling Fede's cell with the PO-lead opener — armed as a deliberate two-step (re-arm + approve), runs on his word.
Deep-dive pages
- Before/after — 66 of the 69 labeled real calls replayed /a/vendor-before-after-2026-07-28
- Retro (why the basic mistakes, why iteration was slow) /a/retro-po-live-test-2026-07-28
- July 22 brief scorecard (17/27 achieved) /a/turnover-brief-scorecard-2026-07-28
- Phase-4 scheduling decision brief /a/phase4-scheduling-decisions
Corrected 2026-07-29 after a source audit.