The July 22 turnover brief — scorecard, one week on
Source: Fede's dictated brief "Turnover workflow — vendor detection, scheduling, and outbound calling" (July 22, 2026; session steered through Saturday July 25). Voice-dictated — quotes below keep the speech-to-text artifacts. Scored on July 28 against what has shipped, what is in review, and what remains.
Legend
- ✅ Achieved — shipped and verified
- 🔵 In flight — built, held on review or live proof
- ⬜ Pending — not started or design-stage
The scorecard
| # | Fede's ask (verbatim, condensed) | Status | Where it stands today |
|---|---|---|---|
| 1 | "for turnovers. It's gonna be the next big priority in the product" | ✅ Achieved | This week's entire push (PO pivot, inbound redesign, outbound PO-lead) is this priority |
| 2 | "gather context… every ADR. And find any gaps… to get this to fully launched" | ✅ Achieved | ADR-0116 (vendor job reference) wrote up the architecture, gaps, phases; published on docs site |
| 3 | "I want to see if there's already a har[ne]ss… if there's no har[ne]ss, we're gonna build one" | ✅ Achieved* | Multi-turn gauntlet + full-corpus replay harness exist (PR #4785). *Built last instead of first — the retro's core finding; the brief called for it on day one |
| 4 | "Go replay. Every single test call we've had and the real calls. To kind of measure where we are right now in terms of quality" | ✅ Achieved | 66 of the 69 labeled production calls replayed call-by-call (the 3 others are 2 emails + 1 SMS); published as the before/after page (recognized-vendor transfer 56% → 26%) |
| 5 | "how well we're capturing the information while we're reasoning about. What the issues are" | 🔵 In flight | Graders measure fabrication, disclosure, transfer shape; information-capture quality per se is partially covered |
| 6 | "How we manage expenses. So like charging tenants. Putting them in their ledger" | ⬜ Pending | Phase-4b spend attribution — not started; also gates switching turnover minting from work orders to POs |
| 7 | "correctly identify vendors. And if those vendors can be matched with the vendors in the system" | ✅ Achieved | PO→job binder: high-tier precision 4/4 = 1.00, coverage 14/15 = 0.93 on the hand-labelled fixture (one-click yield 27%); pinned floors 0.98/0.85, with a PM-confirm lane for the rest |
| 8 | "the concept of a prefer[red] vendor… like miracle cleaners is a prefer[red] vendor" | ✅ Achieved | Built and shipped: resolvePreferredVendors resolver, per-property go-to lists as the default vendors view, trade classification + backfill (PRs #4370–#4510 range). Unit-tested; referenced by the maintenance/turnover eval gauntlets |
| 9 | "we've already done that through the email… if not, find the cloud session" (top-10 vendors) | ✅ Achieved | Recovered; vendor roster data and the AppFolio PO corpus mined (all POs, per the PO-pivot brief) |
| 10 | "make sure we have the right contact information… look for emails. From them. To build like a test suite" | 🔵 In flight | Contact info verified for pilot vendors; the email-based test suite not built (email ingestion pending, see 13) |
| 11 | "really being aware. Of when a vendor is calling us" | ✅ Achieved | Caller-id vendor recognition live; personalization delivers vendor, mode, and open POs to the agent |
| 12 | "aware. That there's either a work order. Or. Something scheduled they're coming to do" | ✅ Achieved | The PO pivot's heart: Clara knows the open POs and states the right one with unit and work description |
| 13 | "we have an email from them with a schedule… start. Detecting those emails and ingesting and reasoning" | ⬜ Pending | Vendor email ingestion not built |
| 14 | "look at the transcription from this week on slack" | ✅ Achieved | Consumed during the context phase |
| 15 | "put them in the calendar so the property management. Team have that. But make it not block. The tour" | ⬜ Pending | Phase-4 scheduling; 3 design decisions still open |
| 16 | "architecture and harness first… writing down the whole. How the whole machine is gonna work" | ✅ Achieved* | ADR-0116 is the write-down; harness exists. *Sequencing was violated in practice — see retro |
| 17 | "The last one's gonna be. Voice. So calling vendors. After a turnover" | ✅ Achieved | Sequencing followed: voice work came after architecture, data model, and binder |
| 18 | "inspection → I need vendor X, y and z → text the property management team confirmation → make an [outbound] call which is probably a new agent we gotta build" | 🔵 In flight | Built (ADR-0059 P1) and upgraded with the PO-lead opener; PRs #4786 and #4782 are merged (2026-07-29); live proof still outstanding |
| 19 | "attack it and stress test it. As a first pass. Find bugs. And have like a baseline" | ✅ Achieved | Robo-call batteries + gauntlet; production baseline measured (58% transfer / 17% Clara-resolved on the 69-call corpus; 61% transfer on the 66 replayed voice calls) |
| 20 | "try to schedule the work and explain what the work is about" | 🔵 In flight | The PO-lead opener does exactly this ("I've got PO 90760 for you, the tub resurface at unit 101, hoping to get it on your schedule") — merged (#4782); live proof outstanding |
| 21 | "on behalf of. The property manager by name and the property and… probably say that we are an AI assistant. To start" | 🔵 In flight | On-behalf-of naming: in the opener. Explicit AI disclosure: NOT in the opener — open decision for Fede |
| 22 | "if… we can't. We would transfer to the human. For help" | ✅ Achieved | Transfer rules in place; today's fix makes Clara offer help once before honoring a person-ask — never a gate |
| 23 | "deep research on the web of. How these phone calls. Actually happen in real life… build expertise on that domain" | ✅ Achieved | pm-domain-knowledge corpus + the mined production call corpus; standing rule: all eval coverage derives from it |
| 24 | "the goal would be. Placing those calls out to schedule jobs" | 🔵 In flight | Merged (#4782, #4786); one live proof away |
| 25 | "if they call back. Having that context of what was said. Having the context of the schedule" | ✅ Achieved | Callback sessions carry context; polish bug found July 28 (stale session outranked the PO book in the greeting) — fix on the branch |
| 26 | "they need to cancel, reschedule. We should handle that. If they have questions, we should handle that" | ✅ Achieved | This week's happy-path work: reschedule, ETA, PO-for-invoice, approval status — measured better-than-production on the replay |
| 27 | "I want to focus on that in the next week" | — n/a | The week is up; this page is the accounting |
Reading of the week
17
achieved
6
in flight
3
pending
1
was the deadline itself
Of 27 asks: 17 achieved, 6 in flight (all on the outbound lane or measurement depth), 3 pending, 1 was the deadline itself. (Corrected 2026-07-28 late: the preferred-vendor roster was already built and shipped — the first published revision wrongly listed it as pending.) The pending three cluster into two themes: money (expenses/ledger, and the WO→PO minting switch it gates) and channels beyond voice (email ingestion, calendar).
The sharpest lesson: item 3 and item 16 — harness and architecture first — were in the brief on day
one and were honored last. The full account of that is the companion
retro.
Corrected 2026-07-29 after a source audit.
Open decisions for Fede
AI-assistant disclosure in the outbound opener— DECIDED 2026-07-28 (Fede): no proactive disclosure — but disclose honestly if asked. Clara opens with her name, the PM, and the company only; if a vendor asks whether she's an AI, she says yes.- Phase-4 scheduling design decisions (2 genuinely open: vacant-unit auto-book autonomy; occupied/vacant flag source. The third — pull outbound voice into Phase 4 — was effectively decided by building it).
When to build ADR-0060 preferred-vendor roster— RESOLVED: already built and shipped (see item 8).