Willows Production Readiness — the Aug 19 Verdict

Written the evening of Aug 19, 2026, after a full day of live manual QA (Fede), an overnight+daytime fix campaign, and the first end-to-end production test harness. Companion pages: Overnight Campaign · Fair-Housing Architecture (parked).

Verdict: READY for staffed production at The Willows, with two watch items.
Every defect found today — five by Fede's hand-testing, four more by the new automated journey testing — has a fix that is merged, or approved and merging tonight. The launch bar (below) is met once tonight's merge queue drains and deploys. The two watch items are quality-trend items, not blockers.

The launch bar (adopted today), graded

Bar	Status	Evidence
No dead ends — every message gets a real response, every promise has an owner and a deadline behind it	MET*	Wrong-address replies answered; closed matters answerable; blank replies root-caused (a safety check could ship an empty message — fixed at 3 layers); the "team will be in touch" felony dead-end now structurally impossible (a code guard opens a tracked question or removes the promise — replay over 31 days of real traffic found 6 customer moments it would have fixed). *Two fixes are approved-and-merging tonight; "met" is true at deploy.
Copy at the coworker bar — everything Clara sends reads like a person	MET*	First-person acks live; decision emails distilled; hold notices rewritten in coworker voice (3 review rounds); staff briefing-voice ("Ok say that…") unwrapped before relay; taught policies distilled into professional standalone statements with "decided by X on date". *Copy + distillation PRs in final review round tonight.
Fair-housing safety net live	MET	Live in production and proven by Fede personally: "no we dont" held, the compliance-override pushback held, the compliant rewrite relayed and taught. Service-animal over-verification fixed and eval-pinned. Screening context now includes prior holds.

What shipped today (all live or merging tonight)

LIVE Fair-housing tripwire — staff answers screened before any person or policy sees them.
LIVE Service-animal rule — no fees, no proof demands; regression-eval-pinned.
LIVE Staff analyst toolbox — any staff member, any channel, real answers; resident-record writes blocked from escalation-answer lanes only (a PM's own work untouched, pinned by the 16-case incident-replay suite).
LIVE Thread-splitting fix — one person, one conversation; the split hit hardest mid-escalation.
LIVE Blank-reply fix — no safety check can ever ship an empty message.
PROVEN Voice callback — the day's headline: Clara now calls a voice-caller back with the staff answer. Proven with five real phone calls; three designs failed on contact with a real line (a 15-second silence timer was killing delivery) before the final shape: the answer rides the first sentence, voicemail gets a short close, text always sent. Final transcript: "here's the answer to what you asked us: … I've also sent it to you in a text." → silence → "No worries — you've got the answer by text too. Take care!" Enabled on the bench Willows; production enablement is a one-command flag.
MERGING Coworker-voice notices + briefing-voice unwrap (4 review rounds) · Policy distillation ("ya thats fine no fee for esa" → titled, professional, attributed policy) · Escalation question scoping (the tracked question is the ask, never a transcript dump) · Staff metrics tool ("what's our occupancy?" → "As of today, 93%").

How it was proven

The journey harness (merged): the whole funnel run as one continuous story in production — prospect asks, staff refuses illegally, pushes back, answers in briefing voice, prospect returns — every message judged as its recipient, plus absence-checks ("did that promise get an owner?"). It found two bugs no scenario test could see, before any customer did.
Real inputs: today's live session phrasings are permanent test cases; ~450 real Camellia emails were mined for the tricky-scenario taxonomy (the corpus itself stays out of the repo — real residents' stories don't belong in version control; a fully synthetic twin set is the follow-up).
Real calls: the callback was only believed after actual phones rang — the rule that caught what every simulation missed.

Watch items (not blockers)

Item	Why it's a watch, not a block
Paired-name fairness probe flagged one uneven run (different offers to "Emily" vs "DeShawn")	Single observation on live generations; needs a repeated-trial study before it's a finding. Scheduled.
One email, one format	Escalations still notify both the office inbox and the owner, in two formats, with the routing token visible in the subject. Cosmetic but user-facing; next copy pass.

Needs Fede (two small things)

Staff registration for your email address — one prepared command registers fede@propflowai.co as staff in both orgs so your emails to Clara are recognized everywhere (your Camellia occupancy email went unanswered for exactly this). The rest of the team roster should follow the same path.
Camellia stays in shadow per your call — revisit when the Willows bar has held for a week.

Numbers of record: 15 PRs shipped today across the campaign; 9 production defects found and fixed same-day (5 by hand, 4 by the harness); 6 real customer moments in the last 31 days that tonight's promise guard would have fixed; 5 real test calls; 1 sensitive-data near-miss caught and contained before merge.