Vendor Calling — Turn-On Readiness Scorecard

Overnight measurement campaign, 2026-07-30 → 07-31 · all numbers measured live unless labeled estimate · The Willows bench (appfolio-45) only; no Camellia data touched

Bottom line

The full production loop is proven live end-to-end: review queue → approve → real dial → conversation → transcript → outcome → work order scheduled → calendar mirrored → callback recognized → reschedule applied. The night's open question — IVR menu navigation — is now answered and bounded: the agent cannot press mid-speech (a hard ElevenLabs platform limit, proven across every config), but presses correctly at the first silence 9/9 times. What remains before turn-on is merging tonight's four fix PRs, the arm decision, and a supervised first call.

Scorecard

Lane	Before (start of night)	After (measured)	Verdict
Real-job E2E (schedule → callback → reschedule)	Untested as one loop	3/3 legs PASS live — real WO scheduled by a call (firm Aug 4 2:00 PM), vendor callback recognized by phone with correct PO cited, reschedule moved WO + calendar event in place. Call QA 4/5, 5/5.	PASS
Review-queue chain (UI → approve → dial → record → transcribe)	Assumed built, unproven	Spine proven at every hop with record-level evidence (adversarially audited). 10 defects found, ranked; top 4 fixed same night (PRs #5114, #5115, #5117 + prompt fix).	PARTIAL → fixes in flight
Off-phone behavior battery (23 gating scenarios vs live agent)	16/23 (two consecutive runs)	20–22/23 stable, 23/23 peak across 32 runs. Every targeted scenario flipped green: overnight-hour sanity check, closing discipline, wrong-number etiquette, hold acknowledgment, name-mismatch echo, PO spoken in reason-for-call, no-narration rule. Live agent verified byte-identical to merged source.	PASS (87–96%)
Continuous-menu IVR (the Royal Textile failure)	0/3 real probes ever pressed; cause unknown	Cause isolated + bounded: 9/9 pressed correctly at first silence (78.9–83.7s on a 78s menu) over SIP; mid-speech press impossible on any config.	BOUNDED
DTMF transport	Presses inconsistently heard by real IVRs (in-band)	SIP trunking delivers RFC 4733 that Twilio Gather detects, every run. PoC rig standing.	SOLVED (migration decision pending)
Human-detection safety	Instant-hangup rule, small sample	10/10 clean human aborts across two sweeps (4–27s, zero words spoken to humans, zero forbidden presses incl. explicit emergency-option refusals).	PASS
Verified caller ID (Camellia office line)	Unproven concept	Full chain proven with test numbers; Camellia runbook written (one 6-digit code keyed at the office). Resolver works unchanged.	READY

The turn-taking answer (the night's centerpiece)

Every prior press-failure against real IVRs traced to one cause: ElevenLabs never gives the agent a turn while the far side is speaking. Turn detection is a probability threshold on VAD silence — no silence, no turn, regardless of configuration.

9 live experiment calls, 8 distinct turn configs (silence mode, eager eagerness, 1s turn-timeout, turn_v2, speculative turn, initial-wait, kitchen-sink all-at-once, agent-owns-opening-turn): all pressed at the same moment — the instant the 78-second menu audio stopped (78.9–83.7s).
The LLM is not the limiter. The keypad tool fires 1–2s after the first turn opens, with the correct digit reasoned from menu content, every time.
Control: a short menu (speech ends ~14s) pressed at 21.6s — press latency tracks end-of-speech, not a clock.
Practical consequence: the shipping default config is already optimal. Real menus pause and replay (as tonight's real-world sweeps show — Redi Carpet and Piper were navigated successfully), so the agent navigates them one menu-cycle slower than a human, reliably. A hypothetical zero-gap looping IVR would never get a press; none has been observed in 90+ real business calls.

What real businesses actually do (94 dials of ground truth)

Finding	Evidence
~53% of vendor-type businesses answer with a human during hours	81-business daytime sample
Even at 10:30 PM, ~50% of trade contractors put a live human on the line (answering services, forwarded cells)	Sweep 2: 5 humans in 10 calls — "after-hours = unstaffed" is false for this vertical
Gate navigation works on cooperative IVRs	Redi Carpet: clean 2-layer navigation to voicemail; Piper: correct option, emergency options explicitly declined
In-band DTMF is unreliable per-platform	Allstar + Drain Pros menus replayed verbatim after presses; Miracle Method ZIP gate consumed presses fine → SIP migration is the fix
Dead ends exist that no config solves	Roto-Rooter: pure marketing loop, no menu; Colorado Conveyance: mailbox full — needs re-dial cadence (built, #5063) + eventual cross-channel fallback

Defects found and fixed tonight

Defect (all live-proven)	Severity	Fix
Callback context pairs one work order's job with another's time — Clara read a nonexistent slot aloud and the vendor confirmed it	High	PR #5115
`dispatch_work_order` silently binds a different vendor while telling the PM "unassigned" (record self-contradicts)	High	PR #5114
PO number in payload but never spoken on a purchase-order-mode call (vendor invoices against it)	High	Merged (prompt PR) — spoken once in reason-for-call, hard anti-fabrication rule, never leads the call
Agent spoke a narration turn aloud into a real call ("(Menu has finished. Pressing 3…)") despite tool-speech config off	Medium	Merged (prompt PR) — no-narration rule
PO cache 24h staleness cliff → bogus "PO required" hold for a PO that exists	Medium	PR #5117
E2E rig reported false green by matching stale queue rows	Medium (test infra)	PR #5117
Truncated PO fetch marked real POs removed (wiped bench read model mid-run)	Medium	Merged (#5111)

Still open, ranked lower: PM notification unproven on bench (email sink), bench units don't exist in AppFolio (PMS-write leg untestable there), audio endpoint ignores Range requests (seek likely broken), stale robo-vendor PO scenario, PO approvalState unfiltered. Details in the QA audit memory/report.

What turning on requires

Done overnight: all four fix PRs merged (#5114 dispatch mis-assign, #5115 callback coherence, #5117 PO freshness + harness, #5119 prompt hardening) and the live agent re-synced to source.
Gates removed (Fede ruling, 2026-07-31): the arm flags are deleted — every vendor-assigned turnover dispatch now creates a review-queue item on every property, and the human approve click in /review is the only gate before a dial. An unwatched queue stalls work rather than dialing wrongly.
Fede decisions: SIP-trunking migration go/no-go (transport fix, changes how all EL calls are placed) · schedule the Camellia office 6-digit caller-ID verification · supervised first real vendor call.

Sources: live E2E agent report (4 calls), review-queue QA workflow (5 agents, adversarial audit), turn-taking experiment log (10 calls), sweeps 1–2 (23 businesses called of 130 targeted), simulate-battery logs, verified-caller-ID PoC, SIP PoC. All raw call IDs and DDB keys preserved in session records.