Everything I could decide myself is decided. What's left needs
you. Each one carries the recommendation of the session that raised it — if you
agree, pick it and press Done. A model's opinion, where there is one, appears on the card
as a receipt line; cards with no receipt were never put to a model. "I'm not sure" is a real answer and becomes work for me.
You don't need to open the session — pressing Done sends your answer back and it picks up.
1maintenance-eval score collapsed 17/21 (09-22) -> 10/21 (09-23) -> 8/21 (09-24). Every newly-failing scenario is a text/SMS conversation where create_work_order must resolve a tenant/occupant identity (all 9 tenant-* presets, plus handyman-create-served-occupied/relay/reverse-relay/close-with-photo, which anchor to an occupied unit). Every scenario NOT needing occupant resolution (ask-no-unit, ask-no-location, out-of-scope-refusal, create-served-VACANT, all 4 voice presets) still passes. Confirmed via Temporal workflow history for tenant-oven-create (maintenance-386ce97a-8c01-45d4-9f0b-2233eafc5d34): gatherActivity found workOrderCount=0 at debounce-flush time, meaning create_work_order was never actually committed during the live SMS turn -- the break is upstream of the debounce/dispatch pipeline, in the Inbound Dispatcher's Sonnet intake rung or the work-order handler's identity resolution, on Vercel (no CloudWatch trace; Vercel function log drain is disabled per #11222). Ruled out: PR #11655 (work-order home resolver) merged 2026-09-24T11:25Z, four hours AFTER today's eval run (07:19-07:44Z) -- can't be the cause, and its gating setting conversations.join_across_homes has no row for org_sandbox (reads off) anyway. The regression window is 2026-09-22T07:44Z -> 2026-09-23T07:44Z. Candidate commits in that window I checked and ruled out (files don't touch the create/identity path): #10984 (work-order BY-ID routes fail-closed), #11023 (org-scope predicate refactor, 8 unrelated routes), #10745 (lapsed-role-row fix, currently dormant/no writer). Did not rule out: #10878 (co-tenants resolve the primary's lease), #10756 (a known tenant never gets a ProspectInquiry/guest card), or ~20 other commits in that window not diffed line-by-line. The eval bench is org_sandbox/appfolio-45, but the code path is shared with every real customer -- this may be silently dropping real tenant maintenance requests since 2026-09-23.
In plain terms. Should an engineer be paged right now to check if tenants texting about repairs are silently not getting a ticket filed?
The session raising this recommends: Page someone now workOrderCount=0 was confirmed via real Temporal history, not just the eval's own grading -- if this reproduces on any real occupied property, tenants texting about repairs are being silently dropped, which is worse than a failing eval and time-sensitive Tried first, unsuccessfully: Read the pointer rows for 3 nights to isolate the regression window and separate it from a pre-existing unrelated handyman-relay bug; pulled real Temporal workflow history for the failing conversation to confirm the ticket was never created (not just a grading bug); checked org_sandbox's cross-home setting (off) and the newest work-order PR's merge time (after the run) to rule both out; read the 3 most identity/work-order-shaped commits in the window and none touch the create path; did not have Vercel function-log access to find the exact refusal, and stopped short of a full manual bisect of ~40 commits given the bounded run