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.
1tenant-resident-services nightly eval: the neighbor-noise case ('upstairs neighbor blasting music until 2am') has failed 2/2 recent runs (2026-09-11, 09-13; 100% on 09-06). Clara replies 'I can't handle neighbor disputes directly, but I'll get this to the property management team' with tool sequence [none] — no forward_to_property_manager call, which capabilities/resident-services.ts:42 ('Promise-must-be-backed-by-a-tool-call') forbids. Eval floor 94% on 8 cases, so one miss fails the section. Cause not pinned: capability changes in the window are #7152 (renewal.ts, 09-06) and #7220 (balance copy, 09-07); neither touches the noise-complaint branch. Fix = a Clara prompt/behavior PR at proof-ladder rung 2 (multi-turn eval + resident-services eval run), HIGH risk per the decider tier, not for an unattended session.
In plain terms. Clara is telling residents she will pass a noise complaint to the property team without actually doing it — should we fix her behavior now as a resident-facing regression, or park it until the next resident-services prompt work?
The session raising this recommends: Fix Clara now — open the prompt PR with evals, treat as a resident-facing regression A resident told 'I'll get this to the team' when nobody is notified is the exact trust failure the rule was written for, and it has reproduced 2 of 2 runs — not a flake. Tried first, unsuccessfully: Read the failing case's grading output (forward_to_property_manager did NOT fire, tool sequence [none]); confirmed the prompt rule that forbids prose-only forwards; checked every capabilities/ and eval change between the 09-06 pass and the 09-11 fail (#7152 touched renewal.ts, #7220 the balance tool copy — neither the noise branch); confirmed it is 2/2 reproducible, not n=1. Stopped there because a Clara prompt change is HIGH risk and needs an attended owner, not an alert decider.