Clara in the renewals meeting

Western Slope's weekly renewals meeting, 11:33 to 12:01 Mountain. From context (the transcript carries no names): one person shared her screen, another dictated the clicks, Jay's prices were relayed, Fede listened and answered one question. Granola's title reads "Pomona and Athalia properties"; "Athalia" is the transcriber hearing "AppFolio".

Questions Clara could have answered in the room (tier 2, voice): what did we offer this tenant and when; is he 6 or 12; has anyone at Pomona replied; who is out for signature and how many have signed; who is on a housing voucher; is 3138 Perkins' ETA updated; did Zander get contacted. Every one is a read on records Clara already mirrors or can read through the robot.

Fede, Oct 27: "there are some dumb things they are doing… let's add some judgment on best practices and AppFolio and help them. For example you could be using AppFolio defaults and not have to uncheck things one by one." Checked against AppFolio's help articles (numbers below are at knowledge.appfolio.com/en/articles/…), the domain notes and Colorado law. Five of eleven habits end with a one-time setup or one written rule.

Transcript caveats: four tenants were named for reminders, not six; the zero-balance remark may not have been about Kenneth; "5/29" was called both the start and the signing date; nobody called the dated month-to-month line a fee. Not found in Colorado law: a 60-day notice at the end of a written lease, or a 90-day voucher rule (the only 90-day rule is the no-fault non-renewal notice under HB24-1098). 2025 changes not checked. Sources on the page's research record.

Two read-only dry runs of the renewal decision engine against Western Slope's real October list (8 leases; writes blocked, no AppFolio calls, model on the subscription), once with Granola's recap and once with the raw transcript. Then each decision against the code.

Why only 8 leases: 12 were left off as "already offered in AppFolio" (the log gives the count, not the names; most likely Kenneth, whose offer went out during the meeting, and the six out for signature), January leases are outside the window, and Renewals is off at Lincoln and Courtyard. So the homes the meeting spent most time on were most likely never in front of the engine. The summary-first failure is the design lesson: the raw transcript is the input, never the recap.

Also exposed: staff typed "original lease start" as the day the lease was signed (5/29/26); recipe v3 uses our lease record's start date. Those can differ. Lane 11's shadow replay (test:harness:renewal-replay) was not run: it issues up to 90 reads against Western Slope's AppFolio.

Verdict (Oct 27 evening, from the evidence, not the meeting)

Fede: "the staff is making it up, they don't have a standard policy." So the recipe was checked against our own record: the 11 finished offers the learner read on Oct 20, the Oct 24 form pre-fill read at 18 occupancies, the Willows proofs, and a fresh read-only look at Kenneth's sent offer and his form today.

What follows. Two fixes, both fix-shaped, no switch: the engine builds the addenda list from empty (remove every pre-filled part except the pinned one), proven once at the Willows by parking an offer whose form pre-fills an extra addendum; and the fee date as a separate step (new end + 1 day), which needs a robot verb to edit the lease's month-to-month line, named here, not built yet. The better help for the office is upstream: set the per-property renewal defaults in their AppFolio so the form pre-fills 39 alone (section 1b, habit A), which ends the manual clearing for them and for Clara; their AppFolio config is their and Fede's call.

Fede, Oct 27: "I would like all the core infrastructure to be the same regardless of the mode." The mode only changes how the raw record is captured. Everything after the first box is one code path, and the Oct 15 ruling already said so ("isn't it the same engine?").

flowchart TB
  subgraph capture["Capture · company setting: meeting mode"]
    direction LR
    N["Notes · silent bot · transcript + summary"]
    V["Voice · same bot speaks · answers questions"]
    F["Video · Tavus face · (60-min plan)"]
  end
  capture --> R["Raw record: transcript, recap, screenshots · → company brain, Internal"]
  R --> E["Renewal decision engine · per-lease decisions, quotes, who said it"]
  E --> C["Email within the hour: 'Here is what I heard' · reply to correct"]
  E -. unknown unit, contradiction, loose instruction .-> Q["One question back to the person who decided"]
  C --> P["Prepare: packets parked in AppFolio, · offer texts and reminders drafted"]
  P --> Rv["Email: 'N leases ready for your review'"]
  Rv --> S["Office presses Send"]

A screen for it, later

Nothing is built, by Fede's Oct 4 ruling ("I don't want to build a UI yet"). The brain page already holds four mockups (#ui: Settings page per audience, company-wide Internal list, Internal panel on a tenant's page, renewal view "what Clara will do and why") and three placements for notes (#notes-ux). The renewal view (mockup D) is the on-screen twin of the "here is what I heard" email: same per-lease rows, same quotes, same reply-to-correct. When a screen is wanted, start there, with the meeting's transcript and screenshots attached to each row.

The one fact that reshapes this: Western Slope's Tuesday 11:30 meeting ("WSPM Rent Roll Management") is a Microsoft Teams meeting on Western Slope's own tenant, no phone dial-in (calendar invite, Oct 27). Other customers will be on Zoom or Google Meet. The pipeline after capture (raw transcript → company brain → decision engine → emails) is ours in every case; this section is only about how the transcript gets to us.

What the join layer must do

1. Name the speaker on every line, so a decision carries who made it (owner, manager, assistant). Fede, Oct 27: without names "we can't distinguish between ICs and managers or owners, or who made the calls."
2. Work without Fede in the room, on whatever platform the customer uses.
3. Ask the customer for as little as possible: one click by the organizer beats a tenant-admin change.
4. Leave a door to voice later: Clara answering "did they reply?" in the room needs live audio in and out.
5. Store the raw transcript, never a recap (global rule, Oct 27).

What "build" really means

Not a bot of our own. Zoom's Meeting SDK page says it "is designed for human use cases and does not support bots or AI notetakers" and points to RTMS (verified Oct 27). Google's Meet Media API is in developer preview and "no longer accepting new signups". Microsoft's real-time media bot needs Windows .NET hosting and Microsoft says it is not recommended for AI agents. So "build" = wire each platform's native transcript (Teams Graph, Zoom RTMS or cloud recording, Meet REST), one small integration per platform, post-meeting only. "Buy" = a meeting-bot API that joins by link on all three and can also carry audio.

Verdict

Names first, vendor last. For Western Slope the cheapest path with names is the one they already own: a staffer presses Transcribe in Teams. That is a build row with almost nothing to build. A bot vendor earns its place only for the voice tier (Clara speaking in the room), which no native path gives, and for customers whose admins will not grant transcript access. Pick the vendor after Nov 3, on evidence: whether Teams names come through, and whether Recall's per-platform speaker pages exist (two were 404 on Oct 27).

Three checks before any purchase: (1) open today's Teams meeting chat and look for a Recap or transcript; if Western Slope already transcribes, the named transcript of today's meeting exists and the native path is proven at zero cost; (2) run one transcribed meeting in our sandbox tenant and fetch the VTT through Graph with the app credential in Keychain, which settles names and permissions without touching a customer; (3) re-fetch Recall's speaker-timeline and Teams-lobby pages.

Recall.ai proof of concept (Oct 27 evening, on Fede's go): done, no human in the loop

Before: Granola on Fede's Mac, transcript labeled only me / them, nothing saved anywhere Clara reads. After: a bot joins the Teams meeting from an API call, the transcript carries the speaker's real name, and the raw transcript is a meeting note in the company brain. Every step ran headless with no person involved; the "human" in the room was our sandbox test user driven by a script.

Two traps for the build: the bot's display name arrives with lookalike (Cyrillic) letters in Teams' screen text, so never match the bot by exact name; the lobby Admit control is not a button, find it by its text. What a product version needs: a "meeting bot" ingest route that takes Recall's webhook (bot done → fetch transcript → save meeting note with the speakers kept raw) behind the company's notes-intake setting, then the renewals decision engine reads the note (hook question in the inspection below). Cost at one meeting a week: free tier covers the POC; production pricing on the grid above.

Sources: Zoom Meeting SDK Linux page; Zoom RTMS docs; Google Meet REST and Media API guides; Microsoft Learn calltranscript-get, application access policy, manage-external-bots and meeting policy reference; Recall.ai meeting-platforms and output-audio pages (fetched Oct 27); Meeting BaaS, Skribby, Nylas, Attendee, Vexa, Tavus and ElevenLabs public pages. Haiku-gathered numbers (Tavus plan minutes, ElevenLabs tiers, Zoom RTMS per-minute) are estimates until re-read. Screenshots: Granola's API returns this meeting's attachments as empty; a bot recording or Teams recording carries the shared screen.

• Notes and transcript: solved today. Fede's Granola API key (Keychain) returns the note with its 300-line transcript and Markdown recap. A poller on the Mac mini already fetches every Granola note within 5 minutes and posts it to #transcripts. Adding a second leg that writes the same text into the company brain as a meeting note for the matched company is a small change to that poller, no new system.
• Screenshots: manual for now. Granola's API has none; a Teams or bot recording carries the shared screen.
• This meeting is saved locally (recap, transcript, note id not_agff32wxc15512) and is the first entry for the brain once the leg exists.

Deep read of every renewals design page against the code at the tip of main and Western Slope's production settings (Oct 27 evening, read-only). Most of the machine is built. Two arrows are missing, and one rule in the code is wider than what Fede said.

The loop, with the two broken arrows

flowchart LR
  C[Clock: daily scan,
90 days out] --> R((Lease's
renewal))
  M[Weekly meeting →
raw transcript → brain] -. never reaches .-> R
  L[Monthly list reply] -. only parked texts .-> R
  T[Tenant text / email / call
to Clara's line] --> R
  A[Tenant or staff text
in the AppFolio inbox] -. filed, never signals .-> R
  S[AppFolio change seen by sync] --> R
  R --> P[Packet parked:
"N leases ready"]
  P --> O[Office presses Send] --> G[Tenant signs,
Jay countersigns] --> K[Reminders until signed]

Tenants and AppFolio already reach a running renewal. Team decisions and AppFolio-inbox texts do not. A meeting decision lands in the brain, which nothing reads; a list decision sits on a row only the send gates and the next scan read. So the Prepare rung either parks a packet 90 days early with policy terms, or holds it forever waiting for a decision that never arrives.

The hook: one renewal signal, then look before acting

Design: one intake call, postRenewalSignal(lease, {source, evidence, decision?}), used by every source (list, meeting, AppFolio text, sync). It writes to a per-lease, per-cycle decision record (quote, speaker, history kept) and signals the lease's running renewal. It never starts one: the daily scan stays the only starter and reads the record at start. On every signal and every reminder wake the renewal runs one "look before acting" step: read the record, AppFolio's offer state, the newest office-side message on the thread, and tenant replies on any channel, then pick one of rebuild the packet / hold or stop reminders / skip the cycle (team has it) / wrap up. This is Fede's rule: Temporal is the alarm clock, Clara checks the state of the world. Two engines, two hooks: the words-to-decisions engine runs when a raw note lands in the brain (never on the clock); the per-lease act / skip / stop logic runs inside the renewal on every signal and timer.

PR chain, in order: (1) decision record + signal + the look-before-acting step with packet re-entry, first producer the list, proven at the Willows (fix) · (2) AppFolio inbox filer and sync writer post signals (fix) · (3) a brain reader runs the engine on new notes and writes "proposed" decisions, staff-only (fix) · (4) the "here is what I heard" email with reply-to-correct (new capability, company setting, default off). No branch is opened until Fede buys the design (D8).

Correction: the tenant rule in the code is wider than Fede's ruling

The AppFolio inbox mirror stops every tenant's text before Clara can draft ("tenant thread, never a turn", merged Oct 21, quoting "i didn't want that"). The raw transcript (session 007, Oct 20, 9:20 PM) shows the ruling was about switching Western Slope's leasing texts to the AppFolio line that night: "we don't want to mess with maintenance or tenants right now", "just leasing". The sentence "a tenant's reply to a renewal text is also not a turn" was the builder's. Before and after, Fede asked for the opposite under supervision: Oct 19 "if someone responds… supervised mode… I would have to approve each one"; Oct 20 "if the tenant replies through AppFolio, we capture the intent"; Oct 21 to Gera "once we're in the inbox and dialed, that also unblocks renewal"; Oct 23 supervised "for renewals, tenants, maintenance, any topic… nothing ever goes out unless I approve it"; Oct 25 "start replying to tenants in the shared inbox, for maintenance or renewals… supervised mode… see what Clara would have replied, on Slack"; Oct 27 he rated Clara's silence on Rachael "Wrong". The shadow-week cards in Slack are that observation feed: 121 tenant texts at Western Slope since Oct 16 and 41 at Situs since Oct 22, Clara answered none, staff answered about 70% (median 8 min at Western Slope, 61 at Situs), 11 have no reply at all.

The fix (D7): a tenant's text on a renewal thread is both a renewal signal and a turn for Clara when the Renewals module is on at that company; Supervised holds whatever she drafts. Renewals-only, riding the existing module switch, so zero effect until Fede turns the module on: fix-shaped, no new switch. Maintenance is untouched (Gera's maintenance tenant door already covers it and is off; widening generically would start held repair drafts at Western Slope tonight, where maintenance is already Supervised).

Top gaps (full list and evidence in the fold)

1. The Prepare packet is never rebuilt after the office decides (one-shot build; held packets stay held). Fix.
2. Team decisions do not reach running renewals: monthly list vs weekly meeting, an approved list takes no edits, leases already offered are off the list. Fix (the hook).
3. No teammate check before a renewal send; Bev's texts never stop Clara. Fix.
4. A tenant's AppFolio reply does not stop Clara's reminders. Fix.
5. The brain is write-only: Western Slope has 12 notes (newest Oct 10), none from the Oct 20 or 27 meetings, nothing reads them. Fix.

Goal Fede set: run the whole mechanism on next week's meeting, hold everything for review, send nothing to a tenant. What has to be true by 11:30 Mountain on Nov 3:

1. Capture: D2; Granola stays on as the fallback.
2. Into the brain: poller leg writes the note (recap + transcript) to Western Slope's company brain. Fix shaped, no switch.
3. Engine: the dry-run reads the brain note for that company and produces the per-lease table. Misses from section 2 become battle-test cases first.
4. Confirmation email: "Here is what I heard" to Kat, Jay and Bev within the hour, with reply-to-correct. New capability: company setting, default off, Fede turns it on for Western Slope. Nothing to a tenant.
5. Prepare, parked: for each confirmed decision, Clara parks the offer at the Prepare rung (needs gates 3 and 4 from Oct 24: one park and cancel on their test property, then Fede's press) and drafts the texts and reminders as held messages.
6. Review email: "N leases ready" with one link per lease. The office presses Send in AppFolio.

Steps 1 to 3 are proof work and can run on Oct 27's recording this week. Steps 4 to 6 wait on Fede's presses named in section 9.

Nov 3 dry run: what happened

Both are right, and both match Oct 27. What the engine missed:

• The free half month for "these three" landed on no lease. Clara would ask which three.
• Most of the renewal talk (countersigns, reminders, Kenneth through March, McKeever listed twice, Xander) was about leases already offered in AppFolio. Those are not on the list the engine reads.

Applications (the 2980 Emerald background check) and maintenance (pre-move-out inspections) came up too, so the extractor needs topics beyond renewals. Next PR: screen frames, to resolve "this one".

The Oct 23 deck (ws-renewals-review) stands; these edits bring in what we saw.

• Slide 2, today: replace "10 minutes of clicking, your team's words" with the measured round: one offer, 7½ minutes, three people; 3 minutes to find what was offered; six reminders by hand.
• Slide 4, template: insurance email: not used. Addenda, month-to-month fee date and a second term: pending the recipe investigation, not from one build.
• Slide 7, level 1: add "Clara sits in your Tuesday meeting, takes the notes, emails back what she heard, prepares every packet and draft, you press Send".
• New slide after 7: "Questions Clara answers in the room" (the list in section 1), as the step after notes.
• Slide 10, start here: unchanged properties; add "decisions you make in the meeting live with Clara, so they are never misremembered" (the office has no standard policy to encode; Clara proposes the default, you confirm).

Proposed — pending review  Fede, 2026-11-03: "we've overbuilt and I want to just get something out"; "the team initiating the engine, not Clara on her own"; "what is the smallest possible thing that adds value and is not intimidating to start". Research: the Oct 27 and Nov 3 raw transcripts, the Sep 30 and Oct 14 calls, every renewals page and the engine code, Fede's rulings since Sep 28.

The customer already described the model. Jay, Oct 14: "getting Clara being like, hey, here's what I want to send out on Thursday, and getting that on Tuesday morning … using that as our fodder for the meeting … kick that back to Clara, and she's good to go." Two moments a week, nothing about the meeting changes (Fede, Nov 3: no pre-read; the meeting is the starting point, Clara sends nothing before it):

Meeting  Clara listens; the team decides; Kat keeps negotiations
After    Clara ──► "here is what I heard": done & reported · drafts for your OK · questions · blocked
         Team  ──► reply "approved"  ──► Clara does the work on the normal rails from wherever it is

Retro in one line. The core works: Camellia end to end since April, the packet proven through a tenant signature, the meeting engine zero wrong calls. The overbuild is around it: seven gates before one text, 14 send paths that each check a mode, 5 company settings + 5 building fields + 3 env switches, 21 pages. Fede's Temporal read is right on two specifics: a counteroffer ends the workflow (renewal-workflow.ts ~2296) instead of handing off and coming back, and a team decision only counts before a renewal starts (renewal-auto-start.ts 1956–1969), never for one already running.

Smallest first step (two PRs, timers untouched). (1) On each new meeting note, run the engine over all open leases, offered ones included, and email Kat, Jay and Bev "here is what I heard", one task per lease, reply-to-approve. New capability, default off. (2) For each approved "prepare", park the offer unsent and email "N ready for review". Then the usual path: an offer that is out gets reminded until signed, whoever started it, with the clock set from the real send date (start at day 45 if that is where it is; take over one a person began).

Local proofs of concept, no PRs (Fede: "just to get vibes"): the after-meeting email rendered from the Nov 3 transcript over all open leases; a "take this over" table for the 13 already-offered leases. Files and the deck: Western Slope deck, Nov 4 version.

Status 2026-10-28: Fede read D2, D7, D8 and the Supervised question in session and will decide at the end of the week (Oct 30–31). Nothing is built or opened until then. Gera's lane census (Smith, Oct 28): Supervised now holds every resident message, the % rollout has zero plans in production, no renewal-signal hook exists anywhere; D7 touches Gera's inbox door and would go to his side through the bridge.