Situs Group — What the Records Show (2026)

Doors come from the unit directory; occupancy from the occupancy report, which covers 341 of the 397. Alton is listed there as a single unit, so it carries no occupancy figure — a question for Eileen.

What we can't see: whether AppFolio's occupancy and days-vacant fields match reality. Eileen: she has walked into units marked vacant and found people living in them.

Priority is a channel-permission gap rather than a habit: the 0% holds at 0.0% in business hours and 0.0% after hours for both phone and portal, so neither the answering service nor a resident can set severity in AppFolio at all.

The assignment lag is an hours effect, not a channel effect — every channel lands at 11–16 hours after 5pm — so replacing the answering service alone would not move it. The call centre's 1.0-hour headline is a business-hours artefact.

Rework is the one measure where the call centre is systematically worse. On apartments alone the portal is the more after-hours channel of the two (46.6% against 36.2%) and re-assigns slightly more often (31.6% against 29.9%), so duplicates rather than re-assignment are what separate them. The portal's 10.7% online-update rate against the call centre's 2.2% points at callers being unable to see the ticket they already have.

Abandonment sits with the bots and two staff creators (25.0% and 18.6% stale at 30 days, against 0.4% and 0.0% for two others), not with the phone.

Contact-pinning error rates are 8.2% call centre (n=364), 10.0% staff (n=10), 12.5% bots (n=8), 0.0% portal (n=6) PARTIAL. The three small denominators establish no difference between groups.

Open work orders quiet for 30+ days: 135 of the 205 quiet for a week; none of them from the call centre. MEASURED

"…using the answering service/feature for after hours / weekend inbound calls / leads that would allow us to retire our MCC line and keep our main line available 24/7… after hours Clara answers all the calls and ports out through the tree when an emergency, etc. arises." — Hugo, email of 27 Aug

"After-hours maintenance goes to the MCC — an offshore call center paid for through AppFolio." — Eileen, 1 Sep. On a live test call the agent could not match "yellow jackets" to the existing "bees" work order, took 8+ minutes, and logged it against the wrong property. That call was on a single-family house, so it sits outside this page's scope — the behaviour still stands as evidence.

Hub page: published office hours Mon–Fri 9–5, closed weekends — 128 of 168 hours a week unstaffed.

What we can't see: call recordings, total call volume, calls that ended without a ticket, the vendor's contractual SLA, and whether the answering service is able to set urgency in AppFolio at all.

What we can't see: the phone. 173 call events across all 658 apartment leads, 7 of them after hours. Nothing here measures the voicemail channel.

Each recovered lease is valued at one month of rent. Excluded: the VA hours (658 hand-keyed leads, 351 confirmation texts, the scheduling events), listing spend, the vacancy days behind a 12.1-day application clock, and the phone channel. The conversion uplift is a correlational reference class, not a causal estimate; Situs's own speed-versus-outcome curve is not monotonic.

• Answer every lead within a minute, 24/7, in English or Spanish.
• Book the tour inside the conversation instead of a 28.9-hour booking lag and a confirm-or-deny text.
• Chase the silent applicant on a fixed cadence.
• Escalate complaints and "can I speak to a person" rather than leaving them unanswered.
• Write the tour note, a real loss reason, and flag a returning phone number.

AppFolio stays the system of record; every action is written back into it.

Prerequisites. 19 of 42 leasable vacant units are simultaneously vacant, rent-ready and advertised with a rent — the verified set she may speak about. Language tagging first, per Eileen. Build dark, prove on a test property; switching Situs on is a separate decision.

"My rent is supposed to have the second half of my lease renewal incentive applied to this month. I signed my lease last month with a $1,000 incentive… My rent is showing 1250 and should only be 750 this month. I don't want to receive a notice for being late. But that is the reason it's late." — tenant, 6 July

"The one that I think would get the hugest ROI… would be the renewals." — Eileen, 1 Sep

What we can't see: the renewal incentive tracker (a spreadsheet outside AppFolio), any recorded reason for a rent cut, and how fresh "market rent" is — 7 upcoming leases carry none and 12 show rent under $300.

Re-pricing the 25 month-to-month tenants at 3% is $11,938/yr INFERRED. The rent floor plus the market gap together are credibly $150K–$200K/yr INFERRED; the two are not additive because some leases appear in both.

Phase two, in order: the floor gate first — no offer below current rent, and none built from an empty or zero rent field. Then a priced offer generated on a schedule with the concession on the offer rather than in a text thread. Then one owner for the target expiry window: Hugo wants renewals steered into spring, Eileen wants November–February penalised, and spring 2027 already holds 87 expirations. Accounts in active collections receive no offer.

Bots open 229 of 400 turn tickets, 137 of 167 common-area and 126 of 166 exterior. The call centre concentrates in plumbing (174 tickets) and appliances (76).

What we can't see: technicians text residents from personal phones, so "3 of 1,884" measures the system of record, not resident contact. Vendor bills sent to an AP email rather than AppFolio are invisible, so $115,072 is a floor.

212 of the 248 flagged billback tickets carry $0; at the $360 median vendor bill, closing those is ~$76K/yr INFERRED.

1,560.3 unpriced hours at $50–75/h is $78K–$117K/yr INFERRED; no labour rate is recorded anywhere in the dump.

Eileen's own estimate — roughly half of the work orders a year should be billed back and fewer than 100 are — is hers, not ours, and includes owner-billable work. Much of the total is owner-side capital work, and Colorado limits deposit charges.

Inspection photos and per-item condition ratings: the scrape captured report structure, not images. Whether move-out statements were produced outside AppFolio. Which of the 44 blank-status reports are abandoned rather than mis-saved.

Eileen's ask: an AI first pass on move-out inspections — flag what's missing, email the tech for it, reconcile charges against the ledger, prepare the move-out statement. She rates move-outs and billbacks her second-highest return.

INFERRED: days vacant ÷ 30.4 × rent, taking the first of scheduled, advertised, new or last rent; units with no rent contribute $0, so the figure is a floor.

The 7 rent-ready unlisted apartments carry $10,425/month of scheduled or advertised rent. Cutting the apartment median from 49 to 40 days across the 92 apartment turnovers is worth roughly $44K/yr INFERRED.

Whether a unit is unlisted deliberately: AppFolio has no such flag, so a redevelopment hold and an oversight look identical. One scrape of the availability page, so no history of how long a unit sits before it is listed. Zillow and Tenant Turner listing state is not visible, only AppFolio's posted-to-internet flag that feeds them.

"AppFolio 'posted to internet' flag generates the property integration ID that feeds all ILSs — unposting a unit kills the feed even on paid contracts." — Eileen, 1 Sep

48 of 92 turnovers record no cost. At the $1,755 mean of the 44 that do, that is ~$84K of turnover spend never reaching a billing decision INFERRED — smaller if some were genuinely no-cost turns.

If half the 20 turnovers with no end date are genuinely open at ~$1,700 rent, that is ~$16K/month of ambiguous vacancy INFERRED.

Orca went live the day of the meeting, so none of its data is in this dump: everything here is the pre-Orca baseline. The vendor rate card is outside AppFolio. Labour, purchase-order, billable and inventory columns parsed empty on every turn row. Inspections carry no type field — move-out scope versus occupied check is inferred from which question template was used, and only a handful of distinct unit tokens could be recovered from the 80 scope inspections, so the 42-of-92 match is a floor. Inspection records carry no property link, so only the turn side of this match is scoped to apartments.

Correction to carry. The failure-rate-by-sender table is unusable: AppFolio strips the author from a message once it fails, so the unnamed sender bucket is exactly the failed set and every named sender reads 0.0%. The overall 8.2% rate stands; a per-person rate is not derivable. MEASURED

Failed-text topics: other 602 · maintenance 506 · money 246 · bulk notice 209 · showing 6 · renewal 10. After a failure, a text followed 19% of the time and an email 41%. MEASURED

Email delivery: 17,248 opened (61.3%), 1,085 hard failures (3.9%). MEASURED

What we can't see: about half of resident communication runs on two employees' personal phones and is not exportable, so every response-time figure measures the visible half. No carrier error code is captured, so a wrong number, a landline and a carrier block are indistinguishable. Whether the 61.3% open rate reflects image-proxy prefetching.

"Sending everything in both English and Spanish offends some tenants… Figure out the language tag first, then all templated comms respect it." — Eileen, 1 Sep

Only 5 rows belong to former residents who have moved out ($13,512). 5 balances are under $500. 7 payment plans exist only as note text. MEASURED

What we can't see: what the Ananeo Colorado LLC concentration represents; why the two AppFolio reports disagree; legal status — no eviction filings, demand letters or notices to pay or quit appear in the dump; the terms of the 7 verbal payment plans; cash and money-order payments (30 mentions across 21 residents).

The rent roll cannot be audited against reality from a CSV. Language detection is keyword-based on message text, not a declared field. The tenant-directory match rate is 557 of 753 names, so untagged counts are floors. Note timestamps are post or last-edited times, not the date of the event described; 150 notes carry a pre-2026 date with a body citing a 2025–26 event.

"Your data is only as good as the people who are updating it." — Eileen, 1 Sep

"~50% of tenant communication happens on two employees' personal cell phones — $50/month allowance, no visibility, no way to bulk-upload their texts." — Eileen, 1 Sep

"The moment that they find out that it's going to record them, the whole system would collapse." — Eileen, 1 Sep

Techs are 1099 workers who invoice by text photo. A tech was set up as a "vendor" in AppFolio purely so the VAs could text him.

What we can't see: the shadow channel itself. Everything measured here is what leaked into AppFolio; the ~50% figure is Eileen's estimate, not ours. No MCC call logs, no Crexendo records, no Orca data, no spreadsheets, no tech text threads. 51% of notes are truncated, so keyword detection under-counts. An earlier claim that 11 outbound texts publish a staff cell number in the signature is not reproducible and is withdrawn.

A record written after the fact is a summary of a conversation, not the conversation. Payment plans, notices to vacate, maintenance commitments and renewal decisions agreed by voice reach AppFolio days later or never, so an agent can contradict a deal a human already made. For residents in the habit of texting a personal mobile, an agent's number is the second channel and silence on it is not silence. Staff resistance is a design constraint: Eileen is explicit that anything reading as surveillance of the on-site team gets routed around.

• Window: 2026-01-01 to the 2026-09-01 capture. Pre-2026 matches appear only in an "earlier history" column and are never added to a 2026 count. Undated records are excluded from both.
• Corpus: 152,955 staff-authored statements from tenant SMS, tenant email, tenant notes, guest-card messages and notes, application messages and decision reasons, and work-order remarks. 84,346 fall in the window; restricted to the 16 multifamily apartment properties and to human authors, 52,992 remain. 36,078 were dropped as commercial or shell (24,109) or single-family (11,969).
• Notes were split out of AppFolio's stacked panel into individual notes with their own author and date — 2,087 notes, 1,217 in 2026. Earlier analyses treated a whole panel as one undated note.
• Assistant Lead is a shared VA login counted as staff. System, PM Assistant, Maintenance Performer and Customer Service are bots and are excluded, as are work-order records themselves. The scope is the 16 multifamily apartment properties only: placeholders, commercial, single-family houses and entity shells are excluded from every count.
• Three counts per rule answer different questions: messages (how often it was said, blasts included), residents (how many distinct people heard it), distinct wordings (how many genuinely different statements sit behind the volume). One blast can show thousands of messages and a handful of wordings.
• Property is attached from each resident record's own unit link, and for prospects from the guest card's property cell. Roughly one message in ten could not be matched to a property and is excluded from the multifamily scope.
• Values are pulled only where the amount is tightly bound to the fee it names ("$50 late fee"), and matches next to balance, prorated or reversal language are dropped. This undercounts rather than inventing a policy out of an account balance.
• Conflict verdicts come from a split test: two values are a genuine inconsistency when the same resident was told both, or the same staff member said both at a shared property in overlapping weeks; they are per-property when the two values share no property.
• Quotes are de-identified in place — phone numbers, email addresses, street addresses and greeting names replaced. Situs staff names are kept, because who said what is the point of the ledger.
• Limits: keyword selection can miss a rule worded in a way nobody searched for, and can over-select a topic, which is why every money topic also reports how many messages actually state a value. Nothing was verified against AppFolio, the phone line, or any document Situs holds outside this export.
• Full ledger with verbatim evidence: ~/situs-data/analysis/POLICY-LEDGER.md. Tables under ~/situs-data/analysis/tables/policy_*.csv.

17 distinct addenda types. The Alton and Bryant Gardens master leases carry no renter's-insurance obligation, no month-to-month fee, no security-deposit amount, no parking addendum and no smoking addendum; all six Alton documents are unsigned, and their notice to vacate is 90 days against 30 everywhere else.

13 tenant: 3-Day Notice · Renewal Notice Letter · Renewal Notice Letter (Section 8) · Non-Renewal Letter · Non-Renewal Letter (Spanish) · Rent Increase · Deposit Disposition Refund · Deposit Disposition Amount Due · Security Deposit Interest · Tenant Payment Information · Tenant Unpaid Charges · 24/7 Maintenance Support Options · Arizona Welcome Letter (English).

6 applicant: Adverse Action Notice · Approval · Approved with Conditions · Denial · Partial Report Denial · Pre-Approval. 3 owner: ACH Setup for Distributions · Owner Bank Account Information · Updated Address for K-1s.

Three carry conflicts. The Rent Increase letter is the unmodified AppFolio default and says rent is due the first of the month; every lease says the 3rd. The Renewal Notice Letter states the $1,000 credit as two $500 payments. The 24/7 Maintenance letter promises "up to a week for non-emergency repairs" — a service-level promise stated in no lease and never quoted to a resident.

Arizona (2) and 6908 Lowell (4) fall short because that is all that exists — every other occupancy at those two properties has only cancelled documents, verified by exhausting the list. Of 64 policy values recovered from documents, 28 confirm what staff say, 19 contradict it, and 16 are written rules staff have never mentioned to a resident.

Replay-ready — an inbound turn that is substantive (more than 3 words, not a bare acknowledgement) and carries an individual human reply: 809 resident texts MEASURED, 47 resident emails MEASURED, and the prospect turns INFERRED by applying the measured 88% answer rate to the non-acknowledgement turns. Total ≈1,502, against 42 rows in today's real-replays-email-sms.yaml.

Verdicts come from a read-only grep and case count over 74 dataset files and 60 promptfoo configs. PARTIAL — a suite could cover a situation under wording not searched for.

Counts MEASURED; the consistency column is INFERRED from the spread, since some variation may be legitimate per-property difference. Six rules were quoted at different values to different residents inside the same year: referral credit, renewal bonus, free-rent weeks, lockout charge, pet fee, partial payments.

Limits. The phone leaves zero transcripts. Email is one-way: 317 inbound of 42,746, of which 95 got an individual reply, so any email eval rests on 95 rows. 60% of the 943 notes posted this year have no usable body. One client, one market — these are Situs's policies, not portable answers. The human reply is a reference, not a gold answer: 22% of the time there was no answer, and the corpus contains 22 delay apologies. Broadcast detection is a heuristic PARTIAL.

Where these differ from the other sections: the 22.3%-never-answered figure elsewhere uses "any human reply within 14 days"; this 22% uses "an individual reply within the next 3 messages". Quote one definition, not both.

Step one needs no client data and no permission: author roughly 40 cases across the four Missing gaps — entry permission, the insurance and liability charge, the pest-prep fee decision, and a blast swallowing a live question. These are Clara gaps regardless of Situs.

Step two needs no data: turn the 16 policy rows into a decision page for Situs, six of them live contradictions.

Step three is the 120-row replay suite, stratified toward the four Missing and six Thin situations plus 30 Spanish rows and 30 mid-conversation continuations, deduped on timestamp and first 120 characters, scrubbed with a hand-verified substitution table, graded by rubric against a property fact block on a subscription runner. It is the only step gated on the consent answer.

1 · Entry permission and a maintenance revisit — resident, SMS, English

Inbound: "Hello! I tried calling and couldn't get through. Can you come by Monday afternoon?"

Human reply (Assistant Property Manager, 21 minutes): "Hi Noel, sorry we missed your call. Yes, we sure can arrange for Monday afternoon. I will let the technician know. Thank you so much!"

Clara: "Hi Noel, thanks for getting back to us! I'll pass along your request for Monday afternoon to the maintenance team so they can confirm the technician's availability and get this rescheduled. Per the lease, you'll get 24 hours' notice before they come by. If you don't hear confirmation soon, feel free to call the office at 555-0100 during business hours (Mon–Fri, 8am–5pm)."

Verdict — acceptable-different: Clara relays the request but hedges on confirmation; the human confirmed Monday afternoon directly. Both factually clean.

2 · Notice to vacate and a month-to-month fee dispute — resident, SMS, Spanish

Inbound (abridged): "Hola buena tarde estoy escribiendo para dejar constancia que llamé e hice saber q ya no renovaré contrato… me dicen q tenía q avisar con 30 días antes… me dice la chica q como son 5 días después q aviso q tendría q pagar el mes a mes y no estoy d acuerdo."

Human reply (Assistant Property Manager, 63 minutes): "Hola Juan y Ximena, Por favor envíen su aviso formal de que no renovarán su contrato de arrendamiento a leasing@example.com. Una vez lo recibamos, les enviaremos las instrucciones para su mudanza, con la expectativa de que desocupen la unidad a más tardar el 31 de agosto."

Clara: "Ximena, lamento mucho su pérdida y los problemas de salud que ha enfrentado… para que quede oficial en el sistema necesito que envíe su aviso por escrito a leasing@example.com. Sobre el cargo de mes a mes y su desacuerdo con los $150, entiendo su punto, pero esa decisión no la puedo tomar yo — voy a escalar su caso al equipo de la propiedad… Y sobre el depósito: por contrato tienen hasta 30 días para procesarlo y 60 días para la disposición final una vez que se desocupe."

Verdict — acceptable-different: Clara directed written notice correctly, acknowledged the dispute, escalated rather than deciding it, and stated deposit timelines accurately.

3 · Availability and pricing — prospect, SMS, English

Inbound: "I would like information about a one-bedroom apartment"

Human reply (leasing, 66 minutes): "Hi! Thanks for reaching out… Here's a link to one of our available 1-bedroom apartments… We're currently offering 6 weeks free on select units."

Clara: "Hi Rowan! We have a great 1BR/1BA available — 620 sq ft at $1,299/month. It's a cozy, well-priced home close to the RTD stop on W Alameda Ave. Would you like to see it in person? We still have that tour slot open for tomorrow at 11am if you'd like to take it!"

Verdict — acceptable-different, human was wrong: Clara gave accurate unit details and a tour offer; the human stated "6 weeks free", which is NOT ESTABLISHED in the property facts.

Names, phone numbers, email addresses and balances are substituted; the mapping stays on local disk.

Individual blocking facts named most often: security deposit amount (5 rows), application screening timeline (5), rental application link (4), application approval status (3), Section 8 voucher acceptance policy (3).

The human reply is a reference, not a gold answer, which is why every row is graded against the property facts as well as the transcript.

• Clara's model: claude-sonnet-5 on all 2,114 rows, resolved at run time from production resolveDraftParams(channel) in conversation-manager.ts, not hand-picked. Judge: claude-sonnet-4-6, the pinned EVAL_JUDGE_MODEL, so scores stay comparable with the repo's other evals.
• Prompt read at run time from evals/promptfooconfig-leasing-response.yaml; the persona block is swapped for a resident persona on resident rows.
• Replay tier: in-process prompt replay. The full-pipeline harness was deliberately not used — its entry point can dispatch a real send, and this is a third party's residents. Nothing was sent.
• Corpus: 2026-only inbound turns that received an individual, non-broadcast human reply. Co-tenant duplicates removed on timestamp plus first 120 characters; bare acknowledgements removed.
• Scrubbing: every resident and prospect name in the whole dump — not just the thread's own participants — plus every phone, email and account balance, replaced through a deterministic local substitution table. The mapping stays on local disk and is never committed.
• Topic tags come from the same deterministic regex taxonomy as the corpus analysis, with no model call, so stratification is exactly reproducible.
• Property facts were built per property from the AppFolio property directory plus the resident or unit record. Rules staff quote inconsistently — referral credit, renewal bonus, free-rent promotion, lockout charge, pet fee, payment-plan authority — are written as NOT ESTABLISHED on purpose, so a reply that invents a number is caught as a failure. That is why missing-fact is 33%: the fact block was deliberately thin. Run 2 tested the obvious next move — a richer fact block — and it did not shift the headline; see below.
• Billing: run on the Claude subscription via OAuth. No metered API key was exported at any point; chooseCredential refuses a metered runtime.
• Cost and time: 40 minutes wall clock for 2,114 rows, two model calls each (draft plus judge). 14,346,910 input tokens, 1,176,898 output.

Sorted by B-vs-before change. A topic under 30 rows is marked thin n and is not a headline. Only three topics at n ≥ 30 improved at all, and every one of them sits at or inside the 0–3 point noise floor.

By language: English 1,480 rows — 55% before, 54% A, 56% B. Spanish 145 rows — 59% before, 54% A, 59% B, with wrong-tone-or-language rising 3% → 6% → 9%.

What the 158 B regressions became, and where they sit. Remember the control: 39 rows broke with nothing changed at all, so a large share of this list is judge churn rather than damage.

One example of each failure type, with the judge's own words:

• Wrong fact — correct → wrong-fact: "Clara claimed the date was 'left out of the English version by mistake,' but the English message clearly states 'Tuesday March' — the date was left out entirely, not just from the English version."
• Missing fact — "Clara couldn't confirm the human's key point that the last proof was outdated/expired, weakening her response compared to the human's."
• Wrong action — "Clara confused Kai as an existing resident rather than a new applicant and failed to confirm receipt or next steps."
• Hallucinated policy — "Clara states the portal accepts any amount the resident enters, a claim not established in the property facts."
• Wrong tone or language — "Resident wrote in Spanish but Clara replied in English, violating the stated policy of answering Spanish-language messages in Spanish."

The full 158-row list for B and the 184-row list for A are in ~/situs-data/analysis/replay/RESULTS-KB.md.

No property-level conflict survived curation: the handful the miner flagged turned out to be one-off statements about a single resident's account, not two standing answers to the same question, so they were dropped rather than resolved. B's addition is exactly these six portfolio rows.

• Same rows. The exact multifamily row set from run 1, selected by id — 1,625 rows across 16 apartment property records / 397 units. No row was re-selected, re-scrubbed or re-built. Single-family, commercial and placeholder rows excluded.
• Same models. Subject claude-sonnet-5 at effort high, resolved from production drafting; judge claude-sonnet-4-6 pinned. Same judge prompt, with the knowledge base appended to the facts the judge is told Clara held — without that, a fact she now legitimately knows would be graded as a hallucination.
• The knowledge base is built only from data already mined: the 23 confirmed stable rules in the policy ledger, the AppFolio unit directory, vacancy detail, itemised rent roll, showings export, and staff statements mined from the message corpus. No new AppFolio access, no network call.
• Staff statements were filtered to standing facts after the same substitution scrub as the rows — dated closures, outage notices and one-off prorated amounts were dropped, because loading an event as a policy would measure the curation rather than the idea.
• Block size: 13 per-property blocks, 1,042–1,496 words each (1,636–2,349 tokens), each sent as its own cached system prefix. Summed across every row served that is 3.36 M tokens under A and 3.64 M under B — what it would have cost without caching.
• The human baseline does not move — it is a property of Situs's 2026 record, not of Clara. The judge flagged the human reply as wrong on 26% of rows before, 28% under A, 29% under B; the human reply never changed, so that drift is the honest noise floor on how precisely the other deltas can be read.
• Concurrency capped at 8 on one subscription seat; resumable, checkpointed every 25 rows. Nothing was sent to any resident, prospect or member of Situs staff at any point.

All three ran on the Claude subscription seat through OAuth — no metered API key was used or exported at any point. Full detail: ~/situs-data/analysis/replay/RESULTS-KB.md.

propflowai-pms-agentic-layer is a frozen snapshot, last commit 2026-04-14 against the live repo's 2026-09-01. Three findings sourced only from that tree are already fixed and must not be actioned.

agents/clara/lib is a second @/lib/* root: tsconfig.json maps the alias to both src/lib and agents/clara/lib. A sweep scoped to src/lib alone misses the files holding the largest prompt findings. The live runner is ~/.claude/appfolio-browser-agent, audited separately.

Language — the full picture for a ~70% Spanish-speaking portfolio

The conversational lane is bilingual: detection, formal usted, mid-conversation switching, a stored-preference ladder, Spanish maintenance vocabulary with English enums preserved, and sendEmail that throws without a resolved language. The gaps are in templated sends and voice specialist config.

An explicit import-chain check found zero production imports from evals/, fixtures or scripts/: no fixture leaks into a production prompt or seed path.

The eval corpus is Camellia-shaped: 145 Camellia mentions in leasing-response-quality.yaml, 95 in renewals.yaml, 46 in knowledge-grounding-gauntlet.yaml. Two are hard gates rather than context — clara-application-link.yaml asserts contains: "https://www.camelliadenver.com/#floor-plans" deterministically, so 100% of its cases fail for any other property, and leasing-response-quality.yaml:216-249 grades against Camellia's real fee schedule ($38 application, $250/$100/$35 pet, $80/$95 parking). Both need re-authoring, not re-parameterising.

clara-sms.ts and clara-text.ts compose the persona with no language rule, but every importer is a test file — not live.

Three absences: no end-to-end Spanish conversation eval; no multi-property disambiguation eval; no call-centre-mediated intake eval, though every Situs harness models direct tenant-to-Clara contact.

• The last ~90 lines of the voice personalization route, and full verification of its ~100 dynamic variables.
• clara-turnover.ts and clara-delivery.ts — checked for Camellia, JP and Denver literals, none found, but not read end to end.
• Collections, renewal-policy and turnover prompt bodies were sampled rather than read line by line.
• The ~30 sibling appfolio-browser-agent-* worktrees, and the runner's workflows and discovery module.

All 70 findings with file:line: ~/situs-data/analysis/CAMELLIA-ASSUMPTIONS-AUDIT.md.

153 leads · 109 showings scheduled · 51 showings completed · 57 move-ins · 50 move-outs · 1,107 work orders (32% of the year) · 84 renewals effective · 63 applications.

Vacancy was the year's worst: a 99-day median, p75 189 days, and 14 of 49 re-let units (28.6%) empty more than 180 days. Work orders peaked in September (401) and leads bottomed out in November (29). Duplicate work-order rate 12.3%, against 10.8% in Jan–Aug 2025.

Window totals, Jan–Aug 2025 → Jan–Aug 2026: leads 423 → 637 (+50.6%) · showings scheduled 302 → 407 · showings completed 156 → 226 · move-ins 106 → 98 · move-outs 96 → 87 · work orders 2,446 → 2,280 · after-hours work orders 830 → 656 · renewals effective 177 → 149 · applications 188 → 248.

Bolded months are the four 2026 has not reached. Applications come from a report with no property column that sees only applications reaching AppFolio — TurboTenant, Tenant Turner and AffordableHousing.com sit outside it.

Move-in rate by arrival, in-hours / after-hours: 25.8% / 23.8% in 2025, 19.7% / 14.1% in 2026. Answered within 1 hour rose 39.3% → 49.3%; waited over 24 hours fell 27.1% → 22.8%.

On voucher leases AppFolio prints the tenant's portion in "Previous Rent" and the full contract rent in "Rent", which fakes a $1,000–$3,000 increase. Leases with a previous rent under $600 are dropped: 16 in each window. The DU placeholder removes a further 19 rows in 2025 and 5 in 2026.

Priced leases: 110 of 177 expiring in 2025, 93 of 149 in 2026. Median rent change +1.75% → 0.00%; p25 0.0% → −2.86%. Median among increases +$50 in both years; among decreases −$150 → −$50. Total monthly lift across priced leases $7,510 → $4,640. Outcomes: renewed 66 → 60, month-to-month 72 → 55, did not renew 3 → 7, cancelled by user 36 → 27. Offers sent 188 → 170, accepted 89 → 91, countersigned 78 → 75, reminders 367 → 403.

Renewal lead time is 71 → 72 days on a symmetrical audit-log definition. The 106-day figure quoted elsewhere came from a 2026-only artefact counting any touch including texts; the two are not interchangeable.

AppFolio's unit-vacancy report is a snapshot of the capture day and no 2025 version exists, so vacancy is rebuilt from the calendar feed: every move-out is paired with the next move-in at the same unit across the union of both years' feeds. Cohort = move-outs in the window; move-outs with no later move-in are right-censored, excluded from the median and counted separately.

Censoring biases 2026 downward. 34 of 87 Jan–Aug 2026 move-outs have no move-in yet, against 0 of 96 in 2025 — and the censored ones are the long vacancies, so 62 days is a floor. A matched Jan–Jun pair gives 76 → 62 days on 68 and 66 move-outs, still censored at 16.

Vacant over 60 / over 90 days: 72 / 38 in 2025, 29 / 7 in 2026. Mean days 83 → 63. Commercial office, split out and never mixed in: 16 move-outs at a 181-day median in 2025, 19 at 106 days in 2026.

Make-ready is a second, independent measure from the unit-turn report, move-out date to turn end: median 63 → 32 days, p90 104 → 62, turns over 60 days 62 → 9, over the 18-day target 95.8% → 81.9%. It measures getting the unit ready, not re-letting it.

1. Vacancy, occupancy, delinquency and unpaid balances as AppFolio reports them — all are capture-day snapshots with no 2025 version to pull.
2. Days to complete on maintenance, cleanly. 18.7% of the 2026 cohort has no completion event yet against 7.3% in 2025, so the 2026 median is a floor.
3. Days vacant, cleanly — 34 of 87 move-outs still censored.
4. Renewal lead time on the earlier definition — no 2025 equivalent exists.
5. Three reports have no CSV export in this AppFolio tenant — box score, property performance, tenant transactions summary — missing in both years.
6. The application funnel outside AppFolio. TurboTenant, Tenant Turner and AffordableHousing.com never enter the applications report in either year.
7. Labour hours and billed amounts parse empty on work orders in both years.
8. Phone calls. Neither year has call records; anything that arrived by phone and never became a guest card or work order is invisible in both windows.
9. Pre-2024 tenant history. Both dumps hold only current tenants plus 2025 and 2026 move-outs; 2025 and 2026 are complete, earlier years are survivor-biased and were not used.
10. September 2026. 53 work orders and a partial month landed on 2026-09-01 before capture; every window stops at 2026-08-31.

Scripts: yoy_lib.py, yoy_renewals.py, yoy_renewal_process.py, yoy_leasing.py, yoy_vacancy.py, yoy_maintenance.py, yoy_tenant_comms.py, yoy_seasonality.py. 21 tables at ~/situs-data/analysis/tables/yoy_*.csv; full memo at ~/situs-data/analysis/memo-yoy.md.

Scope. On the apartments: 2,322 work orders, 658 leads, 417 showings, 189 inspections, 61,563 resident communication events, 1,036 resident records, plus the standard reports. The whole database holds 2,588 work orders, 883 leads, 539 showings and 201 inspections; the difference is the houses, the commercial book, the placeholders and the shells. Raw and event-level, never AppFolio's rolled-up summaries, which is why some counts differ from their own reports.

Definitions. Open work order = new, assigned or scheduled as of 1 September. Duplicate candidate = same property and category within a few days, a heuristic and therefore a ceiling. Loop closed = an outbound message to the resident after the completion event. The shared assistant login is a team account used by people and counts as human; PM Assistant, System, Maintenance Performer and Tenant Portal are AppFolio bots; "Customer Service" is the outsourced call centre.

Exclusions. Everything that is not a multifamily apartment property: commercial (5 records, 191 units), single-family houses (20 records, 20 units — the Denver-University-portfolio houses, 446 Galapago, the Ski House), "DO NOT USE" and migration placeholders (3 records, 19 units, 146 work orders, 13 lease expirations), and entity shells, holdings and land (15 records, 19 units). Grand Junction is in a separate database. Three AppFolio reports carry no property column — the 248-application totals, the lead-source yields and the leasing summary — so those stay portfolio-wide and are labelled where they appear.

What this analysis does not claim: that faster answering causes conversion; that any dollar estimate is a forecast; that the phone channel is sized.