# Newport Beach property research

Observed October 3, 2026. [Open the local dashboard](http://127.0.0.1:8798/).

The dashboard contains observed homes strictly above $5m and below $12m, in Newport Beach city limits plus a two-mile buffer. It supports drawn polygons and rectangles, saved focus areas, specific source-neighborhood filtering, comparable sales/rents, price movement, rent versus buy and public hazard/court evidence.

## Coverage and provenance

Market analysis is restricted to October 3, 2023–October 3, 2026. The oldest in-scope primary sale is 2023-10-05. Earlier property-history events are excluded from market trends and repeat calculations. Legal research retains relevant current proceedings whose filing dates may precede this window.

| Evidence | Saved observations |
|---|---:|
| Active purchase homes | 89 |
| Primary closed sales within $5m–$12m | 786 |
| Expanded $3m–$15m closed-sale/comp pool | 1813 |
| Valid model rows after exclusions | 1778 |
| Residential asking-rent listings | 247 |
| Selected three-year sale-history observations | 2032 |
| Active homes with public APN/address matches | 69 |
| Court cases directly checked | 2 |
| Dated Airbnb search results retained | 10 |

The [Realtor actor](https://apify.com/epctex/realtor-scraper) and [Airbnb actor](https://apify.com/tri_angle/airbnb-scraper) ran against real sources. Eight research runs succeeded; verified Apify usage total was $7 ($7.21218 before rounding). Raw selected fields, inputs, run/dataset IDs and exclusions are saved in `data/` and `data/run-manifest.json`.

Collection used a broad scraper radius, then clipped records to the Census Newport Beach municipal polygon projected to UTM 11N and buffered by exactly 3,218.688 meters. This includes Corona del Mar and Newport Coast where the municipal polygon includes them. [Municipal GIS](https://nbgis.newportbeachca.gov/arcgis/rest/services/EmployeeMasterMap/MapServer) provides statistical areas, Census blocks, associations, parcel identifiers and public hazard geometries.

Source-selected snapshots do not prove complete MLS/deed coverage. Primary-sale deduplication retains one sold-search observation per property ID; selected histories add some repeat transfers but are not a complete transaction ledger. Asking status, coordinates, parcel associations and current structural attributes still need verification. There are no private owner/contact fields or API keys in the served dashboard.

## Neighborhoods and comps

Matching prefers a specific source subdivision code and preserves distinctions such as Balboa Island Main/Little, Peninsula/Point, and subdivisions within Crystal Cove. Broad statistical areas, mapped associations and source subdivisions are different units. Uncoded labels and association fallbacks remain marked. Source labels can disagree with GIS association polygons; flagged comps require boundary review.

Code filters property type, source frontage flags, distance, living-area ratio, beds and transaction age before Jev reviews. Comps are tiered: same source neighborhood, shared mapped association, nearby different/unverified neighborhood, then expanded geography. Within-tier semantic ranking is advisory. The closed-comp pool spans $3m–$15m; market charts stay in the selected strict purchase price band.

A pooled earlier-sales neighborhood/size model estimated a size exponent of **0.686**; neighborhood-cluster bootstrap range **0.625–0.743**. It used 1240 sales through December 2025, controlling for source neighborhood, statistical area, type, frontage and time. A 20% larger home receives about **13.3%** more value in this fitted relationship. Constant price per square foot assumes an exponent of one.

Comp cards expose `comp price × (subject sqft / comp sqft)^beta`, followed by a separately shown pooled time adjustment. At least three same-source-neighborhood comps are required for their adjusted median. These screens omit unresolved view, condition, lot quality and concessions. The fitted relationship is not causal, and its time slope is not an appreciation forecast.

## Model improvement without choosing on later outcomes

The retained model blends gradient boosting with a regularized model using smoother geographic features. Six fixed candidates were compared using four rolling quarters in 2025 and five geographic folds restricted to pre-2026 transactions. Selection required at least 2% relative improvement in a fixed score (70% chronological mean percentage error, 30% geographic-transfer mean error), with neither validation axis worsening by more than 1%. January–March 2026 supplied interval calibration.

On the same **390 April–October 2026 benchmark sales across $3m–$15m**, median percentage error decreased from **13.23% to 12.38%**. Current mean absolute dollar error is **$1,022,342**. Median-error geographic bootstrap range is **11.35%–13.55%**. A nominal 90% calibrated band covered **89.74%** of those sales.

In your strict **$5m–$12m subset (157 benchmark sales)**, median error changed only from **13.74% to 13.58%**. Current mean dollar error is **$1,291,253**, median signed error **-9.25%** and nominal 90% band coverage **88.54%**. Negative signed error means estimates tended low in this selected subset. No blanket correction was fitted to these later outcomes; price-band selection itself can affect residual statistics.

Paired median-error reduction was 0.85 percentage points, with a geographic-bootstrap range of 0.02–1.96. The mean-error reduction interval includes zero. This is a modest improvement, with local exceptions.

| Local area with worse benchmark median error | Trades | Original | Current |
|---|---:|---:|---:|
| M1 | 16 | 15.05% | 23.83% |
| M2 | 10 | 11.37% | 13.14% |

A second fixed comparison tested neighborhood/size, neighborhood/hedonic and a 20% neighborhood blend. The blend improved rolling error slightly but worsened earlier geographic-transfer error and failed the fixed selection rule. It was not adopted even though its already-seen later benchmark median looked better.

| Neighborhood candidate | Earlier rolling mean error | Earlier geographic mean error | Already-seen later median error |
|---|---:|---:|---:|
| current_blend | 16.17% | 17.02% | 12.38% |
| neighborhood_size | 18.15% | 23.10% | 13.52% |
| neighborhood_hedonic | 16.97% | 25.01% | 11.99% |
| blend_neighborhood_20 | 16.03% | 17.62% | 11.60% |

The 2026 benchmark has been inspected in prior comparisons. It is an audit, not a fresh confirmatory test. Future transactions or an independently sourced locked sample are needed for confirmation. Candidate selection did not use those later prices. Current listing attributes proxy sale-date attributes; historical renovation/view/concession evidence is missing. Final active-home estimates refit all valid observations and borrow benchmark error bands; their active-listing or block-level coverage is unverified.

## Standard deviation, tails and consistency

In the strict $5m–$12m later benchmark (n=157), signed percentage-error sample SD is **19.45 percentage points**, absolute-error SD **13.24 points**, and the **90th/95th-percentile absolute errors are 32.57% / 43.82%**. Mean signed bias is -8.55%. A good median does not remove large misses or bias.

The same nine fixed model candidates were audited for dispersion on four earlier 2025 time folds (633 transactions) and five pre-2026 geographic folds (1,240). Sample SD uses n−1. The table reports signed percentage-error dispersion and variation in mean absolute error across folds. Folds have unequal sample sizes, and four/five folds give limited precision. No model or weighting was changed by this new audit.

| Fixed candidate | Mean absolute error: time / geography | Signed error SD: time / geography (pp) | P90 absolute error: time / geography | Fold mean-error SD: time / geography (pp) |
|---|---:|---:|---:|---:|
| baseline_boost | 16.63% / 17.32% | 23.73 / 22.75 | 35.60% / 36.81% | 1.87 / 1.81 |
| regularized_boost | 16.04% / 17.49% | 23.32 / 23.13 | 36.18% / 37.69% | 1.47 / 1.57 |
| smooth_geo_ridge | 17.74% / 18.59% | 25.27 / 24.76 | 39.25% / 38.78% | 1.59 / 1.94 |
| extra_trees | 16.17% / 18.66% | 22.17 / 23.83 | 36.02% / 37.74% | 0.33 / 1.30 |
| blend_smooth_25 | 16.19% / 16.99% | 23.25 / 22.40 | 33.96% / 35.44% | 1.71 / 2.00 |
| blend_smooth_40 (current) | 16.17% / 17.02% | 23.26 / 22.47 | 34.13% / 35.77% | 1.63 / 2.03 |
| neighborhood_size | 18.15% / 23.10% | 25.40 / 29.15 | 37.68% / 45.33% | 1.35 / 2.34 |
| neighborhood_hedonic | 16.97% / 25.01% | 26.27 / 55.93 | 34.55% / 42.49% | 1.82 / 12.04 |
| blend_neighborhood_20 | 16.03% / 17.62% | 23.45 / 23.38 | 34.31% / 37.73% | 1.61 / 1.23 |

The 25% smooth blend has the lowest larger-of-time/geographic error SD (23.2528 pp), nearly tied with the current 40% blend (23.2574 pp). Both sit on the observed average-error/dispersion tradeoff. There is no established meaningful variance improvement from switching. The neighborhood/attributes candidate has 55.93 pp geographic error SD, reinforcing caution about apparent later median gains.

The dashboard recomputes raw price and $/sqft sample SD, $/sqft coefficient of variation (CV), interquartile range and out-of-time error SD inside a drawn focus area, neighborhood or block. CV is SD divided by the mean, times 100. It allows relative comparison across different price levels. Raw sale spread spans three years; local model residuals start in 2025 and use predictions trained before their transaction window. Local residual summaries require 20 observations; raw spread requires five. These measures describe a selected cross-section and do not establish future price volatility.

Least-variability rankings require a source-coded subdivision, 20 qualifying $/sqft sales and at least five in each of the latest/prior twelve-month periods. Error-SD rankings also require 20 out-of-time residuals. Missing evidence is not zero variance. Minimum counts reduce sparse-sample artifacts but do not prove a stable ordering or verified subdivision boundaries. Housing mix, condition, time and price-band truncation can change the ranks.

Lowest observed $/sqft CV across the entire study area and all home types, before any custom focus filter:

| Source neighborhood | Sales | Latest / prior 12 months | $/sqft SD | $/sqft CV |
|---|---:|---:|---:|---:|
| Harbor View Homes (HVHM) | 41 | 14 / 14 | $274 | 16.53% |
| Balboa Peninsula (Residential) (BALP) | 34 | 14 / 13 | $594 | 20.58% |
| Newport Heights (NEWH) | 25 | 9 / 8 | $285 | 21.20% |
| Irvine Terrace (IRVT) | 26 | 8 / 10 | $533 | 24.20% |
| Balboa Island - Main Island (BALM) | 20 | 9 / 7 | $753 | 27.29% |
| Lido Island (LIDO) | 28 | 8 / 11 | $699 | 33.93% |
| Corona del Mar South of PCH (CDMS) | 40 | 14 / 15 | $870 | 36.86% |
| Balboa Peninsula Point (BLPP) | 25 | 5 / 13 | $985 | 39.88% |

## Jev: real piecewise reviews and boundaries

Pinned **jev-1.13.0** produced **10,338 answers across 1,723 comp pairs**, with 0 unavailable pairs. Six independent questions assess view, condition, lot usability, frontage, layout and outdoor amenities. All saved responses passed contract validation.

Answers were 1,544 similar, 2,730 different and **6,064 unknown**. The ten fixed alternate-order diagnostic pairs included **2 with changed answers**. Those neighborhood tiers retain their deterministic ranking. A small ordering diagnostic is not a property-appraiser accuracy test.

Seven predefined synthetic integration cases included **3 disagreements with the expected answers**: English and Spanish view-match examples returned unknown, and a contradictory-condition example returned a difference rather than unknown. Ambiguous source wording can justify abstention; these expectations require human adjudication and are not a measured production error rate. Choice, Noul and Score were all verified through actual API responses. This small diagnostic does not establish appraisal accuracy or general robustness. The dashboard defaults to deterministic comp order; experimental Jev ranking is an explicit opt-in. Feature probabilities remain visible in either mode.

The reusable `typesafe-jev` skill is installed in the Codex skills folder; a project copy is in `skills/typesafe-jev/`. It covers the request contract, Choice/Noul/Score semantics, independent questions, explicit missing evidence, confidence interpretation, numeric/date/counting weaknesses, context filtering, adversarial source content, option ordering, retries, caching, model pinning and evaluation leakage. Primary references: [API](https://docs.typesafe.ai/api), [state](https://docs.typesafe.ai/concepts/state), [confidence](https://docs.typesafe.ai/confidence), [version weaknesses](https://docs.typesafe.ai/model-jaggedness/jev-1.13), [feature-discovery cookbook](https://docs.typesafe.ai/cookbooks/autoresearch_feature_discovery) and [vendor skill](https://github.com/typesafe-ai/skills/blob/main/skills/typesafe-ai/SKILL.md).

Confidence summarizes the returned distribution. It is not property-specific validation or overall workflow correctness. No exact price/adjustment is reconstructed from Jev scores. Source-stated text is not an inspection finding. Pair-ranking weights are fixed heuristics, not validated price effects. Jev answers remain outside the numerical valuation model because the source descriptions have not been verified as available at the historical sale date.

## Rent versus buy economics

The calculator starts with equal initial wealth and a common monthly cash budget. The renter invests the buyer’s down payment, purchase costs and renovation allowance, then invests or withdraws ownership-minus-rent cash differences. Buyer wealth at exit is sale proceeds after selling costs, remaining debt and the entered terminal-tax scenario. Principal builds equity and is not counted again as an economic expense.

Defaults: $8m purchase, $25,000 monthly rent, 10-year hold, cash purchase, 3% appreciation, 5% net investment return, 3% rent/cost growth, 1.1% initial property tax, 2% assessed-base growth, 0.75% maintenance reserve, $25,000 annual insurance, $800 monthly HOA, 1.5% acquisition costs and 5% selling costs. Tax benefit, sale tax, renovation and uninsured-loss defaults are zero. These are adjustable scenarios, not financing/insurance quotes or forecasts.

| Example ($8m / $25,000 monthly rent / 10 years) | Required annual appreciation | Break-even rent per rented month | Buyer minus renter wealth at 3% appreciation |
|---|---:|---:|---:|
| Cash, 12 rented months | 4.17% | $31,883 | $-1,219,897 |
| 50% mortgage at 6.5%, 12 rented months | 4.86% | $36,273 | $-1,998,084 |
| Cash, 3 rented months per year | 6.85% | $126,616 | $-4,535,171 |

Buying makes financial sense in this model when net buyer wealth exceeds the renter’s investments at the chosen exit. At the default ten-year cash scenario, that occurs at roughly 4.17% annual appreciation, or if comparable rent exceeds about $31,883/month while appreciation remains 3%. Actual financing, tax treatment, repairs, insurance and an equivalent executed lease can change this threshold. No automatic jumbo mortgage-interest deduction or home-sale exclusion is assumed; see [IRS Publication 936](https://www.irs.gov/publications/p936).

Partial-year renting charges only the selected whole number of rented months, placed first in each year. Ownership costs continue all year. The model assumes a rental/stay permitting that term and does not earn rental income on an owned second home.

Across 85 active homes with supported asking-rent candidates, median ask is $7,650,000, median matched monthly rent $17,750, median price/annual rent 33.3×, and median required ten-year cash appreciation 4.90%. Only 8 of those homes favor buying at default assumptions. These are selected listing screens, not independent rent valuations or market-wide investment conclusions.

Rent bases: 42 same specific source neighborhood, 7 mapped-association fallback, 36 adjacent/unverified-neighborhood fallback, and 4 without supported rent. Seasonal/short-term mentions are excluded from automatic matches. Raw candidate asking rents are not adjusted for size or condition; Jev semantic reviews do not change the rent median. Obtain executed luxury lease comps before relying on a home-specific threshold.

Airbnb results used a November 1–December 1, 2026 request for two adults, entire home, at least three bedrooms, capped at ten. Source results did not echo dates, and displayed prices can disagree with calculated breakdown totals. They support furnished-stay context, not verified checkout totals, long-term rent, occupancy, income or permit rights.

## Price movement and local support

The latest complete-quarter pooled adjusted index is **109.5 in 2026Q3**, versus 100 in 2024 Q1. Its geographic-bootstrap range is 104.5–114.9. This adjusts a selected $3m–$15m sales pool for observed structure/location and quarter effects. It remains vulnerable to missing transactions, price truncation and attributes changing since the sale.

Drawing or selecting a neighborhood filters raw medians, counts, selected repeats and economics. Quarterly medians require five trades; direct year-over-year $/sqft changes require ten trades in both twelve-month periods. Sparse blocks/neighborhoods suppress direct changes. The first collection quarter and the three days of October 2026 are partial. The regional adjusted index is clearly labeled and does not recalculate for a custom polygon or block.

Selected history records contain 42 qualifying repeat pairs across 39 properties. Their median gross annualized resale change is 8.65%; renovations, redevelopment, transfer selection and transaction costs are not adjusted. This is not an appreciation forecast.

## Hazards, court evidence and insurance

Official flood, landslide, liquefaction and 2025 fire geometries identify follow-up questions. Hazard screening uses public parcel overlap where APN and site address match; other records use listing coordinates. Shared parcels, condominiums, boundaries and buffer jurisdictions need verification. Some scraped wildfire probability text contradicted its numerical score, so those probability claims were excluded.

Conditional risk associations use regularized city-sale regressions with structure, location and quarter controls, plus 80 geographic bootstrap samples. They are not causal price discounts, loss probabilities, claim histories or insurance quotes. Positive associations can reflect expensive waterfront/location exposure rather than safety.

The Orange County portal’s case registers and hearing calendars were directly checked. Selected observations:

- [30-2023-01365757-CU-EI-NJC: Robert Welbourn v. The City of Newport Beach](https://civilwebshopping.occourts.org/ShowCase.do?showCaseForm.index=0&showCaseForm.tab=0&showCaseForm.number=01365757), filed 2023-11-29 (Eminent domain / inverse condemnation). Galaxy Drive slope failure; regional connection reported by the Dover Shores association. Complaint documents and subject-parcel linkage have not been reviewed.
  - 2026-04-29, ROA 369: Amended judgment for defendant pursuant to granting of summary judgment filed by Repipe-California, LP.
  - 2026-07-22, ROA 484: Request for dismissal without prejudice, party, filed by Dover Shores Community Association. Scope requires the document.
  - 2026-07-23, ROA 486: Request for dismissal without prejudice, party, filed by Repipe-California, LP. Scope requires the document.

- [30-2023-01353566-CU-EI-NJC: Adam Holiday v. The City of Newport Beach](https://civilwebshopping.occourts.org/ShowCase.do?showCaseForm.index=0&showCaseForm.tab=0&showCaseForm.number=01353566), filed 2023-09-20 (Eminent domain / inverse condemnation). Court docket explicitly links this case to 01365757. Subject parcels and allegations require the complaint.
  - 2026-06-12, ROA 432: Notice of entry of judgment filed by the City of Newport Beach. Scope and underlying judgment have not been reviewed.
  - 2026-06-09, ROA 434: Answer to cross-complaint filed by Dover Shores Community Homeowners Association.

A party’s judgment entry or dismissal request does not establish resolution of every claim. Paid complaints/orders and settlement scope have not been reviewed. Neither case has a verified listed-home/APN linkage. No property-level insurance payout record was obtained; absence of an obtained record does not mean no loss or claim.

## Additional sources to add

1. [Licensed CRMLS history](https://go.crmls.org/idx-resources/): closed/off-market/relist events, original prices, concessions and attributes/text as of each sale. Confirm historical analytics rights; IDX alone does not establish them. This is the highest-priority improvement for coverage and valuation backtests.
2. [ATTOM/deed/title feed](https://www.attomdata.com/solutions/delivery/property-data-api/) and [Orange County Recorder](https://www.ocrecorder.com/recorder-services/obtaining-official-record-copies): transaction reconciliation, APN chains, nonmarket transfers, easements, liens and lis pendens.
3. Executed broker/MLS luxury leases, supplemented by [RentCast asking-rent coverage](https://developers.rentcast.io/reference/rental-listings-long-term): term, furnishing, condition and exact neighborhood. Executed leases matter more than another asking-rent scrape.
4. Seller-authorized [CLUE](https://consumer.risk.lexisnexis.com/request), carrier loss runs and [NFIP history](https://nfipservices.floodsmart.gov/frequently-asked-questions-about-nfip-policies-and-claims-data), repair completion and bindable coverage quotes. Public redacted claims cannot establish an exact seller’s payout history.
5. Permit finals, engineering reports, elevation certificates, retaining-wall/slope inspections, coastal restrictions and surveys. Start with [Newport Beach floodplain/building records](https://newportbeachca.gov/government/departments/community-development/building-division/floodplain-management-information), [California geological hazards](https://www.conservation.ca.gov/cgs/geohazards/eq-zapp) and [official fire zones](https://osfm.fire.ca.gov/what-we-do/community-wildfire-preparedness-and-mitigation/fire-hazard-severity-zones); match the exact parcel and jurisdiction.
6. Exact seller/builder/HOA [court searches](https://www.occourts.org/online-services/case-access), complaints, orders and association litigation/reserve/special-assessment/insurance documents. A nearby case alone cannot identify a defect in a listed home.
7. [Redfin aggregate downloads](https://www.redfin.com/news/data-center/downloads/) and [NOAA sea-level data](https://coast.noaa.gov/slrdata/) for broader context; no Redfin historical file or parcel-level NOAA exposure model was ingested in this run. Short-term use requires separate [city permit verification](https://nbgis.newportbeachca.gov/gispub/Dashboards/ShortTermLodgingDashmobile.htm).

## Reproduction and checks

`data/raw/` stores source snapshots; `data/processed/` stores geography, cleaned records, model comparisons, neighborhood study, Jev summary and economics. `data/research/` stores the legal register and actual Jev responses/cache. `exports/` contains current public CSVs. `scripts/` contains collection, screening, modeling, evaluation and report/build steps. Keys stay outside the dashboard and these exports.

Numerical finance and focus-geometry checks passed in Python and JavaScript. Statistical checks verify sample SD, quantiles, missing observations, heavy tails, constant-but-biased residuals, earlier-training provenance and minimum sample support. Contract/evidence checks cover malformed answers, unknown outcomes, pinned model versions, strict price/date scope, comp IDs/tiers and option-order fallbacks. All 1,723 actual pair responses passed validation. Browser checks verify preserved focus areas, specific-neighborhood filtering, selected-home synchronization, adjusted comps, expandable Jev reviews, model tables and finance calculations. These checks verify implementation behavior, not real estate appraisal or claims accuracy.

The local server binds `127.0.0.1:8798` and serves the dashboard, cached photos, report, public exports and fixed property/preference API contracts. Run `node serve.mjs` from this workspace to start it if needed. Rebuilding uses `scripts/build.py`. The numerical comparison scripts need the local `.deps` libraries. No licensed MLS, private claims database, paid court-document packet or executed lease feed was added.

## Live application updates - October 4, 2026

The application is deployed at [newport.aventaria.com](https://newport.aventaria.com/). A property-map click highlights one home with a gold ring and shows its photo and key facts alongside the map. The full property dossier opens through an explicit details button. Compact screens also show a map popup preview. A selected tour property outside the current inventory filters is explicitly labeled.

All 89 observed active homes have descriptions and listing attribution. The application caches all 3,949 source photos, approximately 283 MB of image files, and uses the source-provided higher-resolution images. 87 homes contain a public professional phone and 67 a professional email, which can belong to an advertised office. Contact fields do not verify current representation or authorize messages.

Tours use a [road-travel-time matrix and road route geometry](https://project-osrm.org/docs/v5.24.0/api/) to optimize visit order. Up to seven homes use exhaustive order evaluation; eight to ten use a heuristic. The first chosen home is fixed as the starting visit when no separate starting pin is supplied. Fixed appointments, visit durations, a driving buffer and parking time affect feasibility. Actual returned road-leg times are reapplied to the displayed schedule. The map displays the road line and numbered stops, with property previews, navigation links and a button to display the same route on the property map. Changed tours clear the previous route. No road line is invented when geometry is unavailable.

A labeled three-home example demonstrates route optimization without saving homes as user preferences, exporting calendar appointments, creating inquiries or triggering reminders. In the verified example, the entered order Lido, Newport Coast, One Ford Road was reordered to Lido, One Ford Road, Newport Coast: 10.2 road miles and 26 buffered driving minutes. These are example routing estimates, not confirmed appointments or live traffic.

Personal tours support calendar-file generation with visit/departure entries and alarms, editable agent-inquiry drafts, and optional notifications while the app remains open. Calendar generation and schedule constraints passed numerical checks; no calendar was imported, no reminder delivery was observed, and no showing was requested or confirmed. Automatic sending and booking remain unconnected.

Optional preference memory uses a separate Newport profile service, the Gorse 0.5.11 content engine selected for Entregale, and Hindsight. Canonical explicit preferences control ranking and permissions. The content model returns suggestions within the current map filters; it has no established Newport preference-prediction accuracy and does not change valuation. Hindsight extraction runs through a durable Newport-owned queue using synchronous retain calls, with bounded retries and revision-specific banks. Old or forgotten banks are deleted; asynchronous shared-memory processing is not required.

Synthetic tests on the public site verified independent cookie profiles, consent and origin checks, actual Gorse responses, actual Hindsight extraction and recall, corrected-note recall excluding obsolete notes, stale-revision rejection, and test-profile/bank deletion. Test data were removed. A live property-record query for 3001 Harbor View Drive returned five exact normalized site-address active permit matches. This does not establish permit completion, title, litigation status or insurance history.

The local and public applications expose only intended assets and fixed API contracts. Deployment excludes source snapshots, private credentials, skills and personal profile files.