Bay Area New-Home True Price

What does a new Bay Area home actually cost once the builder's incentives are counted, and what do the county's records say about the homes already sold in that community?

For production-builder communities in eight counties in and around the Bay Area (Contra Costa, Alameda, Santa Clara, San Mateo, Solano, Sonoma, Napa and San Joaquin; San Francisco and Marin showed none on the builders we read), this project collects move-in-ready homes with street addresses or lots, floor plans and incentive fine print from the builders' own sites. It turns each incentive (flat cash, closing credits, permanent rate buydowns, 2/1 buydowns, ARM promos) into a dollar value, and shows the result as a true-price range next to the county assessor's records for homes sold in earlier phases.

Built for the Meta Global AI Developer Hackathon (Indie track, "Build for Your World").

Live: https://bay-area-true-price.netlify.app: map, community pages and an agent you can ask ("cheapest 4-bed in Contra Costa after incentives, ready now?").

In numbers (data read 2026-11-04; refreshed every morning)

• 581 move-in-ready homes in 146 communities from 17 builders; every home has a street address or lot. Another 83 communities (20 of them coming soon) show floor plans with nothing ready to move into, and are listed but not counted. Sold-out ones are left off.
• 570 homes priced: incentives cut the asking price by 2.7% with "up to" offers at $0, 2.9% with them in full, at a 7.24% market rate. Savings a builder already shows in the price are never counted twice.
• 37 communities show 453 earlier homes' county records (assessed value as published, no estimates); 27 of them have enough for a median.
• Extraction accuracy on 43 hand-checked pages: home recall 99.3%; price, beds and sq ft 100%. Hold-out builder (never tuned on): 151/151 homes.
• Agent, on frozen 11-03 data: Muse Spark 20/20 hand-checked questions, Muse Spark Open (the live model, on Baseten) 20/20 (report). It scored 18/20 until a review caught "biggest discount" ranking by percentage; the 20/20 comes after that fix, so it isn't a blind score. The other miss, a citation to a home id that doesn't exist, didn't recur, which may be chance. Fireworks' 18/20 for the same model predates the fix and wasn't rerun (report).
• Total Meta model spend for the whole build, daily price checks included: about $11.

Meta models used

Spend from the build's own ledger (llm_calls, repeat calls served from cache cost $0):

Job	Model and mode	Calls	Cost
Extract homes, plans and incentives from ~340 builder pages, one extractor for every site, plus daily re-reads of changed pages	Muse Spark 1.3, contributor	1,141	$5.04
Re-extract pages that fail validation	Muse Spark 1.3, standard	43	$1.93
Extraction eval on open weights	Muse Spark Open	43	$1.04
Buyer agent with 7 read-only database tools (native tool calls), eval runs	Muse Spark Open	588	$1.35
Same agent, eval runs	Muse Spark 1.3, contributor	797	$0.16
Read room labels off floor-plan images (vision)	Muse Spark 1.3, contributor	432	$0.10

Contributor mode lets Meta train on the request and is about 12× cheaper. It only ever sees public builder pages and my own test questions.

Muse Spark Open (open weights, on Fireworks) ran the same 43-page extraction eval and also passes the 90% gate (report, $1.04). Against Muse Spark contributor on the same pages (report):

Field	Muse Spark (contributor)	Muse Spark Open
Home recall / price / beds / sq ft	99.3% / 100% / 100% / 100%	99.3% / 100% / 100% / 100%
Baths	99.3%	98.5%
Incentive kinds: recall / precision	95.7% / 100%	96.7% / 97.5%
Plan beds	100%	95.1%
Pages failed	0 of 43	1 of 43 (output ran on to the 32k-token limit)

The buyer agent on Muse Spark Open passes all 20 of the same hand-checked questions, after the ranking fix above, with every dollar figure traced to a tool result (report; python -m tools.eval_agent --purpose agent_live). The site shows the tool calls behind every answer. The eval runs on a frozen copy of the 11-03 data, so the daily refresh can't move its answers; python -m tools.derive_agent_truth --freeze rebuilds the key from newer data with SQL written apart from the agent's tools. Since 2026-10-28 the live site's agent runs on Muse Spark Open, on Baseten since 2026-11-04, which doesn't store model inputs or outputs by default; /api/health names the model. It moved from Fireworks that morning, when a one-word call took 133 seconds there and new questions timed out. Visitor questions are capped at $2/day.

Every model call in the product goes to a Meta model. The code itself was written with Claude Code.

How it works

builder sites ──fetch (robots.txt, 2 s/host)──► preprocess (embedded JSON + visible text)
        ──► Muse Spark extract (JSON schema) ──► validate (ranges, quotes must match the page)
        ──fail──► Muse Spark standard ──fail──► review queue
        ──► SQLite: homes, plans, incentives ──► incentive pricing (FRED rate) ──► true price range
county assessor layers (ArcGIS) ──► link parcels by street ──► each home's assessor record + community median
SQLite ──► export ──► static site (Netlify)       SQLite ──► agent API (Flask on Railway, /api proxy)

Daily price check

Every morning python -m tools.daily_refresh (a launchd job on the build machine) fetches every builder again, re-extracts only pages whose prices, amounts or rates changed (pages change other text daily: office hours, tracking tokens), reprices, records each home's asking and true price in price_history, and redeploys. Model spend is capped at $10 a day on the Meta API; a page that misses extraction keeps its last one, as does a builder whose site fails that morning (dropped connections are retried first), and a day that loads under 70% of the previous day's homes is not published. Each home keeps a permanent id, so links and cited answers stay on the same home. The site shows each home's price history, and the whole series is a CSV: https://bay-area-true-price.netlify.app/data/price_history.csv

Data and rules

• Public data only: builders' own public pages, county assessor open data, FRED.
• Only derived facts and a source link are stored. Raw pages, photos and floor-plan images never leave the build machine, and incentive fine print is linked, not copied.
• Earlier homes show their county records as published: the assessed value (land + improvements) once it has been reset to the last sale. Nothing is estimated. A value under 60% of its community's median per sq ft is withheld as possibly part-built, and the page says so (7 today).
• robots.txt honoured. Builders behind a bot challenge or JavaScript-only data were skipped and logged.

Limits

• Listings and incentives are re-read every morning at 05:30 Pacific; a builder can change either in between. A page that could not be fetched keeps its last reading for up to 3 days, marked as carried.
• County records cover Alameda, Contra Costa and San Joaquin only. Santa Clara has no free assessor values, and San Mateo, Solano, Sonoma and Napa haven't been pulled. Sonoma's builders list no move-in-ready homes today, so it shows communities only.
• Some builders aren't covered: FCB, D.R. Horton and Richmond American need JavaScript or show a bot challenge, and a few small builders (Landsea, Integral and others) weren't added. Most Alameda communities selling now have no closed sales on the roll yet, and Alameda's transfer list ends 3/2025.
• An assessed value on a fully assessed new home is usually the sale price, but it's the county's figure: we have no recorded sale prices today. San Joaquin publishes no sale dates, so its homes count once built before the roll's lien date; a builder still owning a finished home would show its construction value.
• "Up to" incentives may pay $0, so rankings count them at $0.
• A lower rate is valued at what it would cost to buy: one point (1% of the loan) per 0.25% below the no-points rate. That rate is the market rate for a fixed loan and an estimated 0.50% lower for an ARM, because no free daily ARM rate exists; published gaps in 2026 ran 0.31 to 0.83. On a $1,074,533 loan at 4.99%, that range moves the value between $61,033 and $83,384.
• Every figure is at 20% down. A financing offer may not apply at every down payment, so the site doesn't pretend to rework it.
• The 0–3% "worth asking" note on community pages is our estimate of what builders may give beyond what they advertise. It isn't from any builder and isn't counted in any price.
• Home history in price_history.csv before 2026-11-03 used the earlier rate method (payment savings), so compare asking prices across those dates, not true prices.
• A rate offer is only treated as fixed when the builder says it's fixed. Meritage's terms page confirms its 4.99% is a 7/6 ARM, so it's valued that way. Trumark's 5.25% (5.993% APR) names no loan on any page we read, so it's marked "term not stated" and valued like an ARM; that may be too low. KB's 4.99% is an FHA 5/1 ARM and its terms page gives no conventional rate, so it's valued like any loan at 20% down; that leaves out FHA mortgage insurance and may be too high. True prices in the history before 11-03 used the older reading of these offers.
• Not financial advice. The builder's own page is the source of truth; every figure links to it.

Status

Each piece when it passed its gate, with the counts of that day. Current figures are at the top.

Piece	State
Schema (migrations/001_init.sql, tools/db.py)	done
Incentive pricing (tools/price_incentives.py)	done, 17 unit tests
Market rate (tools/fetch_market_rate.py, FRED MORTGAGE30US)	done
Assessor pulls (tools/pull_assessor.py): Alameda, Contra Costa, San Joaquin	done, 155k parcels + 14.5k priced Alameda sales
Builder page fetch (tools/fetch_builder_pages.py): 19 builders	done: 209 community pages + 73 single-home pages; 3 builders skipped (JS-only / bot challenge), logged
Page preprocessing (tools/preprocess_page.py)	done: 0.1–5 MB of HTML → 3k–31k tokens per page
Model client (tools/llm.py): cache, cost ledger, budget guard	done
Truth labels (tools/build_truth.py, no model): 43 pages, 136 homes, 81 plans	done
Extraction (tools/extract_listing.py) + eval (tools/eval_extraction.py)	Muse Spark contributor passes the 90% gate on 43 pages (report)
Validation (tools/validate_records.py) + routing (contributor → standard → review)	done; 6 of 286 pages in the review queue
Coverage (tools/coverage_report.py)	Day 3 gate PASS: 463 unique homes, 100% with address or lot (report)
True prices (tools/load_records.py): 462 homes loaded, 456 priced	Day 4 gate PASS (report)
Floor plans (tools/read_floorplan.py, Muse Spark vision): 107 Lennar plans	98.7% of printed room labels read exactly (report)

How incentives become dollars

Assumptions: 20% down, 30-year amortization, savings discounted monthly at the market rate (FRED MORTGAGE30US; jumbo loans add a spread). Each rate offer gets two figures:

• Payment-equivalent price cut (buyer view): the price cut that gives the same monthly payment at the market rate.
• Builder cost (lender view): the present value of the monthly savings over the period the promo actually lasts. That's 30 years for a permanent buydown, 24 months for a 2/1, and the fixed period for an ARM.

A rate with no loan named isn't assumed to be a 30-year fixed (see Limits). "Up to" amounts are ranges starting at $0. Incentives the builder offers as a choice ("your way") count once, at the biggest option. The true price is a range: asking price minus the high and low values.

Run

uv venv --python 3.12 .venv && uv pip install --python .venv/bin/python -r requirements.txt
.venv/bin/python -m pytest -q
.venv/bin/python -m tools.db                  # create data/trueprice.db
.venv/bin/python -m tools.fetch_market_rate   # latest 30-yr rate
.venv/bin/python -m tools.link_parcels --spot-check   # assessor records per community + live assessor checks
.venv/bin/python -m tools.agent "4 bed under $1.1M true price, ready by January, Contra Costa or San Joaquin"
.venv/bin/python -m tools.eval_agent          # 20-question agent eval -> eval/reports/<date>__agent.md
.venv/bin/python -m tools.geocode_communities # map pins for communities the assessor roll can't place yet
.venv/bin/python -m tools.export_site_data    # web/data/*.json + data/site.db (the live agent's slim DB)
.venv/bin/python -m tools.server              # site + agent API at http://localhost:8787
.venv/bin/python -m tools.build_bundle        # allowlisted deploy folder for the agent API -> .tmp/bundle/
.venv/bin/python -m tools.daily_refresh --no-deploy   # the daily price check, without publishing

Licence

Code: MIT. Listings, incentives and county records belong to their publishers; follow the source links.

What existed before the event

Before the 2026-10-26 kickoff: research only, no code. That covered a county ranking from Census population and permits, a roster of builders currently selling in the four counties, and a check of which county assessors publish free parcel data. All code in this repo was written during the building period.

Build log

• 2026-10-26 (after kickoff):
  • Repo scaffold, SQLite schema (13 tables), incentive pricing engine. Its unit tests were written first and seen failing (17 failed against a stub) before the code went in.
  • FRED market-rate fetcher. First pull: 2026-10-22, 7.32%.
• 2026-10-27:
  • Assessor pulls from public ArcGIS layers: Alameda 80,524 parcels plus 41,603 transfers (14,520 with a single-parcel price), Contra Costa 61,689, San Joaquin 12,951. Every run checks its row count against the server's count.
  • Builder page fetcher (robots.txt read with our own User-Agent, 2 s per host): Taylor Morrison 11 communities, Pulte 12, Trumark 18, with zero errors and zero robots refusals.
  • Muse Spark client with disk cache, cost ledger and a hard budget guard; page preprocessor; model-free truth labels for 43 pages; extractor with JSON-schema output; eval with a 90% gate.
  • First Muse Spark run (contributor mode, 43 pages, $0.07): home recall 99.3%, precision, price, beds and sq ft 100%, incentive kinds 95.7% recall / 100% precision. The first scoring failed at 80.8% kind precision; reading the misses showed the model was right and five truth rules were wrong (fixed test-first, listed in workflows/extract_listing.md).
  • Day 3 (same day): 16 more builders; discovery now reads the city from the page itself when URLs don't name it, follows Lennar's master plans and Meritage's single-home pages. FCB, D.R. Horton and Richmond American skipped (no data without the site's JavaScript; D.R. Horton shows a bot challenge). Routed extraction over 286 pages: 10 escalated to Muse Spark standard, 6 left for review. Hold-out check on Lennar (truth parsed from its own homesite data, prompt never tuned on it): 151/151 homes, price/beds/baths/sq ft 100%. Meta spend to date $2.34.
  • Day 4 (same day): extraction loaded into SQLite (462 homes, 456 priced). Offers don't stack unless the fine print says so; cuts already in the asking price and buyer-group offers are shown, not counted. Average discount 3.6% guaranteed to 6.0% at most, at a 7.32% market rate. Floor plans: Muse Spark vision read 107 Lennar plans, 906/918 printed room labels exact, $0.09. Meta spend to date $2.44.
  • Day 5 (same day): communities linked to assessor parcels by their homes' streets, near the right city, in new-home book-pages, in the community's size band (San Joaquin). 27 communities show past sales (5 Contra Costa / Alameda, 22 San Joaquin). Alameda's 14.5k recorded prices confirmed the Prop 13 rule used everywhere else (value = price × 1.02 per July-1 roll since the sale, 2025-26 roll). Most Alameda communities selling now have no closed sales yet (lots still use code 0800).
  • Day 6 (same day): buyer agent on Muse Spark with native tool calls (search homes, find/get community, plan, past sales, compare, re-price with the buyer's own down payment). Rankings and budgets use the guaranteed end of the true price, since an "up to $250K" offer may pay $0. The re-pricing tool reproduces all 456 stored true prices at 20% down. 20-question eval: 17/20 on the first run, 20/20 after adding address lookup and tightening the prompt; 5 held-out questions also checked correct by hand. Whole eval $0.05 in contributor mode.