Repository avatar
Other Tools
v1.0.0
active

appeal-draft

io.github.2016judea/appeal-draft

Is my Twin Cities property tax assessment too high? County value vs. recorded sales, when to appeal

Documentation

Appeal Draft

Live: https://appeal-draft.vercel.app

Type a Twin Cities address. You get the county's value for the home (blue), what recorded sales of homes like it say it is worth (green), your ZIP's sales per square foot (sand), and one comparable sale. For $90 you get every comparable sale behind the green number and a one-page appeal filled in with your county's form fields, steps and dates, ready to print or save as a PDF.

The next appeal window opens with the March 2027 valuation notices: local boards and Open Book meetings run April 1 to May 31, 2027, county boards in June 2027. A Tax Court petition on the 2026 value is open until April 30, 2027. Sources and read dates: data/deadlines.json.

Covers houses and townhouses in Ramsey, Anoka, Dakota, Scott and Washington counties and the City of Minneapolis (618,892 homes, 48,604 recorded sales since July 2024). Not covered: Hennepin outside Minneapolis, Carver, condos. See DATA_MODEL.md for every field and what reality could not fill.

Ask an AI: the connector

MCP: https://appeal-draft.vercel.app/mcp (Streamable HTTP, no auth). For the homeowner who asks Claude or ChatGPT "is my property tax assessment too high, should I appeal?". Three read-only tools:

  • check_property_assessment: address in; the county's value, what recorded sales of similar homes say, the gap, a verdict, the yearly tax a cut would save, the 2027 dates, and a link that opens this page filled in for that home (/?pk=…&zip=…&sh=…&src=mcp, plus &v= when the owner passes the value from a newer notice). No ZIP? The Census geocoder finds it; out-of-area addresses get a plain "not covered", never an estimate.
  • property_tax_appeal_dates: the 2027 cycle per county, every date with its official source (data/deadlines.json → cycle_2027, read 2026-11-19).
  • assessment_ratio_by_area: the fresh number below, for a ZIP, city or county.

Every number comes from the same shards and files the page reads; a model is never the source of a number. Nothing is sold in the chat: the comp set and the letter stay on the site. Listing copy and what needs a click: SUBMISSION.md. What it records: ANALYTICS.md.

Assessment ratio: the fresh number

Live: https://appeal-draft.vercel.app/assessment-ratio/

County value ÷ sale price for 17,606 recorded sales, each tested against the value it did not go into (sales from the October before to the June after the January 2 valuation date). 31% of homes that sold were valued above their price; the typical home at 0.958. By ZIP the share runs from 14% (55449, Blaine) to 47% (55130, Saint Paul) among ZIPs with 80+ sales, and the cheapest quarter of Ramsey homes is valued above its price 49% of the time against 29% for the dearest. Every county passes the IAAO median and COD tests; Ramsey (PRB −0.051) and Washington (−0.058) sit just past the PRB line. scripts/build_ratio.py → data/ratio.json; the page, llms.txt, llms-full.txt and the sitemap come from scripts/build_connector.py.

Expectation

Written 2026-11-19, before the connector is in any directory.

  • Base rates. heistclips' connector links brought 151 people in 4 days (11/16–11/19), 54% of whom made a clip, because clippers already ask an AI that exact question. The control, the Brick & Mortar business connector on the same channel, got zero human calls in 7 days. Directory review has taken weeks (Claude) and three rounds (ChatGPT).
  • Expected by the read date: between now and March 2027, about zero human calls (the control's rate; nobody has a notice in hand). In the two weeks after notices mail: 10 to 40 people landing from src=mcp links and 0 or 1 paid, if at least one directory has listed it by then.
  • Would surprise us, upward: 100+ src=mcp people by the read date, or any paid letter from a src=mcp visit before March.
  • Would surprise us, downward: zero non-test mcp_tool_call events between 2027-03-15 and the read date while a directory lists it. That says homeowners do not take this question to an assistant, even on the date.
  • Read on 2027-04-01 (two weeks after mid-March notices): mcp_tool_call by client and verdict, people with src=mcp landings and their checkout_open / paid, and chatgpt.com / bing referrers on /assessment-ratio/. Queries: ANALYTICS.md.

How good is the green number

Every home that sold itself since January 2025 (30,224 of them; a home is never its own comp) is a test: the median sale landed at 1.00× our number, and 53% sold within 10% of it. So a gap under 10% is a phone call to the assessor, not a fight. The comp rule ignores lot size and water frontage, so a lake lot or acreage reads too low; the page says which sales it used. Recomputed on every build (site/public/data/meta.json → backtest).

Price

$90, once, per home. A hypothesis. Among the 86,790 homes where the county sits 10% or more above the comps and a tax bill is on the roll, the median yearly saving if the value fell to the comp number is $954 (tax × the cut, assuming tax moves with value; it moves less). A tenth of one year's saving is about $95; $90 is the round number under it. Measured 2026-10-19. Counter-signal: AppealDesk sells a DIY evidence packet for $49 flat (their own comparison page, read 2026-10-19). If round 1 shows checkout opens and no pays, $49 is the next test.

Distribution

Round 1 is drafted, not sent: 50 recipients in outreach/round1/ (gitignored), each with a /for/<token>/ page that opens on a home of their own — 40 listing agents (their live listing), 10 small landlords (a house their LLC owns), and 5 MN appeal services on hold (no home of their own to open on). Subject bricks / <company>.

A/B, the same single variable on all seven shelf products so results pool: A states the price in the email, B leaves it to the page. Alternated within each segment; the variant rides in /data/for/<token>.json and fires with preview_open, never in the URL. Hypothesis: stating the price in the email lowers clicks but raises paid-per-click. Readout on day 3: visits, checkout_open, paid and replies, by variant, by segment and pooled.

Run it

make build    # closed_system.duckdb (read-only) -> site/public/data + private/bundle, the ratio study, the pages
make page     # just the pages (index, /assessment-ratio/, /privacy/, llms*.txt, sitemap, api/_mcp_data.json)
make test     # pytest + node handler tests (paywall and /mcp), logs/*.log
make serve    # localhost:4321 with the API handlers and /mcp
set -a; . ~/Desktop/workspace/bricks/.env; set +a; make deploy   # then pings IndexNow
python3 scripts/build_chatgpt_plugin.py   # chatgpt-plugin.zip for the ChatGPT plugin portal

Only site/public/ is served statically. The paid comp sets ride up as api/_full/, bundled with /api/data and never under a public URL. Webhook goes through ~/Desktop/workspace/stripe-router. scripts/fulfil.py reprints a lost password or mints a comp one.

MIT licence for the code. The data is public record from the county assessors and recorded sales; the owner name is never printed, the homeowner types it on the letter.