How the verdict is computed

Locating the parcel

The address is normalized and matched to the appraisal district’s roll by situs address (trigram similarity, house number first). We read the district’s market value and its land and building split, the capped appraised value where a homestead cap applies, living area, year built, quality grade, neighborhood code, land use and the prior-year value.

subject = roll.find(situs ≈ address, same house number first)
require subject.living_area > 0 and subject.value > 0

Selecting comparable homes

Comparables come from the same roll, never from listings. The filter widens in fixed steps until at least 8 homes qualify. Candidates are ranked by similarity (size, age, distance, neighborhood) and the closest 20 are used.

pool  = roll.filter(same county, same assessment year, same land use,
                    same neighborhood code or distance ≤ radius,
                    |sqft − subject.sqft| ≤ tolerance, |year − subject.year| ≤ tolerance)
comps = top 20 of pool by similarity (size, age, distance, neighborhood)

# Each comparable adjusted to the subject with the district's own values (Texas equal-and-uniform grid):
adjusted_i = comp.market + (subject.land − comp.land) + (subject.sqft − comp.sqft) × comp.building / comp.sqft
           = subject.land + comp.building_per_sqft × subject.sqft
median_adjusted = median(adjusted_i)
unequal_pct     = (subject.market − median_adjusted) / median_adjusted

Land is swapped rather than scaled because it follows location, not house size. Comparing raw total $/sq ft instead would count a bigger or better-located lot as over-assessment; on our Harris County test home that error more than doubled the estimate. Where a roll has no land and building split, we fall back to total $/sq ft and say so.

The three tests

Worked example: 1215 Marconi St A, Houston, a single-family home on the 2026 Harris County roll.

sales      = recorded sales of similar homes ≤ 12 months, arms-length screen:
             drop sales under $10,000 and $/sq ft beyond 1.5 × IQR
indicated  = median(sale_price / sale_sqft × subject.sqft)          # needs ≥ 5 sales
implied    = subject.assessed / county_ratio                       # PA: STEB ratio, 100% after reassessment
market_pct = (implied − indicated) / indicated
jump_pct   = (assessed − prior) / prior   vs. the comparables' median change   (supporting only)

ground     = unequal appraisal in Texas (market value shown as support)
           = market value elsewhere; boards decide one ground, so we never average them
confidence = 0.4 × count_score + 0.4 × spread_score + 0.2 × source_score   # 0–1
verdict    = insufficient  if |comps| < 8 (or < 5 sales, or no verified county ratio)
           = fair          if pct < 0.03
           = likely_over   if pct ≥ 0.07 and confidence ≥ 0.60
           = borderline    otherwise   → "we'd wait", no packet sold

Over-assessment and savings

requested_value = median_adjusted (Texas) or indicated × county_ratio (market states), rounded to $100
over_assessment = value the board sets (market value) − requested_value
display cap     = 25% of value unless confidence > 0.85
tax_cut         = the part of the reduction below your capped appraised value (Texas homestead cap)
annual          = tax_cut × county effective tax rate
range           = annual × [1 − spread, 1 + spread],  spread = 0.10 + 0.30 × (1 − confidence)

Sources and limits

1. Appraisal district certified rolls, read on the date shown beside each number (Harris County: download.hcad.org bulk files).
2. Recorded sale prices from deed records where the state discloses them. Texas is a non-disclosure state, so market tests there rely on listing-derived sales, are labelled as such, and only support the unequal-appraisal ground.
3. County assessment ratios from the state where one applies (Pennsylvania: State Tax Equalization Board common level ratio certifications). A county without a verified ratio gets no market-value verdict.
4. County effective tax rates, all taxing units combined, for the latest tax year we hold; the rate used is printed in every estimate.
5. Limits: we do not adjust for condition, because the roll does not record it, or for age beyond keeping comparables within the year range shown. Homestead exemptions and caps change the tax, not the value comparison. Savings shown are capped at 25% of the assessed value unless confidence is above 0.85.
6. Coverage tiers: “Full scan + packet” where we hold the county’s bulk roll; “Scan from licensed data” elsewhere, with lower confidence; “Deadline + forms only” where neither is available.