Nassau County, NY · Property tax grievances · Methodology record

How we pulled every grievance record in Nassau.

From Nov 5 to Nov 11, 2026, a plain Python script queried the county's public Land Records Viewer (LRV) once per parcel. It made 429,722 lookups in three passes and collected 6,460,852 appeal records going back to the 1988 roll. Separately, 156 official assessment-roll PDFs were downloaded and parsed for owner names and tentative and final assessed values. Then everything was joined on a normalized parcel ID. This page covers what was pulled, how, when, and by which agent, along with what's still unverified.

Methodology record · Public county data
LRV lookups429,722407,648 GIS parcels + 15,591 condo units + 6,483 roll-only IDs
Appeal records6,460,852Every roll year, type, ref #, status, updated AV and agent
Roll PDFs1563,201,706,905 bytes, SHA-256 checked, 12 county index pages
Parcels joined429,437426,538 with a 2027/28 tentative AV · 428,920 with an owner of record
Parsed roll rows5,361,2675,088,295 after de-duplication
Pull window6 daysNov 5, 12:24 PM ET to Nov 11, 11:11 AM ET
Public sources
  • Nassau Land Records Viewer, appeals endpoint: https://lrv.nassaucountyny.gov/getappeals.php?id=<lrv_id>
  • Nassau GIS parcel layer Parcels_2026_Clean: gis.nassaucountyny.gov/server/rest/services/Hosted/Parcels_2026_Clean/FeatureServer/0/query. This was the parcel universe, 407,648 features.
  • Nassau County Dept. of Assessment roll PDFs: the tentative and final roll index pages on nassaucountyny.gov (5819, 5831, 5739, 5754, 5677, 5689, 5471, 5517, 5122, 5411, 5121, 5184). Each document's URL and SHA-256 is in roll_sources_sha256.csv.

FOIL was put on hold at Rae's direction and was not used. No paid data was used. Nothing was filed, posted or sent.

Coverage
  • Towns and cities: Hempstead, Oyster Bay, North Hempstead, Long Beach and Glen Cove, which together make up all of Nassau County.
  • Final rolls: 2021/22 through 2026/27, 6 years.
  • Tentative rolls: 2022/23 through 2027/28, 6 years. No 2021/22 tentative roll was posted on the archive pages.
  • Appeal history (LRV): back to the 1988 roll year. Assessed values exist only for 2021/22 through 2027/28.
  • Roll year mapping: LRV roll year Y = tax year (Y−1)/Y. LRV 2028 = the 2027/28 tentative roll, which has no final roll yet.

Methodology, step by step

  1. Confirm the public endpoint

    Xavier X Research tested the LRV appeals page live. It returns an HTML table per parcel with these columns: Roll Year, Appeal Type / Reason, Reference No, Status, Updated Assessment, Agent Name. There is no free countywide bulk download, so the plan was to make one request per parcel.

  2. Build the parcel universe from GIS

    pull_nassau_grievances.py pages the Parcels_2026_Clean feature layer 2,000 records at a time (the layer's maxRecordCount). Fields kept: objectid, parid, parcelkey, address, class, LUC and school district. GIS reported 407,648 features. Condo units are not in this layer.

  3. Map each parcel ID to an LRV ID

    Rule: replace the whitespace in the GIS parid with ++. For example, 01111 01390 becomes 01111++01390, and 15 H 02510 becomes 15++H++02510. Condo units use different encodings, which were spot-checked live on Nov 10. For example, 15-198- -0028-U-CA0111-00070 becomes 15198++0028UCA011100070, and lot-group units drop the ++.

  4. Pilot, then the full run

    The pilot did 30 parcels in about 30 seconds: 28 had appeal history and 23 showed Pending. The full run was launched with --rate 2 (0.5 s delay). The county server limited the actual pace to about 0.9–1 parcel per second. Each response is parsed into rows and written to a SQLite checkpoint, so the run survives restarts. Results export to parcel_status.csv and appeals_history.csv. Every request identifies itself with a research User-Agent. No model reads any parcel; it's a plain script.

  5. Condo-unit supplemental pull

    The gap was found Nov 10. Unit IDs were taken from pdftotext -layout extracts of the 11 2026/27 final rolls. The official condo regex found 11,552 units, and a broader start-of-line pattern added 4,039 letter-block and lot-group units, for 15,591 units total, 0 conversion failures and 0 duplicate LRV IDs. This pull ran at 1 request/sec under nice 19, with its own checkpoint. A 20-minute pace watch was set to pause it if the main pull's speed dropped more than 30%.

  6. Roll-only and utility supplemental pull

    6,483 IDs appeared on the rolls but in neither pull. Most were RS5 special-franchise and RS6 utility IDs (e.g. 1260004005), plus RS1/RS8 parcels missing from GIS. They were pulled at ≤1 request/sec, nice 19, using a restart-on-crash loop.

  7. Retry every error once

    Errors were retried at ≤1 request/sec with the same endpoint and parser. Main pull: 202 errors resolved, ending at has 370,039 and none 37,021. Condo: 4 resolved (15,547 has, 44 none). Supplemental: 6 resolved (2,079 has, 4,404 none). Every network, timeout and parse error was cleared. The 588 GIS features with an empty parid can't be looked up.

  8. Download the assessment rolls

    download_rolls.py discovered the current roll index pages, saved HTML snapshots, and downloaded every listed PDF serially at low priority (nice 15, ionice -c3, about 0.5 MiB/s cap). It validated each file as a real PDF and recorded its SHA-256 in manifest.json. One file timed out (150/156) and was retried. Result: 156 of 156 validated, 0 failed.

  9. Parse the rolls

    Each PDF was converted with pdftotext -layout.

    • Tentative rolls: single-line tables, sliced on each page's header positions. Fields are PARID, OWNER1/2, IN_CARE_OF, owner address 1–3, property address, village, LUC, class, LAND_AV, TOTAL_AV and MARKET_VALUE. 0 parse failures.
    • Final rolls (report AA315NY): multi-line blocks. Fields are SBL, owner, full market value, land and total AV, a transitional "T" amount, exemptions, taxable values, property address and Liber/Page.
    • Reconciliation: parsed counts match each roll's own RECORD COUNT footer wherever it could be read. The 3 known one-record shortfalls and 793 flagged rows are listed in qa_parse_flagged_rows.csv.
  10. Normalize and join (no address matching)

    parid_key = the ID in uppercase alphanumerics only. GIS 01111 01390, tentative 01111 01390, final 01-111- -0139.0 and LRV 01111++01390 all become 0111101390.

    • Final-roll duplicates: final rolls list wholly-exempt parcels twice. Duplicates were collapsed, conflicts were flagged, and the result was 5,088,295 de-duplicated roll rows.
    • Address joins: none were used.
    • Owner names: the roll's owner of record, kept by year.
  11. Validate

    • Match rates: 99.87% of main GIS parcels (406,535 of 407,060) were found in a roll. 100% of condo units were. 99.73% of 2022–2028 appeals (2,121,458 of 2,127,115) matched a roll row in the same year.
    • ARC updated assessment vs final AV: the two are equal for over 99.9% of ARC parcels each year. For 2027, 203,429 are equal and 61 differ.
    • Hand check: condo unit 15-198--0028-U-CA0111-00070 shows a 2026/27 final AV of 645 and an LRV 2027 ARC "Settled" of 645.
    • Rebuild comparison: the final build was compared against the 7:49 AM interim build.
  12. Value definitions

    • Tentative AV: TOTAL_AV on the tentative roll, the published baseline before the Assessment Review Commission (ARC) acts.
    • Final AV: the full ("actual") AV on the final roll. For phased-in parcels, the printed total is the transitional value plus T.
    • ARC updated assessment: the lowest numeric "Updated Assessment" on an ARC row for that year.
    • Limits: a tentative-to-final drop is not assumed to be caused by an appeal or agent, and no tax savings are claimed.

Timeline

All times are Eastern (ET). Pull times come from each run's SQLite fetched_at column and log lines, roll times from project_state.json (converted from UTC), and conversation milestones from the agents' chat history. Where no exact time was logged, the entry says so.

Main pull, parcels by last-fetched date

From full/checkpoint.sqlite. Each parcel keeps only its latest fetch time, so the 202 parcels retried on Nov 11 are counted on Nov 11.

The records

This is a representative random sample of 2,500 parcels, stratified by town in proportion to its parcel count (fixed seed 20261116), taken from the full joined table of 429,437 parcels. The full table is in nassau_records_full.csv.gz (16 MB gzip, 73 MB unzipped). Owner names come from the public assessment roll and are the owner of record, not necessarily the beneficial owner. No phone numbers, emails or mailing addresses are included.

Columns: SBL as printed on the 2027/28 tentative roll · property address · town · class · owner of record · tentative AV 2027/28 · tentative AV 2026/27 · final AV 2026/27 · lowest ARC updated assessment 2026/27 (LRV 2027) · LRV result (has/none) · 2027/28 grievance pending · appeals on file, all years. AV = assessed value. Market value ≈ AV ÷ level of assessment (0.1% class 1, 1% other classes).

Limits & known gaps

  • 588 GIS features have an empty parid and couldn't be looked up. 107 of them have a parcelkey that matches a roll parcel, and those parcels were covered by the roll-only pull.
  • Court outcomes: SCAR and certiorari (court) outcomes that come after the final roll aren't reflected in roll values, and LRV doesn't publish the reduction amount for court settlements.
  • Utility parcels: the 2021/22 final roll is the only final roll that includes utility (RS5/RS6) values. For 2022/23–2026/27, utility parcels have tentative values only. No 2021/22 tentative roll was posted.
  • Parse gaps: 3 records are missing from roll parsing (2022/23 Oyster Bay SD 17–3306 sections 1 and 8, and 2021/22 Hempstead section 6), and 793 rows are flagged. 525 grievance parcels and 25 latest-roll parcels didn't match.
  • Snapshot in time: statuses are as of the fetch dates, Nov 5–11, 2026. Pending matters will change. The original methodology note says "Nov 9–11," but the checkpoints show main-pull fetches from Nov 5 on, so this page uses the checkpoint dates.
  • Exact times: chat history has no timestamps. Rae's original request (Nov 5), the Nov 7 authorization, the Nov 9 sample parse and the Nov 11 roll parse have dates without exact times.
  • File times: file modification times on disk are Nov 15, 12:39–12:46 AM ET for nearly everything, which reflects a later copy, not when work happened. They weren't used.
  • Where it ran: the launch message said the runner would be copied to the P16. Later messages and all output files place the run on the Grok Bot box under /workspace/nassau-grievances/. A copy of the script is also on the HP.
  • Owner names: these are the owner of record on the roll. No LLC members, phones or emails were added.
  • Joins: joins use parcel IDs only. Tentative-to-final differences aren't attributed to any appeal or agent.