# Reproducible NY/NJ airport ambient-weather screen ## Provenance and scope Analysis prepared November 1, 2026 from public NOAA observations, not Score field records. NOAA's June 23, 2026 notice says ISD/Global Hourly was superseded and has no updates beyond August 24, 2025; its old HTTPS distribution retired July 31, 2026: https://www.nesdis.noaa.gov/news/service-location-change-integrated-surface-data-global-hourly The notice links the official NOAA Open Data Dissemination archive: https://noaa-global-hourly-pds.s3.amazonaws.com/index.html Use GHCNh for a future current-data study; do not extrapolate this archived series. We chose 2024, the last full calendar year before that cutoff. The selection was made before evaluating screen counts. One year at two airports is not a climate normal, trend, forecast, statewide sample, or a prediction for a particular roof. Inputs, retrieved November 1, 2026: - https://noaa-global-hourly-pds.s3.amazonaws.com/2024/72502014734.csv Newark Liberty International Airport, NJ US; USAF 725020 / WBAN 14734; latitude 40.68275, longitude -74.16927; file elevation 1.9 m. - https://noaa-global-hourly-pds.s3.amazonaws.com/2024/72503014732.csv LaGuardia Airport, NY US; USAF 725030 / WBAN 14732; latitude 40.77945, longitude -73.88027; file elevation 3.0 m. Identity is checked against every row's STATION, NAME, LATITUDE and LONGITUDE, allowing 0.02 degrees coordinate variation, not inferred from the filename alone. These are airport observations, not roof-height site observations. The JSON audit saves exact input SHA-256 hashes, byte counts, row counts, flag frequencies and coverage; compare hashes on rerun to detect upstream revisions. Raw files (~8 MB each) are cached outside version control, not shipped with the app. The small monthly CSV and audit JSON are retained and served with the article. ## Decode and quality policy Reference: https://www.ncei.noaa.gov/pub/data/noaa/isd-format-document.pdf See the control section's UTC/report-type definitions and mandatory fields for wind speed, air temperature and dew point (printed pages 4–12). Only FM-15 routine METAR reports are used. FM-12 synoptic, special reports and other report types are excluded to avoid increasing sampling density in changing weather. TMP and DEW encode a signed value in tenths of Celsius plus a quality flag. WND's fourth component is tenths of m/s and fifth is the speed quality flag; direction flags are not used to judge speed. +9999/9999 means missing. Only flags 1 and 5 are accepted: passed all QC, with 5 denoting an NCEI source. Gross-limit-only flags 0/4/9, suspect 2/6, erroneous 3/7 and all letter flags are excluded, even when NOAA permits some special flags as valid. This is an explicit stricter selection, not a claim that every excluded observation is wrong. Malformed fields are unusable. Broad documented field-domain limits are checked; dew point above air temperature is unusable rather than silently corrected. All three measurements must be usable; otherwise the whole selected bin is unknown. ## Time grid, duplicates and missingness The source year is UTC, January 1 00:00 through December 31 23:59, 2024: 8,784 expected hourly bins (leap year). Both files start January 1 00:00 and end December 31 23:51 UTC, but endpoint coverage does not prove completeness. The JSON separately reports actual routine-report bins, absent bins, unusable bins and bins with multiple routine reports across that whole UTC grid. Floor each timestamp to its containing UTC hour. If more than one routine report exists in a bin, select the latest timestamp BEFORE applying QC. Do not select a favorable earlier value. Equal-time duplicate rows with identical TMP/DEW/WND count once. Conflicting selected-time values make the bin unusable. We do not merge measurements from different reports. Convert bins to America/New_York with Python zoneinfo (UTC-5 standard time, UTC-4 daylight time). Select local Monday–Friday 08:00–16:59, holidays included, for 2,358 expected working-hour bins per station. These daytime shifts avoid the repeated/missing local hours at DST changes. The UTC-year boundaries omit only nonworking local hours, not any eligible 2024 work-shift bin. Absent reports and present-but-unusable reports are separate columns. No interpolation, forward fill, or inference from a special report is used. Both unknown categories fail the screen and break consecutive sequences; they do not imply that adverse weather was observed. Denominators remain the complete expected grid. Working-hour coverage: - Newark: 6 absent + 7 unusable + 2,345 usable = 2,358. - LaGuardia: 10 absent + 3 unusable + 2,345 usable = 2,358. ## Thresholds and sequences These are arbitrary, transparent editorial screening assumptions for comparison, NOT product specifications, safety limits or a validated installation model. All comparisons are inclusive and made in the original SI units: - Baseline air temperature: 10–35 C (50–95 F). - Sensitivity air temperature: 15–35 C (59–95 F), changing only the lower limit. - Air temperature minus station dew point: at least 3 C (5.4 F). - Reported wind speed: at most 4.5 m/s (~10.1 mph). The temperature range compares warmer/cooler ambient scheduling assumptions; the spread introduces a near-saturation screen; the wind threshold tests a low-wind assumption. None is presented as a product-approved threshold. The air-minus-dew-point spread is explicitly NOT the substrate-minus-dew-point margin required by a roofing system. Station air is never a substrate proxy. The original ambient columns have NO precipitation screen. The separately labeled comparison below adds limited precipitation evidence, not a snow-cover, gust, roof dryness, roof moisture, frost, solar radiation, cure or forecast model. A run is a maximal sequence of passing adjacent UTC-hour bins on one local working day. Failed/unknown bins, missing hours, nights, weekends and date boundaries break it. A run of four or more bins contributes its full length once to hours_in_runs_4plus; runs_4plus counts these maximal sequences, not overlapping four-bin windows and not days. Four sampled bins do not prove four hours of continuously favorable conditions. Reports occur at varying minutes; weather between observations is unknown. The CSV has 2 stations x 12 months x 2 scenarios = 48 rows. expected_hours = no_report_hours + unusable_report_hours + usable_hours. passing_hours includes isolated passes; hours_in_runs_4plus is a subset. Scenarios overlap; never add their totals. The article imports the generated JSON for its comparison table. ## Reproduce Python 3.9+ with IANA timezone data; no pip packages, credentials or APIs required. From the repository root: ```sh python3 artifacts/score-spray-foam/scripts/roof-weather.py python3 artifacts/score-spray-foam/scripts/test_roof_weather.py ``` Defaults: raw cache .cache/roof-weather; outputs artifacts/score-spray-foam/public/resources/downloads/roof-weather-2024.{csv,json}. Use --refresh to fetch again. Compare the JSON's source hashes with the committed audit before accepting changed results. Identical raw inputs produce byte-identical CSV/JSON outputs; no changing run timestamp is embedded. The downloadable script is the same source. Outside the repository, explicitly choose local directories (the downloaded method remains a separate companion): ```sh python3 roof-weather-analysis.py --cache ./raw-noaa --output ./derived ``` The script validates CSV headers before caching, so an HTML migration page or wrong station fails loudly instead of becoming an empty or misleading analysis. Tests cover QC, sentinels, threshold boundaries, missing/gapped sequences, overnight boundaries, winter/summer conversion and conflicting duplicates. Precipitation fixtures in scripts/fixtures/precipitation.json cover trace, overlapping accumulations, absent groups/reports, missing depths, present weather and conflicts. Additional tests check condition flags, endpoint overlap, duplicate invariance and a nine-bin sequence split into two four-bin runs. ## Separate precipitation-aware comparison ### Source investigation The same NOAA ISD format document's **Precipitation Data / AA1–AA4** section defines period in hours (00–98, 99 missing), depth in tenths of millimetres (9999 missing), condition, and quality. AA1 is not inherently a one-hour field: the period component must be read. Our routine records include 1-, 3-, 6- and 24-hour accumulations. They overlap and must not be added or divided into invented hourly rainfall totals. Condition 2 means trace: even depth 0000 is not a precipitation-free report. Condition 9 describes a missing condition code, **not a missing numeric depth**. We accept an explicit numeric zero with condition 9 and QC 1/5 as a reported zero, not proof that sub-resolution precipitation was absent. Conditions 1 (impossible/inaccurate), 3/4 (begin/end accumulation), 5/6 (deleted period), 7/8 (missing period), E (estimated), I/J (incomplete) remain unknown here. We do not reconstruct flagged accumulated/missing periods. Period 00 cannot establish a one-hour interval; period 99 cannot locate one. The **Present Weather / AU1–AU9 and AW** sections distinguish precipitation types from obscuration and other phenomena. AU precipitation 01–09 includes drizzle, rain, frozen forms and unknown precipitation; 00 means no precipitation in that report, not throughout a preceding interval. AW 21–25 describes precipitation in the preceding hour, not at observation time. AW 40–48, 50–58, 60–68, 70–78, 80–87, 89, 92/93/95/96 and diamond dust 11 are precipitation evidence. AW 26/90 is ambiguous thunderstorm evidence (with or without rain); it remains unknown. AW 91/94 explicitly lacks precipitation; AW 99 is a tornado, **not** a missing sentinel. Neither implies safe roofing conditions. The document lists AW1–AW4, but these files also contain AW5; we decode the same field format through AW9 and retain frequencies in the audit. AT manual weather groups are empty in the routine rows in both inputs. Remarks, special reports, daily summaries and snow-depth groups are not decoded. ### Rule, timing and overlap This is a **retrospective reported-zero lookback screen**, not a dry-roof rule: 1. Keep the original ambient report selection, thresholds and columns unchanged. Independently select the latest routine report per bin for precipitation. Conflicting precipitation fields at the selected timestamp make that bin unknown even if its ambient readings agree. Never substitute earlier favorable readings. A missing routine report is unknown. 2. Require an explicit valid **one-hour** zero in AA1–AA4 at that selected report. Its interval is `(report timestamp minus one hour, report timestamp]`, not the floored clock-hour bin. Longer zero totals cannot fill absent one-hour reports. Empty optional groups do not mean zero. 3. Index every nonempty AA accumulation in **all routine reports**, including reports outside working hours and later in the year. Treat its duration as ending at its report timestamp. If any valid positive/trace accumulation overlaps the selected one-hour interval by positive duration, exclude that bin as `precipitation_evidence`. A six- or 24-hour total does not identify which sub-hour was wet: this deliberately conservative exclusion means **overlapping evidence**, not rain observed in every excluded bin. Touching endpoints alone do not overlap. Do not sum, distribute or subtract overlapping amounts. Identical duplicates do not change the result. 4. An overlapping known-duration but unusable accumulation, or any supplied unusable selected AA field, leaves the bin unknown unless positive evidence already excludes it. Unknown-duration records cannot be projected onto other bins; at their own selected report they block acceptance. 5. Selected-report AU/AW precipitation also vetoes acceptance, including AU vicinity reports as a conservative choice, without calling it on-roof rain. Supplied malformed, unsupported or non-1/5-QC weather fields are unknown. Empty optional AU/AW fields neither veto an explicit AA zero nor establish one; absence of weather codes alone can never pass the screen. Precedence is selected-report absence/conflict → unknown; otherwise positive evidence → exclusion; otherwise uncertainty → unknown; otherwise explicit one-hour zero → reported_zero; otherwise unknown. No prior/following-year records are fetched, so accumulations reported outside the source year cannot be used. This is not a real-time scheduling algorithm: later reports may exclude earlier bins. Variable report minutes can leave gaps between lookbacks; even consecutive accepted bins do not prove continuous precipitation-free time. There is no assumed 24/48-hour drying delay and no inference about evaporation, ponding, absorbed moisture, frost, substrate condition or coating cure. ### Results and accounting The existing 48 monthly rows gain separate precipitation columns, while `passing_hours`, `runs_4plus` and `hours_in_runs_4plus` remain ambient-only. For every row: - expected_hours = precip_reported_zero_hours + precip_evidence_hours + precip_unknown_hours (independent of ambient eligibility). - passing_hours = precip_screen_passing_hours + ambient_pass_precip_evidence_hours + ambient_pass_precip_unknown_hours. - Combined accepted bins require both screens; unknowns and evidence exclusions break runs under the original same-day rule. Verified against the original input hashes and ambient results: | Airport / minimum | Ambient passes | Evidence exclusions among passes | Unknown among passes | Combined passes | Ambient → combined bins in runs 4+ | Ambient → combined run count | | --- | ---: | ---: | ---: | ---: | ---: | ---: | | Newark / 10 C | 841 | 245 | 21 | 575 | 587 → 418 | 92 → 66 | | Newark / 15 C | 685 | 191 | 15 | 479 | 481 → 336 | 74 → 51 | | LaGuardia / 10 C | 633 | 163 | 6 | 464 | 358 → 279 | 61 → 47 | | LaGuardia / 15 C | 526 | 126 | 2 | 398 | 298 → 235 | 51 → 40 | Across all 2,358 working bins, Newark has 1,425 reported-zero, 865 evidence exclusions and 68 unknowns; LaGuardia has 1,480, 845 and 33. These categories repeat across the two temperature scenarios and must not be added together. Run counts need not decrease in general: an exclusion can split one long run into two shorter qualifying runs. Total accepted bins and bins within runs cannot increase. These results quantify this specific conservative rule, not actual rainfall duration or lost installation days. ## Interpretation Baseline: Newark 841 passing bins, 587 in qualifying runs; LaGuardia 633 and 358. With the 15 C minimum: Newark 685 and 481; LaGuardia 526 and 298. This shows why isolated ambient readings and sustained sampled sequences should not be conflated, and why assumptions and station choice affect a planning screen. It does not identify the best month, establish a causal airport difference, or estimate sprayable days. Rain may occur in ambient passing bins; even the separate reported-zero screen is limited to the evidence and intervals above. Airport exposure and building microclimates differ. Actual work requires site measurements and all manufacturer, project, safety and forecast requirements, including a sufficient coating/cure window.