# Methods How every number on the page was produced, what it actually counts, and where it lies. `README.md` covers what the app shows and how to run it — this file is the provenance and caveat record. When a definition changes, change it here too. Time base: everything is bucketed to **whole hours** and displayed in **HST (UTC−10, no DST)**. Source APIs return UTC or HST-with-offset; the fetch scripts normalize on the way in. The map's clock spans the union of every layer's window, so a layer that starts later simply has no marker until its first hour. --- ## 1. Rainfall — `map_data.json`, `rainfall_{detail,summary}.csv` **Source** [HCDP Hawaiʻi Mesonet](https://hcdp.ikewai.org), `/mesonet/db/measurements`. Requires `HCDP_API_TOKEN`. **Script** `hcdp_rainfall.py` **Variable** `RF_1_Tot300s` — rainfall accumulated over each 5-minute interval, in **millimetres**. It is a *total per interval*, not an instantaneous rate, so summing intervals is the correct aggregation. ### Derivation 1. Fetch raw 5-minute observations for the window. 2. Drop any row with `value == null` or `flag != 0`. A nonzero flag is HCDP's QC rejection. In the current pull every row is flag 0, so nothing is dropped — the filter exists for future pulls, and `--demo` asserts it works. 3. Bucket into hourly bins by `floor((timestamp − t0) / 1h)`, summing the twelve 5-minute values that land in each bin. `t0` is the first observation truncated to the hour. 4. Join station metadata (`/mesonet/db/stations`) for name, elevation, lat/lng. Stations with no coordinates are dropped — nothing to map. 5. `total` = sum of the whole series. `series[h]` = that hour's rainfall. The browser computes cumulative-through-hour as a prefix sum at load time, so scrubbing is O(1) per frame. ### Definitions | Term in the UI | Means | |---|---| | **storm total** | Sum over the **entire fetch window**, not over the storm | | **hourly rate** | mm that fell during the selected hour | | **through this hour** | Prefix sum from window start to the selected hour | ### Caveats - **"Storm total" is a misnomer.** The current window is **176 h, Aug 9 12 PM → Aug 16 7 PM HST** — about three days of ordinary windward rain before Lala's first band on Aug 14. The sidebar header states the window explicitly for this reason. A true storm-only total would need a `storm_start` field; not built. - **77 gauges sample, they do not cover.** Windward slopes are undersampled relative to their rainfall. The IMERG overlay exists partly to show the ocean and inter-island gaps where there are no gauges at all. - Gauge undercatch in high wind is real and uncorrected. Treat peak-hour values as a floor, not a ceiling. ### Island assignment (derived, not from source) HCDP's `location` field is the string `"hawaii"` for **all 77 stations** — it means the *state*, not the island. Every island label on the page is derived from coordinates by `island(lat, lng)` in `hcdp_rainfall.py`, a first-match-wins list of bounding boxes. Order matters where boxes overlap, so the list is ordered deliberately: - **Molokaʻi** clips Maui's northwest corner near −156.75…−156.68 → tested first. Without this, Keōpukaloa (−156.729) would read as Maui. - **Kahoʻolawe** sits entirely inside the Maui box → tested first. No gauges there currently, but the box is correct if one appears. - **Lānaʻi** is west of Maui's western edge, so no true overlap, but it is ordered before Maui anyway. Result: Hawaiʻi 30, Maui 22, Oʻahu 12, Kauaʻi 6, Molokaʻi 5, Lānaʻi 2. No gauge falls outside every box. **Caveat:** boxes, not coastlines. A gauge on an offshore islet or in a channel could land in the wrong box. Nothing in the current 77 does. The upgrade path is point-in-polygon against Census island shapes, which adds a real dependency — deliberately not taken. --- ## 2. Streamflow — `streams.json` **Source** [USGS NWIS instantaneous values](https://waterservices.usgs.gov), `stateCd=hi`. No key. **Script** `usgs_streams.py` **Parameters** `00060` → `flow_cfs` (discharge, cubic feet per second), `00065` → `stage_ft` (gage height, feet). ### Derivation NWIS returns irregular 5–15 minute values. Each is placed in its hour bin; hours with no reading stay `null` rather than being interpolated. `flow_cfs_peak` and `stage_ft_peak` are the max over the window. 102 sites in the current pull. ### Caveats - **Stage is gage height, not flood stage.** It is measured from an arbitrary local datum, so 22.73 ft at Wailuku is not comparable to 8.87 ft at Manoa-Palolo. Only the *rise* is meaningful within one gauge. Adding NWS AHPS flood categories is the top item in `TODO.md` and would fix this. - **Marker fill is share of that gauge's own peak**, deliberately. Coloring by absolute cfs would just draw a map of which basins are large. - Provisional data. USGS revises instantaneous values for weeks afterward. - Gaps are gaps. A `null` hour means no reading, which during a flood often means the gauge was damaged or submerged — not that flow was zero. --- ## 3. Storm track — `track.json` **Source** [NHC/CPHC ATCF best track](https://ftp.nhc.noaa.gov/atcf/btk/bcp012026.dat), storm ID `cp012026`. No key. **Script** `nhc_track.py` ### Derivation Parse b-deck rows where field 4 is `BEST`. The file repeats each synoptic time once per wind-radius threshold (34/50/64 kt), so rows are deduplicated by timestamp and the radii merged into one `radii` object of four quadrant values each. Latitude/longitude arrive as tenths of a degree with a hemisphere letter. - `wind_kt` is **1-minute sustained** wind, the US convention. `wind_mph` is `wind_kt × 1.15078`, rounded. - `cat` is Saffir-Simpson from `wind_kt`: TD < 34 ≤ TS < 64 ≤ Cat 1 < 83 ≤ Cat 2 < 96 ≤ Cat 3 < 113 ≤ Cat 4 < 137 ≤ Cat 5. - ATCF writes unmeasured integer fields as blank or `0`; both are read as `null`, not as a real zero. This matters for `rmw`, `roci`, and `eye`. - 29 fixes, 6-hourly. The eye marker between fixes is **linear interpolation** — a drawing convenience, not data. ### Peak-intensity tile `track.peak` is the fix with the highest `wind_kt`: **75 mph / 988 mb, Cat 1, Aug 15 8 AM HST, at 17.9 °N 155.0 °W**. The position is rendered by `where(lat, lng)` in `index.html` as **83 mi SSE of Ka Lae (South Point)** — nearest of seven named places, with distance from Leaflet's `map.distance()` and a 16-point compass bearing from a flat-earth `atan2` with a `cos(lat)` correction on longitude. Over ~100 mi the bearing error is well under one compass point. A best-track fix is a point in open ocean; raw coordinates tell a reader nothing, and this is how CPHC writes advisories. ### Caveats - Best track is **post-analysis** and gets revised. It is not what was forecast at the time, and the app makes no forecast/cone claim. - Wind radii are quadrant maxima, not a smooth footprint. The wind-field overlay drawn from them is a schematic. --- ## 4. Power outages — `outages.json`, `heco_snapshot.json` **Source** Hawaiian Electric (HECO / MECO / HELCO). **Script** `heco_outages.py` (live snapshot only) **Units** the file's own `units` field: `"customers without service"`. ### What "customers" counts **Metered accounts — service connections, not people and not addresses.** - A single-family house is **1** customer regardless of how many people live in it. - Apartments are usually one meter per unit, but a master-metered building can be **1** customer for dozens of households. - A business, school, or pump station is **1** customer each. So the figure systematically undercounts people affected. HECO serves roughly 500k accounts for about 1.4M residents (~2.7 people per account), which would put the 250,900 peak somewhere near 650,000 people — an **inference, not a reported number**, and deliberately not shown in the UI. ### Why the series is hand-entered HECO's outage map is a Blazor app over `https://outagemap-api-heco.azurewebsites.net/api/v1`. Auth turned out to be free — `GET /api/v1/access-token` takes no credentials and returns a ~635-char JWT valid 120 minutes; send it as `Authorization: Bearer ` and all three companies answer. `heco_outages.py` does exactly that. **But there is no history parameter.** The endpoint is a live snapshot, overwritten roughly every 15 minutes, so it cannot reproduce the Aug 15–16 peak after the fact. poweroutage.us blocks scripted requests (403) and gates history behind a paid API. So the seven timestamps in `outages.json` are **transcribed by hand from HECO numbers given to the press** (Star-Advertiser, Hawaii News Now, Hawaiʻi Public Radio — URLs recorded in the file's `sources`). The final point comes from the live API. This is the weakest dataset here and the honest label is "sparse". One useful property for future work: every active outage in a snapshot carries `outageStartTime`, so a *single* snapshot reconstructs "customers still out, by the hour their outage began" back to before the storm. ### Kauaʻi — KIUC (`kiuc_snapshot.json`) **Solved, and easier than HECO.** KIUC is a member-owned co-op absent from the HECO API, but its outage map (`https://kiuc.outagemap.coop`, an Angular app on NISC's hosted platform) reads **three plain static JSON files off a public bucket** — no key, no token, no referer check: ``` https://outagemap-data.cloud.coop/kiuc/Hosted_Outage_Map/summary.json /config.json /outage.bounds.json ``` `summary.json` has five top-level keys: `totalServed`, `lastUpdate`, `configurationId`, `regionDataSets` (per-zip and per-county rollups), and `outages` — each with `nbrOut`, `timeOff`, `estimateTime`, `crewAssigned`, `lifeCycleStatus`, `x`, `y`. Timestamps are **epoch milliseconds**. `kiuc_outages.py` fetches and tidies it. **There are more than three files, and probing cannot tell you that.** The bucket denies listing and returns an identical `403 AccessDenied` for keys that do not exist, so "everything else 403s" is not evidence of absence — an earlier draft of this file drew exactly that wrong conclusion. Reading the client bundle instead shows names are built as `{configName}/{name}.json`, and the app requests `summary`, `config`, `outage.bounds`, and `region.{omscounty,omszip}[.bounds]`. The last four return 200 and carry delta-encoded Esri polygon rings in absolute Web Mercator — real zip and county geometry, if a choropleth is ever wanted. ### What "no history" does and does not mean Three separate statements, with different strengths. Keep them apart: 1. **Object versioning is off — proven.** `summary.json` responds with `x-amz-version-id: null`. A versioned object returns a real version id and would let you fetch prior copies via `?versionId=`. It is null, so every write overwrites and there are no earlier versions on the bucket. This is server evidence, not inference. 2. **The client never requests history — proven, but limited.** No history or archive call appears anywhere in the app bundle. That describes what the UI asks for, not what the server holds; a server can expose routes its own UI never calls. 3. **No history endpoint exists anywhere — NOT proven, and not provable by probing this host.** Every miss returns the same generic `403 AccessDenied`, so `/api`, `/history`, `/archive` are indistinguishable from real-but-denied routes. This is the same trap that produced the earlier wrong "only three files" claim; do not read 403 as absence. **Open item:** whether the Internet Archive holds captures of `summary.json` is unresolved — the Wayback CDX API was returning 503 during an Internet Archive service outage when this was checked (2026-08-16). Odds are low for a JSON file no page links to, but "low odds" is not "checked". Re-run: ```sh curl "https://web.archive.org/cdx/search/cdx?url=outagemap-data.cloud.coop/kiuc/Hosted_Outage_Map/summary.json&limit=20" ``` History that may exist off-platform: poweroutage.us (co-ops included, paid API, blocks scripted requests), EIA-417 emergency incident filings (public, coarse), and KIUC board minutes / press releases — the same route the HECO figures came from. Going forward, `power_watch.py`'s archive solves this prospectively; it is not currently pointed at KIUC. **`x`/`y` are recoverable positions**, at 1 m resolution. They are **metre offsets from the lower-left corner of `config.json`'s `boundaryExtent`, in Web Mercator** — not the State Plane Hawaiʻi zone 4 frame that `mapSpatialReference` advertises. That string is a stored-but-unused leftover describing the utility's own data, and it *cannot* describe these numbers: HI-4 has a false easting of +1,640,416 ft, so Kauaʻi would be near +1.6M feet, not −17.8M. (`boundaryExtent == fullExtent` in this config, so either works here; the client reads `boundaryExtent`.) Add the extent origin and invert Web Mercator. Verified four ways: 1. The vendor's own decoder in the app bundle does exactly this (`jsonToSummary` adds `boundaryExtent[0]`/`[1]`; the map component hardcodes `spatialReference = webMercator()`), so this is the documented format, not an inference that happens to fit. 2. All outages land inside Kauaʻi's real bounding box, 21.87–22.22 N, −159.71 to −159.30 W. 3. All of them return `"Kauaʻi"` from the independent `island()` boxes in §1. 4. Points agree with `regionDataSets`' own zip rollups: the 498-customer outage sits in Princeville (96722, 507 out), the 119 in Kekaha/Waimea (96796, 139 out), the 52 in Kōloa/Poʻipū (96756, 97 out). Three for three. The flipped-Y alternative would put the 119 on the Nā Pali coast, which has no road and no customers — so the orientation is settled, not assumed. So the Kauaʻi outages **can** be mapped, with size, start time, and crew status. That layer is not built yet. Same snapshot limitation as HECO — `config.json` sets a 180 s refresh and there is no history parameter — but again every outage carries its start time, so one fetch reconstructs the curve. Snapshot taken **Aug 16, 7:53 PM HST**: | | | |---|---| | Customers served (Kauaʻi) | 36,344 | | Active outages | 100 | | Customers out | 968 | | Largest single outage | 498 customers, began Aug 16 11:01 AM | Customers still out, by the hour their outage began — 949 of the 968 began on Aug 16, with a sharp spike in the 11 AM hour (537): ``` Aug 16 06 AM 119 11 AM 537 04 PM 2 07 AM 14 12 PM 56 05 PM 6 08 AM 23 01 PM 20 06 PM 5 09 AM 32 02 PM 51 07 PM 6 10 AM 69 03 PM 9 ``` **Do not read that curve as an outage level.** It is "customers *still out at 7:53 PM* whose outage began in hour H" — survivorship-filtered, so it is a lower bound at every hour and it systematically over-weights long-lived outages. An earlier draft of this file claimed the 11 AM bin corroborated the HECO statewide peak hour. **That claim is withdrawn.** 498 of the 537 customers in that bin — 93% — are a *single* outage still unrestored 8.5 hours later. The bin is one long-lived event, not evidence of an 11 AM island-wide hit. The median outage in the snapshot is 1 customer and 76 of 100 are single-customer, so the total is dominated by a handful of large survivors. Falsifying this would need repeated snapshots, which is what the archive in `power_watch.py` is for. 19 customers trace to outages that began as far back as Jul 30 and are unrelated to the storm. ### Zip and county polygons (`kiuc_regions.geojson`) **Script** `kiuc_regions.py` → 13 zip polygons + 1 county polygon, joined to the snapshot's `regionDataSets` counts and rendered as the map's optional "Kauaʻi outages by zip" choropleth. Note which files these are. `region.omszip.json` and `region.omscounty.json` carry the named features. The sibling `region.*.bounds.json` files are *not* the same thing — each is a single unnamed feature whose first ring is a world-sized rectangle, an inverse mask the app draws to grey out everything off-territory. No names, no per-region geometry; ignored. **Wire format — delta-encoded Esri rings.** Each ring is one flat array `[x0, y0, dx1, dy1, dx2, dy2, …]`. The first pair is an absolute Web Mercator coordinate in metres; every pair after is a delta **from the running vertex**, not from the origin. Integers, so 1 m resolution. Ring winding carries meaning as in any Esri polygon: clockwise rings are exteriors, counter-clockwise rings are holes in the preceding exterior — `kiuc_regions.py` reproduces that with a shoelace test and emits GeoJSON MultiPolygons with holes attached correctly. One zip (96716) genuinely has two disjoint parts. Validated: all 13 zip polygons fall entirely inside the independent `island()` box for Kauaʻi from §1. Only the county ring pokes outside, at 4 vertices of ~24,600 — correct, since county boundaries extend offshore. The layer is **off by default and not on the clock**. It is one snapshot, and a polygon that sits still while the slider moves would read as "no change here" rather than "no data here". The legend says so inline. Worth reading off it: Princeville (96722) at **12.7% out** and Waimea (96796) at **12.9%** dwarf everywhere else — Līhuʻe (96766) is 0.26%. The damage was concentrated on the north and west shores, not island-wide. ### Three traps in this feed - **Planned outages are mixed in.** `lifeCycleStatus` is `"Planned Outage"` on a couple of records. `kiuc_outages.py` excludes them from `customers_out` and reports the count in `planned_excluded` — currently 2 customers, immaterial during a storm but not on a quiet day. - **Most ETAs are already expired.** `estimateTime` is never null in the live feed, and ~87 of 101 are in the past as of `lastUpdate`. Read naively they look like forecasts. Each record carries an `eta_expired` flag for this. - **`lastUpdate` is the authoritative snapshot time** — use it, not wall-clock fetch time. It is now carried through as `snapshot_time`. ### Two numerators, two denominators The feed does not agree with itself, and neither figure is silently preferred: | | Out | Served | |---|---:|---:| | Point records (planned excluded) | 969 | — | | `regionDataSets` county rollup | 931 | 35,654 | | `regionDataSets` zip rollup | 931 | 35,651 | | `totalServed` | — | 36,344 | The two rollups agree with each other and disagree with the points by 38; the served counts sit ~690 below `totalServed`, and `config.json` sets `useConsumerCountForTotalServed`, which suggests meters-vs-consumers. KIUC publicly reports 33,000+ billed meters and ~38,695 member-owners, so 36,344 is plausible between them. Quote it as **"KIUC's own served count"**, not a verified customer total. ### Why KIUC is not summed into `statewide` `statewide` in `outages.json` is **HECO service territory only** (HELCO + MECO + HECO), matching what HECO reports to the press. KIUC is a separate co-op with its own reporting, so it is carried as its own series and excluded from the sum. The file records this in a `statewide_scope` field, and the chart tooltip says "HECO territory out" rather than "statewide". This rule was applied to fix a real inconsistency: the Aug 16 14:30 row summed the three HECO islands (191,800, correctly excluding Kauaʻi's 1,824) while the 19:31 row had folded Kauaʻi in (161,933 instead of 161,024). All eight rows now satisfy `statewide == hawaii + maui + oahu`, asserted after each edit. It **does not change the peak-outage tile** — still 250,900 at Aug 16 11 AM, a row where Kauaʻi was never reported. The chart tooltip and the peak tile now say "HECO territory" / "HECO only" so the scope is visible rather than implied. ### Caveats - **Kauaʻi is not in `outages.json`.** The KIUC snapshot is a separate file and is not yet merged into the statewide series, so the chart's `statewide` line remains a HECO-only total and understates the real one by roughly 1,000 at the time of the snapshot. - `null` means "not broken out in that advisory", not zero. - Seven points across two days is not a time series. The chart draws it sparse on purpose rather than interpolating a smooth curve that was never observed. --- ## 5. Coastal — `coast.json` **Script** `coast.py` ### Surge — NOAA CO-OPS tide gauges **Observed water level minus predicted tide**, which isolates the part of the water the storm put there. Both series are pulled at datum **MLLW**, metric, GMT, and differenced timestamp-by-timestamp; the hourly value is the **max** difference within that hour. 7 stations: Nāwiliwili, Honolulu, Mokuʻoloe, Kahului, Kawaihae, Hilo, Midway. This is a *residual*, not a formal storm-surge product: it also contains any error in the tide prediction, plus wave setup and seiche. It is the right first-order signal, not an official surge value. ### Waves — NDBC buoys Significant wave height (`WVHT`, column 9 of the `realtime2` fixed-width text), in metres, hourly max. `MM` means missing. 14 buoys probed, 9 reachable; unreachable ones are dropped rather than drawn empty. `realtime2` carries no position, so buoy lat/lng comes from line 2 of `latest_obs/.txt`, parsed from degrees + decimal minutes. That is the buoy's *current* position — a drifting buoy would be plotted slightly off. Peak observed: 5.5 m (18.0 ft) at buoy 51206, southeast of Hawaiʻi Island. --- ## 5b. Gridded rainfall — `rain_grid.json`, `rain_grid/*.png` **Source** HCDP interpolated daily rainfall raster, same token as §1. **Script** `rainfall_raster.py` ``` GET https://api.hcdp.ikewai.org/raster ?datatype=rainfall&production=new&period=day&extent=statewide&date=YYYY-MM-DD ``` 77 gauges *sample* the islands; this **covers** them, at ~250 m. It is the answer to the undersampling caveat in §1 — with the caveat that it is itself an interpolation of those same gauges, not independent measurement. - **`period=day` only.** `month`, `hour`, and `6hour` all 404, so this cannot be put on the hourly clock. The map's "That day's total" mode swaps at HST midnight and holds; the key names the day being shown. - **~1 day lag.** The current day 404s until processed. Aug 16 — the day of the Oʻahu outage peak — was not yet available, which is why the overlay currently shows rain on Hawaiʻi Island and almost nothing on Oʻahu. - **Per-island extents exist** (`bi`, `mn`, `ka`, `oa`); statewide covers all. - **Nodata is `-3.4e38`**, the float32 fill. Unmasked, every statistic returns `-inf`. Roughly 8% of the statewide grid is land; the rest is that fill. **Reprojection is mandatory.** Source is EPSG:4326 at 0.00225°/px; Leaflet's screen space is Web Mercator and `L.imageOverlay` stretches an image linearly between two corners, so a raw 4326 grid lands ~10 km out in the middle of the image — coastlines visibly off the basemap. GDAL/rasterio does the warp (bilinear, nodata-aware). An earlier draft hand-rolled a nearest-neighbour version of this; GDAL was installed the whole time. Output is one RGBA PNG per day plus a summed total, 16–40 KB each — smaller than the JSON that describes them, because most of the frame is transparent. Daily peaks: **Aug 13** 16.3 mm · **Aug 14** 255.4 mm (10.1 in) · **Aug 15** 621.4 mm (24.5 in). Three-day total peak **674 mm (26.5 in)** — higher than any gauge reports for the same window, which is what an interpolated surface between gauges is supposed to show. ## 6. Satellite overlays — `satellite.json` **Source** [NASA GIBS WMTS](https://gibs.earthdata.nasa.gov). No key. **Script** `satellite.py` Three layers: GOES-West ABI GeoColor (hourly), VIIRS SNPP corrected-reflectance true color (daily, ~1:30 PM overpass), and IMERG precipitation rate (hourly, mm/h). **The archive has holes.** Requesting a timestamp GIBS does not have returns 404 and paints an empty overlay, so the script **probes one cheap tile per candidate frame** (z5/y13/x4, covering the main islands) with a HEAD request and writes only the hits. The app then snaps the clock to the nearest existing frame. GOES currently has 137 of 176 hours; IMERG is denser; VIIRS is 8 daily frames. IMERG is a **satellite estimate**, not a measurement — it is included precisely because it covers the ocean between the islands where no gauge exists, but it should never be read as gauge-equivalent. --- ## Derived-in-browser quantities Everything below is computed in `index.html` and exists in no source file. | Quantity | How | |---|---| | Cumulative rainfall | Prefix sum of `series` at load | | Peak 1-hour rainfall | `max(series)` per station — **94.9 mm (3.74 in) at Kaiholena, hour ending Aug 15 8 PM HST** | | Island label | `island(lat, lng)` bounding boxes (§1) | | Peak-fix location | `where(lat, lng)` nearest-place + bearing (§3) | | Stream marker fill | `flow_cfs[h] / flow_cfs_peak` for that gauge | | Inches | `mm / 25.4`, display only — ramps, radii, and all internal math stay in mm | | Eye position between fixes | Linear interpolation | --- ## Known limitations, ranked 1. **Outages are seven hand-transcribed points**, and the Kauaʻi/KIUC snapshot is not yet merged into the statewide series. 2. **Stage has no flood-stage context**, so hydrographs are unreadable as severity. Fixing this (NWS AHPS) is the top `TODO.md` item. 3. **"Storm total" spans the whole 176 h fetch window**, including pre-storm rain. 4. **Island labels are bounding boxes**, not coastlines. 5. **77 gauges undersample** the windward slopes that got the most rain. 6. USGS values are **provisional** and will be revised. 7. Surge is an **observed-minus-predicted residual**, not an official product. --- ## Changelog Newest first. Record what changed about *method or definition*, not cosmetics. - **2026-08-16** — KIUC feed re-audited after adversarial review. Corrected three errors in the entry below: `x`/`y` **are** convertible to lat/lng (Web Mercator offsets, not State Plane), the bucket holds **more than three files** (403 there is indistinguishable from missing, so probing proved nothing), and the claim that KIUC's 11 AM bin corroborated the HECO peak hour was **withdrawn** as survivorship bias. Also picked up `lastUpdate`, `lifeCycleStatus`, expired-ETA flagging, and the feed's two disagreeing numerators. - **2026-08-16** — `statewide` defined as HECO territory only and made consistent across all rows; Kauaʻi carried as its own series rather than summed in. - **2026-08-16** — KIUC (Kauaʻi) outages located and fetched: NISC's hosted map serves `summary.json` publicly with no auth, and its start times reconstruct a lower-bound curve. - **2026-08-16** — Peak-intensity tile gained a location (`where()`); the fix had lat/lng all along and simply was not rendered. - **2026-08-16** — Island names derived from lat/lng bounding boxes. Previously every gauge displayed HCDP's `location` field, which is `"hawaii"` for all 77 and made Maui and Oʻahu gauges read as Hawaiʻi Island. - **2026-08-16** — Sidebar and header now state the rainfall window explicitly, because "storm total" was being read as storm-only. - **2026-08-16** — Inches toggle added as a display-time conversion; all internal values stay in mm.