- Departures at the chinook sites can be borrowed from the gauges whose years PCIC's model says move most like the stream's (#147, #167). Each site report gains a section 6 that scores
covariance_pickagainst the five nearest gauges at the box's PCIC-covered gauges, under a rule registered before any run; it ships at Maxan, Naver and Gosnell creeks, and Swift Creek keeps the nearest. The comparison article carries each stream's shipped route, says in plain words how donors are chosen, and states Maxan's 10-year 1951–1980 base;wet_departure_carry()'s help page gains the step-by-step. PCIC's per-year files now come from a per-machine cache filled from the public bucket, each file kept only at itspcic.seriessize, withWET_PCIC_YEARSstill naming any other directory (#135).
- The README's Ecosystem table lists crate, the canonical schemas and provenance receipts wet's published build writes through.
- Article maps rebuild byte for byte, and the article bundles read the live build (#165).
scripts/vignette_map_lib.Rtakes lakes and stream segments from fresh's fetchers, orders every query and aggregate, sorts names by radix and simplifies polygons in sf: PostGIS'sST_SimplifyPreserveTopology()under GEOS 3.9.0 drew the same lake differently on repeated calls in one session (research/map_rebuild.md).vm_write()leaves a committed map untouched when a rebuild changes nothing (#155).data-raw/segment_vignette_data.Rruns the segment article's group recipe from the article itself and refuses unless it returns the published picks, and the data scripts install httpfs before loading it (#74). The bundles carry the live build's receipt. Routed PCIC is now picked on 99 % of the Nechako Reservoir groups' segments, so the article no longer names the Nechako. The README's Ecosystem table is rendered from the shared package registry, with links to published sites (#97), andwet_station_snap()'s docs say why it usesfwa_indexpoint()(#55).
- Each published discharge build now records itself, and the fit is published with it (#164). The build writes crate's provenance receipt,
receipt.json, besidemanifest.json. The manifest stays atmanifest_version = 1and gains optionalreceiptandfitentries, whichwet_catalog()returns as attributes and an older wet ignores (#161; crate is in Suggests).fits.rdsand each watershed group's sample are published as parquet underdischarge/fwa/v1/fit/, soWET_GROUP_FROM=published Rscript scripts/segment_group.R <WSG>rebuilds one group on a machine without the province run, given a checkout holding the build's commit (#160).wet_db_load()records each loaded part indischarge.wet_loadinside the load's transaction, keyed by dataset, scenario and part. The table comment is unchanged (#88). The manifest's daily description now names the February fold beside the clamp (#108). The live build's 512 parts are byte-identical to the previous build's.
- One refit chain for six changes that each staled the fit (#163). The CGIAR AET crop is cached under its archives' md5s, checked against figshare's on a rebuild, and the GLO-90 DEM under its tiles' ETags, so a revised upstream source rebuilds them (#23). The climr check against ECCC normals averages only the 30 observed years (#51).
scripts/wb_run_compare.Rcompares two province runs file by file and exits 1 on any difference (#24); the rerun under the rekeyed inputs equalled the previous run in every file, and every tracked fit report reproduced byte for byte (#21).wet_donor_mean()treats a donor within1e-9 * median(d)of a row as the row itself, so a calibration gauge keeps its own shares when its upstream means differ in their last bits, and one group's rebuild now matches every published row (#158). The water-balance gate also scores nested gauges, where the main stems are, under a rule registered before it was scored; the shipped zone adjustment is kept by its 1.0-point margin, and the per-fit override is retired (#47). The published routed PCIC flow moved to fwapg's rebuild that places the Nechako Reservoir: picks moved on 59,206 segments, and routed annual error at 290 gauges fell from 24.2 to 23.1 %.
- One watershed group's balance rows and pick can now be rebuilt alone from the shipped fit, in about 20 seconds instead of a province run (#157).
wet_ws_reach()finds the polygons a group's upstream sums need, cut from the upstream-closed set under the group's roots (wet_ws_closure(),wet_ws_roots());scripts/segment_group.R <WSG>builds over that set and compares every row with the published product. SALR matches it exactly; BULK matches except at 11 calibration-gauge watersheds in months 1–12, by up to 5.8e-5 relative, where the donor shares' exact-zero distance test turns on the last bits of the upstream means (#158). Publishing the fit, so the build runs anywhere, is #160.
- Maxan Creek now runs from its row of the Chinook site table, like Swift, Naver and Gosnell creeks, so the record extension, the borrowed departures and the snow regression have one code path (#146). The two Maxan pilot scripts are retired. Their reports stay as the record, and the site runs must reproduce them line by line before writing anything (
tests/testthat/test-departure_site_parity.R). The short-record article's data now come from the site run, with the same values.
- The two flow-change articles now say what each does (#152), at unchanged addresses. "Measured Stream Flow Against Its 1981–2010 Normal: Two Long-Record Gauges" compares 2016–2025 at Buck Creek and the Bulkley near Houston with the 1981–2010 normal, cited as a WMO standard normal and as the period
wet's published flows are normals of; mean flow was 5–48 % below normal in every window, and every window at both gauges was below normal in each year from 2023. "Rebuilding Stream Flow History Where Gauges Are Few" compares 1951–1980, the decades before warming sped up, with 2016–2025 at Maxan, Swift, Naver and Gosnell creeks, and adds the lowest seven-day flow and snow: peak snow over each catchment held 6–9 % less water and melted 9–16 days earlier. Both articles define the chinook salmon life history timing window from what the fish do. A resampled range is called a 90 % range only where its coverage, measured on simulated no-change series, reaches 0.87 (scripts/departure_interval_coverage.R). All three articles' references are now written bydata-raw/vignette_bib.R.
- The "Chinook Flow at Two Stations" article is now "How Stream Flow Has Departed from Its 1981–2010 Normal: A Case Study" (#145), and its old address redirects to it. It opens with the answer: over 2022–2026, mean flow in the chinook salmon migration and spawning windows at Buck Creek and the Bulkley near Houston was 25–62 % below its 1981–2010 normal. It then shows where the gauges are, each window's flow in each of the last five years, the mean and median day over the whole year and over July to October, and each window's change in mean flow. The trend tests and the year-by-year strip are gone. Its references are written by a new
data-raw/vignette_bib.Rfrom each package's owncitation()and the Zotero group library.
- The "Monthly Flow per Segment" article is merged into "Stream Discharge per Segment", one article for the yearly, monthly and daily values (#149); the old address redirects to it. It opens with a province map of which source gives each river's default flow, scores routed PCIC and the held-out water balance on the same 279 gauges in one table (and the balance on all 315), shows both sources through the year at three Bulkley gauges with the chinook salmon life history timing windows, and says what reading a timing window from monthly means rather than the published days costs. Routed PCIC is described at a high level, with VIC-GL, Raven and the
fwapgfork's notes cited, and every package the article uses is cited inline fromvignettes/segment-discharge.bib. A build check for a claim the article no longer made is gone (#80).
- A new article, "How Stream Flow Has Changed Where Gauges Are Few", runs the Maxan Creek method at three more Chinook streams, fixed in advance: Swift Creek (08KA012) and Naver Creek (08KE014), extended from short records, and Gosnell Creek in the Morice, which has no gauge (#126). The spawning-window drop and the flat migration window hold at all three. Carried spawning-window mean flow fell 12 % at Swift, 41 % at Naver and 35 % at Gosnell, against 46 % at Maxan. The method travels less well: record extension fails in the sparse upper Fraser box, and the borrowed departures score below the 0.5 bar in the spawning window everywhere. The article takes the Maxan article's place in the Articles menu and the README; the Maxan article is still built at its address. The streams are keyed on a site table,
scripts/departure_sites.csv, read byscripts/departure_site.Randscripts/departure_site_snow.R(research/chinook_sites.md).
- The Maxan Creek article's opening now says how Chinook use the upper Bulkley through the year: adults spawn in August and September, the eggs stay in the gravel until the fry emerge in spring, and the young rear in the river and its spawning creeks, Maxan Creek among them (Gottesfeld and Rabnett 2007, pp. 29 and 355). Its map draws and labels Foxy Creek, Maxan's tributary, and labels Maxan Lake, the two ends of the creek's spawning reach. The snow figure is gone, and its paragraph stays. The limits section, now "Where could it be wrong?", says how each filled year is made and that the Bulkley at Quick also takes in the Morice River.
- The Maxan Creek article now opens with two sentences on chinook spawning there: DFO's spawner surveys found them in some years between the late 1980s and the early 2000s, and Gottesfeld and Rabnett (2007) suggest where the main spawning grounds lie and note the creek's low flows. Both sources are cited, and the article's body-prose cap rises from 600 to 800 words. The fuller account waits on a per-stream table of salmon records (#137).
- A new article, "Chinook Flow at a Short-Record Stream", shows how flow in two Chinook windows has changed at Maxan Creek, gauged for only five summers: the rest of its years filled from nearby gauges, and the window flows set beside each year's snow. Its spawning-window flow fell by more than half from 1931–1980 to 2011–2024 (#130). The four articles now cite the Freshwater Atlas, upstream
fwapgand New Graph's fork of it separately, the fork only where it places PCIC's routed flow.
wet_routed_years()routes PCIC's per-year sub-basin days to segments, andwet_departure_carry()can weight donors by PCIC's modelled covariance (#122).scripts/departure_pcic.Rfound donors picked by that covariance carry departures best (median held-out NSE 0.72, against 0.67 for the nearest donors). PCIC's own per-year departures (0.48) and a blend of the two (0.71) do not ship (research/pcic_hydrology.md).
scripts/station_snap_sweep.Rmeasures what NewGraphEnvironment/fwapg#41'sfwa_indexpoint()rework changes onwet_station_snap()'s path (#127). It does not migrate the database: it tries the migration inside a transaction that is rolled back, through the internalwet_pg_trial(). Over HYDAT 2026-07-17's 490 stations only 08KD006's snap moves, to the upper segment at the vertex it is tied on (watershed 8483126 → 7713002, accepted either way).fwa_locatealong()places every candidate measure (114 missed before, 110 of them at stream heads), and two picked catchments that came back empty or raised an error come back. Live tests pin four cases. They skip on an older database, run against the migration underWET_FWAPG_TRIAL, and fail underWET_FWAPG_CURRENT=true. The report isdata/checks/station_snap_sweep_20260717.txt. The shipped stations file and fit stay as they are until the next HYDAT release's refit.
wet_snow_annual()returns each year's ERA5-Land snow from cd's annual layers, such as peak snow water equivalent and the 50 % melt date, over one watershed's dissolved polygons, 1950–2025 (#121). Its example reads Maxan Creek's outline and cd's layers, bundled ininst/extdata/snow_pilot/.scripts/departure_snow.Rregresses the Maxan pilot's Chinook window flows on them: more peak snow and a later melt mean more migration-window flow at every series, and snow explains almost nothing in the spawning window (research/station_flow_departure.md).
- Two functions estimate window flow at a site with a short record or none (#120).
wet_window_extend()extends a short record of window statistics from the nearest long-record donors that cover each year (OLS, MOVE.1 or a ratio of means), with an interval per year.wet_departure_carry()carries donors' log departures, each from the donor's own 1981–2010 mean, onto a site.scripts/departure_pilot.Rchose OLS under a rule registered before any candidate existed, and found carried departures usable (median held-out NSE 0.63 on the Chinook migration-window mean, against 0.5 set in advance). At Maxan Creek (08EE018), extended to 1931–2024 from five gauged years, the spawning-window mean was 56 % lower in 2011–2025 than before 1981 (90 % interval −73 % to −28 %). The report isdata/checks/departure_pilot_20260717.txt.
- After an upload,
scripts/discharge_publish.R(WET_PUBLISH=1) rewrites the publish report's live-build line itself once its anonymous verify passes, so no second BUILD follows a republish (#117). The report now records the md5 of its own tree's parts and of the live parts (wet:::wet_catalog_digest()). PUBLISH uploads a tree only under a committed report naming that tree's md5, and the vignette data scripts refuse a report that names the live commit and part count but describes other parts (#94).
- The published discharge is rebuilt on the first province run under #76's order-free sums (
92b893c4da), with the fit refit on it: every decision is unchanged and every value moved by rounding only (at most 7.2e-11 relative). The articles' data bundles are rebuilt from that build; no number in them changes.
- Upstream sums no longer depend on the order rows arrive in (#76).
wet_upstream_sums()adds polygons that share(wscode, localcode), and each watershed's irregular pairs, inwatershed_feature_idorder; shuffled Fraser and Columbia inputs moved the old sums by up to 4.4e-16 relative and leave the new onesidentical(). The fetches on the province path (wet_ws_fetch(),wet_upstream_irregular(),wet_ws_geom(),wet_upstream_pairs()) order on a unique key, so a run's stored files come out in one row order, andwet_station_snap()breaks an exact score-and-distance tie onlinear_feature_id.scripts/wb_upstream_order.Rreplays two province runs' stored row orders through the current sums. The change alters the province run key, so it reaches the published balance with the next province run.
- The three articles name, link and cite the software they run on (#110). Each ends in a "Software and citation" section computed at render from every package's own citation, through the internal
wet_software_cite(): the version, the first DOI its citation carries (wet's concept DOI) and the commit when the install recorded one. Their closing provenance now says which numbers were checked against what: routed PCIC against fwapg's tables, the balance remade from its fit where the bundle holds it, the scores held out against HYDAT, and the rest on the build report. The monthly-flow article's Data line gains the balance's inputs. The segment-discharge article drops its Validation section and its pointer to an open fwapg issue, states the zone adjustment as kept by decision, and puts the score table's note in a footnote. The station-flow article defines its Theil–Sen and Mann–Kendall trend tests and cites them, withKendallandzyp. The word caps of those two articles count body prose up to the citation section.
- The monthly-flow article gives each Chinook window's flow from the published days as well as from monthly means (#107), with the gauge's own monthly means against its daily record as the cost of reading a window from months; its data bundle carries the three Bulkley segments' days, checked as the publish checks them (routed PCIC routed again from fwapg's sub-basin days, the balance made again from its months and the fit's daily shape). The article is rewritten for its reader: a locator map of the three gauges and their nested catchments, the months and their scoring before the windows,
fwapg(Norris 2026) and the windows' source (Gottesfeld and Rabnett 2007) cited, the Nash–Sutcliffe score defined and rated (Moriasi and others 2007) under the group table throughkableExtra(now in Suggests), and the build checks out of the article, which run in the data script.scripts/vignette_map_lib.Rmakes the map layers for both station articles; the gq registry gains thebulkley_maplayer.
- Per-segment daily flow is published as
discharge/fwa/v1/segment_daily(#99): every FWA segment's mean flow on each day of a 365-day year (29 February folded into 28), for both sources, one parquet part per source and watershed group (2,886,046,240 rows in build 77585a8). The balance's days are its published months laid on the fit's daily shape (#70), so each month's days average to it. Routed PCIC's days are routed from fwapg's sub-basin climatology and segment weights, clamped at zero and rounded as fwapg makes its months; the months rebuilt from them equal fwapg's on every segment, and their plain mean need not equal the published month (the clamp, and February's fold).wet_segment_daily()reads them for named groups,wet_db_load(doys =, groups =)loads a subset, andwet_values_compare(time = "doy")compares them. The segment-discharge article shows the read. - The package is cited by its Zenodo concept DOI (
Config/DOI, read bycitation("wet")).
- The water balance's monthly shares come from donor gauges (#102): each watershed takes the inverse-distance mean of the observed monthly shares of the five calibration gauges most like it in elevation, location, snowmelt timing and peak snow water equivalent (
wet_share_fit(method = "donor");"blend"and the twelve regressions remain as methods). A rule registered inscripts/wb_validate.Rbefore any candidate was coded chose it: held out, the median per-gauge share NSE rises from 0.874 to 0.939 (0.935 with each gauge's upstream and downstream neighbours removed), and on 110 short-record gauges no fit uses from 0.772 to 0.837. The freshet, the Liard's winter and non-snowmelt regimes gain most; four zones (17, 06, 16 and 26) score lower under the donor. The annual flow and every segment's default source are unchanged. The published product ats3://fresh-bc/discharge/fwa/v1/was rebuilt and republished with the new months (build 8ca2dee), and on a calibration gauge's segment the balance now carries that gauge's own monthly pattern. The monthly-flow article is rewritten on it, including the balance's late-summer overshoot at 08EE003.wet_shape_predict()'s donor shares its weighting with the shares (wet_donor_mean()); its output is unchanged.
- Routed PCIC is read from fwapg's PNWNAmet historical 1981–2010 run instead of 1951–2012 (#103), so both sources on every segment, and the gauge scores, describe the same period. The published product at
s3://fresh-bc/discharge/fwa/v1/was rebuilt and republished from it: 4,733 segments changed their default source (3,013 to routed, 1,720 to the balance), and annual skill at the gauges barely moved (routed mean absolute error 24.2 %, 24.1 % before). Routed's spring-low, summer-high shape survives on matching periods (March–April −11.5 %, July–August +15.8 %), so the period explains part of it and PCIC's model the rest (#67). The monthly-flow article's period section is gone, and both articles, their data and the README map are rebuilt on the new build.
wet_segment_monthly(from = "fwapg")and the routed scripts read PCIC's routed flow from fwapg'spcic.discharge_monthly, for the PNWNAmet historical 1951-2012 run, instead of the frozenwhse_basemapping.fwa_stream_networks_discharge_monthly(#101). The run equals the frozen table row for row. The table and run are named once, in the internalwet_segment_sources(). A table without the run's rows is an error that names fwapg's load files, and the routed fingerprint must equal the run's row inpcic.runs. A segment's routed month 0 is now the mean of its months summed in month order, which moves its last bits in 11.4 % of routed segments; the published product is unchanged until #103 republishes it.
- The water balance has a daily shape within the month (#70).
wet_station_doy()gives each calibration gauge's day-of-year record on a 365-day common year (29 February folded into 28);wet_shape_fit()andwet_shape_predict()(donor,basis),wet_shape_flat()andwet_shape_spline()give within-month weights that sum to 1 per month;wet_shape_daily()lays monthly means on them, so every calendar month's mean stays the published one andsegment_monthlyis untouched.scripts/wb_shape_daily.Rscores the candidates held out on the shipped fit under a rule registered before any candidate existed, and donor gauges (k = 5) ship tofit_20260717/shape_daily.rds: median per-gauge NSE 0.982 against 0.899 for flat months (data/checks/wb_shape_daily_20260717.txt). Nothing is published yet;segment_dailyfor both sources is #99, blocked on NewGraphEnvironment/fwapg#25.
- The README is rewritten in plain language (#75). It says what wet gives: modelled monthly flow on almost every BC stream, published openly, and flow and water temperature at stations in the seasons that matter to fish. It shows a province map of where each river's default flow comes from, drawn by
data-raw/readme_figure.R, and a read of the published flow that needs no credentials. It lists all three articles. link can now read wet's published discharge (NewGraphEnvironment/link#331). The function reference is grouped by purpose, and DESCRIPTION names the two published sources.
-
data-raw/segment_vignette_data.Raccepts the segment report that the tracked publish build read, so a rerun that leaves every published part identical needs only a passing BUILD, not a republish (#93). It requires that report to be identical to its copy at the build's commit, and the segment article's data now record its commit (segment_report_commit). The values are unchanged. -
citation("wet")has the keywetand a noteknitr::write_bib()keeps, so a report can cite@wetand get the publisher and the installed version.
- The monthly-flow article's data are rebuilt from the shipped fit after its rescore under the AET rule that keeps cfu on a near-tie (#91). The values are unchanged.
- The citation names New Graph Environment Ltd. as publisher and funder:
citation("wet")now reads frominst/CITATION, andAuthors@Rgives the author's affiliation. The small logo keeps its name clear of the icon.
- Scripts, data scripts and live tests now read a named HYDAT release, from one per-user cache (
tools::R_user_dir("wet", "cache")/hydat/<release>/), instead of whichever release a machine last downloaded (#46). The internalwet_hydat()downloads ECCC's current release into the cache, and checks every file is whole and holds the release it is named for.WET_HYDAT_RELEASEnames the release, by default the shipped one;WET_HYDATis retired and refused. Station-flow outputs record their release.wet_station_*()still default to tidyhydat's own directory for interactive use.
- The published discharge at
s3://fresh-bc/discharge/fwa/v1/now carries the snow-informed monthly balance (build acb820f, #77). Annual flow and each segment'spickare unchanged. The monthly-flow article's data were rebuilt from it, and the article now describes the snow terms. The balance reads high only in April and May, by a median of 7 % to 11 %, and is within 5 % in the other months. The freshet is still flattened.data-raw/monthly_vignette_data.Rreads the held-out months each fit stores, instead of recomputing them (#68).
- The water balance's monthly shares now also use snow: ERA5-Land's 1981-2010 day by which half the snowpack has melted and its peak snow water equivalent (capped at 3,000 mm for glacier points), from cd through the new
wet_snow_normals(), averaged upstream like the other predictors. A rule fixed before scoring kept them: the held-out median share NSE rose from 0.856 to 0.874 on the shipped fit. Winter and early-spring flow is much less overpredicted, the spring peak is a little less flattened, and zone 27 is worse. The annual balance is unchanged (#16). Each fit now stores its held-out monthly balance (#68). The published discharge still carries the earlier monthly values until #77.
wet_segment_monthly()reads flow per FWA segment and source in the shape ofwet_station_monthly()(linear_feature_id, month, q_m3s, source), so a gauge comparison is a join onmonth(#59). It reads the published product, from Postgres afterwet_db_load()or from the parquet parts through duckdb, or withfrom = "fwapg"fwapg's routed table, whichscripts/pcic_routed_compare.Rscores before a publish. A segment with no value comes back as 13NArows.wet_flow_window()turns monthly means into mean flow over month-day windows, labelledbasis = "monthly"; the monthly article's read and its Chinook windows now go through both, with its numbers unchanged.
- The segment-discharge article is simpler (#71). It now says what the published product is, which value to use, how to read it and how good it is, in that order.
pickhas its own early section with its rule and the province counts: routed PCIC on 3,152,838 segments, the balance on 1,380,005, neither on 374,598. One province map of the sourcepicktakes on main rivers replaces the Salmon River maps and the scatter. The map's picks are read from the live published parts bydata-raw/segment_vignette_map.R, under the guards the values script uses. The rename ofpicktois_defaultat v2 is recorded inresearch/discharge_publish.md.
- A new article covers months 1–12 of the published product (#65): how to read one stream's twelve months, both sources' hydrographs at three Bulkley gauges with the Chinook windows, the flow in each window beside the gauge's daily record, and month-by-month scores at the calibration gauges. Held out at all 315, the water balance reads 8–16 % high from January to May and within ±5 % from June to December, and 67–96 % low in January–March at the 25 northern gauges routed flow does not reach (#16). Routed PCIC, at 279 gauges, reads 20 % low in March–April and 19 % high in July–August, and its 1951–2012 period explains only part of that (#67). The values are read from the published parts and checked against their sources before they are bundled (
data-raw/monthly_vignette_data.R), under guards now shared with the segment article (scripts/vignette_data_lib.R).
- The segment-discharge article now introduces the published product (#64): where it is, what a row holds, how to read it with
wet_catalog(),wet_db_load()or duckdb, whatpickselects, how each source was made, and how both score at the 315 calibration gauges, with the Salmon River as the one worked group. The values it shows are read from the published parts and checked against their sources before they are bundled.scripts/discharge_publish.Rnow reads each part it writes back and compares the flows, and the segment rows' runoff, coverage andpick, value by value with what they were built from: on build 7c62b9e, 555,857,796 values with none differing, and all 67 parts identical to the live upload. The fwapg parity and sampling results moved toresearch/fwapg_mad_method.md.
- Per-segment and per-watershed discharge have a published layout at
s3://fresh-bc/discharge/fwa/v1/(#7): Hive-partitioned parquet keyedscenario/forcing/source, one part per FWA top-level drainage with one row group per watershed group, and amanifest.jsondescribing every part, because the bucket allows reads but no listing.wet_catalog()reads the manifest.wet_db_load()loads one scenario of a dataset into Postgres (discharge.fwa_v1_segment_monthlyordischarge.fwa_v1_watershed_monthly), checking each part's size and md5 first, so fresh can joinmonth = 0 AND pick.scripts/discharge_publish.Rconforms each part to its crate schema (crate 0.7.0), builds the tree and writes a tracked report. Uploading is a separate, explicit step and has not been run yet. One dataset ID,<variable>/<network>/<version>/<unit>_<timebase>, names the S3 prefix, the crate schema and the table (research/discharge_publish.md).
- FWA segments now carry both estimates where each exists, PCIC's routed flow (NewGraphEnvironment/fwapg#6) and the open water balance, labelled by
sourceandperiod, with a logicalpickmarking the default (#57).scripts/segment_output.Rwrites them per basin aslinear_feature_id, month, runoff_mm, q_m3s, coverage, source, period, pick. The default is routed unless its mean annual flow is under 0.2 of the balance's, and only a segment on its watershed's own line can fall to the balance, because the balance reaches a segment throughfwa_streams_watersheds_lutand off that line it is another line's flow. 224,199 of 3,374,133 segments with both sources fall to the balance, 26.7 % of them in the Nechako Reservoir groups, where PCIC's outlets are almost all unplaced (NewGraphEnvironment/fwapg#24). At the 290 calibration gauges routed flow covers, the pick's mean absolute error is 22.9 % against routed's 24.1 % and the balance's 27.6 %. The 369,217 segments with no lut row get neither. The balance stays the independent estimate both are scored against.
- The segment vignette compares the open water balance with PCIC's routed flow (NewGraphEnvironment/fwapg#6) instead of fwapg's older annual table (#58): the table of strengths and limits, the scatter, the province coverage outline, the Salmon River pair and the Bulkley gauges. At the 290 of 315 calibration gauges routed flow covers, its mean absolute error is 24.1 % against the held-out balance's 27.6 % (median 13.0 against 18.0); by mean error the balance is closer in the Peace and the Fraser. Where the older annual table also had values, the two PCIC products tie once 7 far-low gauges are set aside, so what routed flow adds is coverage: the coast and the largest rivers. The scores come from
scripts/pcic_routed_compare.R, whose report records the fwapg commit and the table's fingerprint, and the vignette's data script refuses any other build.
wet_temp_fill()fills gaps in daily water temperature at stations, loggers included (#40). Air2stream is run open-loop on daily air temperature fromcd::cd_extract_daily(), and a Kalman smoother carries each station's departure from it across a gap, from both ends, adding the error its peers in the samepoolsaw that day. Held out at 999 station-years, it beats air2stream alone on every gap: 0.48 against 1.31 °C on 7 days, 0.76 against 1.43 °C on 30 days, 0.84 against 1.14 °C on a seasonal logger's spring and autumn. Over a whole missing season its GSDD is no better than air2stream alone (130 against 129 °C-days), the one criterion of a rule fixed beforehand that it fails.wet_temp_gsdd()gives growing season degree days per station-year inwet_window_stats()'s long format, with the share of each season that was filled. At the 17 Skeena stations, every site's GSDD lies inside the network model's 95 % interval (hill_etal2025Spatialstream).
- The dry-interior gauges are scored against licensed water use above them, under a rule fixed before any licence was matched to a gauge (#53). BC water rights licences, placed upstream of each calibration gauge on the FWA network, account for 0–2 mm/yr above the zone-24 gauges against overshoots of 26–245 mm; returning all of it moves zone 24's held-out error by half a point. Diversions above the gauges do not explain the overshoot. The shipped estimate does not change.
- Climr precipitation is scored against the River Forecast Centre's plateau-elevation precipitation gauges and snow courses, under a rule fixed before any product value was computed (#50). climr is at most modestly high on the dry plateaus relative to the rest of the interior (dry over contrast 1.09, not decided), and part of #45's gap between climr and PNWNAmet is PNWNAmet's. Precipitation alone does not explain zone 24's overshoot; the next check is gauges with water taken out upstream (#53). The shipped estimate does not change.
-
The open water balance's overshoot in the dry interior (hydrologic zones 15, 17, 23 and 24) is diagnosed against PCIC VIC-GL's split of precipitation and evapotranspiration, under a rule fixed before any PCIC value was computed (#45). Neither term is named: our precipitation runs 1.4–1.6 times PCIC's there, and our AET offsets most of the difference. The dry zones differ from the rest in precipitation, not AET, so the next check is precipitation observed at plateau elevation. The shipped estimate does not change.
-
The README and package description now cover the water-temperature path, and no longer claim seasonal or climate-scenario output, which no script builds.
-
The open water balance is refit on HYDAT 2026-07-17 (#43). The pipeline now keeps one fit per HYDAT release (
WET_HYDATnames the file), and the 2025-10-14 fit is reproduced exactly beside it. The new fit has 315 calibration gauges and a held-out mean absolute error of 27.5 % (headwater 30.7 %, nested 17.6 %). Pooled zones get no adjustment, settled on short-record gauges no fit uses, which retires the "mildly optimistic" caveat. The zone adjustment ties raw P − AET on headwater gauges and is kept by a recorded decision because it corrects the major rivers (#47 revisits the gate). -
The segment discharge article compares the two estimates in a table of each one's strengths and limits, with guidance on which suits which job, in place of a single recommendation. PCIC's model fits gauges more closely where fwapg has a value; the open water balance covers all of BC and ships its code.
- The segment discharge article is reordered for readers: what mean annual discharge is, which estimate to use and why, how close the open water balance comes at 290 gauges, the Salmon River and Bulkley River as maps that hold every gauge they name, and limits in order of what affects a user. fwapg (PCIC) is now scored at the 168 gauges it covers: it is closer there (25.5 % against 30.5 % mean absolute error), and the article recommends the balance everywhere on the strength of its coverage and open code, stating that gap (#39).
wet_temp_daily()starts awet_temp_*family: daily mean, minimum and maximum water temperature at about 300 hydrometric stations from the water-temp-bc archive, in each station's local standard time, in the shapewet_window_stats()takes. Departures are against 2016-2025, since few stations have year-round records before the 2010s (#36).
- Hex sticker, and a README that follows fresh, link and drift: install, prerequisites, a station example, vignettes and the package ecosystem.
-
Package scaffold, PCIC VIC-GL OPeNDAP subsetting, annual/monthly runoff aggregation, upstream area weighting and an fwapg MAD parity check (#1).
-
Join-free upstream accumulation over FWA codes (
wet_upstream_sums(),wet_upstream_irregular(),wet_ws_fetch()): whole basins, no order-8 skip.wet_upstream_mean()now takes(ws, values, denom, irregular_pairs, upstream_area)instead of a pairs table.wet_pcic_annual()fetches PCIC one year at a time;scripts/mad_basin.Rbuilds the Fraser (#2). -
An open, province-wide water balance after Chapman et al. (2018), the BC Water Tools method, validated out of sample at 290 HYDAT stations (#11).
- Inputs: CGIAR AET (
wet_cgiar_aet()), a GLO-90 DEM (wet_dem_glo90()) and climr 1981-2010 normals (wet_climr_normals()). - Stations:
wet_station_select(),wet_station_monthly(),wet_station_snap(), shared with #6. - Multi-layer sampling and accumulation:
wet_ws_sample()now takes several layers;wet_upstream_means(). - The Chapman fit and monthly shares:
wet_wb_fit(),wet_wb_adjust(),wet_wb_pooled(),wet_share_fit(),wet_share_predict(). - Blocked cross-validation:
wet_cv_folds(),wet_flow_validate(). - Out of sample, the zone adjustment improves annual runoff only slightly
(blocked-CV mean absolute error 33.1 % against 34.7 % raw), and passes its
gate only under a variant added after the pre-set specification failed.
The
scripts/wb_output.Rprovince output applies it, with the monthly shares. Seeresearch/water_balance_method.md. wet_mm_to_m3s()gainsdays;wet_month_days()is new.
- Inputs: CGIAR AET (
-
The water balance's annual AET is now CGIAR floored by a Fu-Budyko AET, chosen from four candidates by a rule fixed before scoring (#15). Blocked-CV mean absolute error on annual runoff drops from 33.1 % to 27.7 % (headwater basins 38.1 % to 31.2 %; 32.3 % under a fully nested selection), mostly on the dry interior plateaus.
- New:
wet_pet_hargreaves(),wet_aet_budyko(),wet_aet_landcover(),wet_chapman_table3(),wet_landcover_nrcan(),wet_terraclimate_aet(). - Chapman's land-cover AET ratios, built from their published Table 3,
make the interior worse; see
research/water_balance_method.md. scripts/wb_validate.Rscores one AET variant per run, andscripts/wb_aet_compare.Rapplies the rule. Shipping is tied to the comparison and to the current scoring code.
- New:
-
MOD16A3GF v061 annual ET (
wet_mod16_aet()) was scored as a challenger to the shipped AET under a rule fixed before scoring (#18). It improves on CGIAR alone but not on CGIAR floored by Fu-Budyko, which stays: headwater MAE is 36.3 % for MOD16 and 34.6 % for max(CGIAR, MOD16), against 31.2 %.scripts/wb_aet_compare.Rnow re-runs #15's comparison, which must reproduce, before the #18 one. Seeresearch/water_balance_method.md. -
The input builders
wet_cgiar_aet(),wet_dem_glo90()andwet_climr_normals(), and the hydrologic-zones grid inscripts/wb_inputs.R, now key their caches on a method version (and the normals on climr's version too), so a builder change never reuses an old file (#19). Existing caches rebuild once under the new names; re-run under them, the water balance reproduces every shipped number. -
Flow departure at hydrometric stations (#25).
wet_station_daily()gives one daily flow series per station, from approved HYDAT through ECCC provisional daily means (the water-temp-bc archive) to real-time. Each source continues the one before it, from its whole record. Ported fromngr::ngr_hyd_q_daily().wet_window_stats()summarises any daily series over month-day windows once per year, in cd's long format, andwet_windows_calendar()supplies the calendar windows. Departure and trend come from cd.scripts/station_departure.Rruns it for real-time stations. At Buck Creek and Bulkley nr Houston, 2023-2026 summer and early-fall flow ran 54-87 % below the 1981-2010 mean. Provisional winter flows are uncorrected ice readings; seeresearch/station_flow_departure.md.
-
A pkgdown site and the first vignette, Chinook flow at two stations (#27). It maps Buck Creek and Bulkley nr Houston and their nested catchments, then compares the last five years with 1981-2010 in the three open-water Chinook windows (migration, spawning, fry migration), from bundled data (
data-raw/station_vignette_data.R,data-raw/station_vignette_map.R). Every window at both stations was below its 1981-2010 mean in 2023-2026, and 21 of the 24 station-window values from those years were drier than nine baseline years in ten. -
A second vignette, Mean annual discharge per segment (#28). It reproduces fwapg's discharge on the 9,000 Salmon River (SALR) segments it gives a value, and shows how area-weighted sampling moves small watersheds. It maps discharge per segment for SALR and for the Bulkley (BULK), where fwapg has none and the open water balance is the only estimate, and plots the balance's blocked-CV skill at 290 HYDAT gauges by zone. On SALR the balance runs a median 1.47 times PCIC. At the four gauges within 30 km of SALR and its outlet gauge, the balance is high at all five. PCIC is within 5 % at the two small basins and low at the three large ones. Data from
data-raw/segment_vignette_data.Randdata-raw/segment_vignette_map.R.