Every species is retuned, graded and compared on link’s new provincial default build, and that run is promoted into the published tables (#145). The build models neither rainbow trout nor kokanee, so knw_tune_run(modelled =) writes not_modelled for a species with no access value in the evidence, and rainbow trout rearing’s published rule is retired. Against the previous run, bull trout spawning, chinook rearing (11 to 13 %) and grayling rearing move outside its region, and bull trout rearing’s move is a reshuffle among near-equal rules (research/habitat_thresholds_tuned.md). knw_build_record() and 0300 read only the rows of link’s run log that belong to the build, since a provincial run leaves the previous run’s rows in place, and 0300 takes SETTLE_TAG= for a build compared before it is promoted. Both articles follow the new run.
knowledge 0.4.26
The chinook rearing article leads with the answer and the Bulkley map, then one before/after table, a figure of why a nearby river matters (the fitted chance of a catch by distance to a river, per stream flow), how sure the choice is, and how it was done in a table that links the methods article (#143). The methods article opens with what bcfishpass’s map was for and what tuning adds, and now holds the one-SE and cross-validation citations. Several statements are now exact where the old prose was loose: the Bulkley map’s three causes of change, a coded site out of reach is dropped, and each site and group count names its population. No code change.
knowledge 0.4.25
knw_tune_anchor() sets each tuned key’s pick, from link’s default and from its flow version, beside the rule the data score best: how far behind it sits, in deviance and SE, and how many parts (size test, gradient, exception) apart (#141). It also measures each pick against the best rule a pick may be (in the region, with no exception under d_floor), since a best a pick may not be can never be picked. scripts/0320 writes these columns into every summary and pick_report.csv, and data-raw/tune_anchor_check.R reads saved tables. On the published tables the starting point steers grayling, steelhead and coho rearing, and on the minimal network bull trout rearing too (research/baseline_anchoring.md, with grayling rearing’s “before” for #140).
knowledge 0.4.24
Every species was tuned on link’s provincial minimal network (#120). That network rejects none of the published rules: each stays within one standard error of the best rule. Where a pick moves, it moves because the pick is the rule closest to link’s default, so each tuned rule is only as good as its starting point; recording what each baseline rests on is #139, and grayling’s first (#140). The results are research/minimal_network_tuning.md, and the candidate run’s tables are archived (data/archive_index.csv), not promoted. No code change.
knowledge 0.4.23
The tuning runs on a second link build beside the published one (#120). 0310 writes a build of another schema into its own EVIDENCE_DIR, with a tagged ledger, and refuses the default directory to it; 0300 takes SCHEMA and POLL_MIN, and 0310 takes SETTLE= and PARENT_DIR, joining each visit’s minimal-network parent through the new knw_parent_join(). 0320 takes DEV_DIR, GRADIENT=parent_gradient and test floors (MIN_WSG, FOLDS), which knw_tune_promote() refuses; knw_link_access_check(by_group = TRUE) gives the check per group. knw_build_record() reads link’s output tokens from <schema>.log, where link writes them, so a tokened build no longer records none. The four-group test on link’s minimal network is research/minimal_network_tuning.md.
knowledge 0.4.22
inst/NOTICE.md gives the PSCIS snapshots under data/<wsg>/ their own section: incidental working copies of an “Access Only” catalogue layer, not a distribution of PSCIS (#76). EcoCat, bcfishpass and the modelled flows stay under “Terms not yet settled”. No code change.
knowledge 0.4.21
The repository is public (#56). The README’s Learn more links point at the published site: the chinook rearing article, the methodology article (newly linked), the function reference and the news page. No code change.
knowledge 0.4.20
A published article bundle can be rebuilt on any machine (#119). scripts/0320 writes a tracked record beside each grade table (<key>_grades_record.csv, read with knw_grade_record_read()), which the bundles pin instead of the gigabyte tables, and knw_tune_promote() refuses a key whose record is missing or describes another table. A run’s grade tables, records, curves and segments, and its evidence build once, are kept in a private, versioned archive (knw_archive_plan(), knw_archive_put(), scripts/0336-archive.R), indexed by version id in data/archive_index.csv; PULL=TRUE (knw_archive_pull()) restores the run a checkout’s records name, md5-checked, and scripts/0335 archives each promotion. knw_build_record() reads link’s output-vocabulary tokens. The methodology article’s limits carry the measured site gradient against link’s segment gradient (#121), computed by knw_site_gradient_visits() and knw_gradient_flips().
knowledge 0.4.19
R CMD check runs on every pull request and push to main through New Graph Environment’s shared workflows (NewGraphEnvironment/flows), so a broken example, test or documentation page shows on the pull request (#98, #99). Its first runs cleared one warning and two notes: knw_tune_smoke() documents folds, the fixed-effects fit calls stats::.lm.fit(), and seven examples are wrapped under R CMD check’s Rd line width. No change in behaviour.
knowledge 0.4.18
knw_compare_rules() logs which rules’ predicates read the near-river table and returns bin_minutes, each worker’s timed minutes with and without default’s own (and without the walk where no other predicate reads that table), so a run projects what reading default’s flags instead of regenerating them would save (#101). On #112’s runs default took 51 to 74 % of a two-rule key’s classify, overlay and connect worker minutes. The cache is parked until link’s reworked stream layers are built (research/database_builds.md).
knowledge 0.4.17
scripts/0330-compare.R compares link’s default and the reported structure (door) alone unless STRUCTURES=all (#112): the rules it no longer regenerates took 29 to 33 % of the worker minutes on three keys. A reported-only run matches the every-structure run’s default and door rows (data-raw/compare_structures_check.R). knw_compare_rules(structures = "reported") resolves the structure, refuses one outside the species x rule’s own set, and returns each rule’s minutes (worker_minutes). The chinook article no longer quotes the free structure.
knowledge 0.4.16
Retuning no longer touches the tracked tables or the articles until the operator promotes a run (#111). DEV=1 0320 is the check after a change upstream: minutes, no database, ending in a pick report (knw_tune_pick_report()) of each key’s pick beside the published one. A full 0320 and 0330 TUNE_DIR= write ignored candidate runs, and scripts/0335-promote.R (knw_tune_promote()) moves one in, whole or by key, archiving what it replaces so a promotion can be undone, then reruns 0325 and the article bundles. Before anything moves it refuses a set that 0325 or the bundles would refuse. Summaries record their run and evidence ledger, which the readers take through knw_run_ledger() rather than the newest ledger on disk.
knowledge 0.4.15
A database-bound run checks its connection’s round trip before any work (#110). knw_db_latency_check() stops 0310, a grading 0320 and every stage of 0330 over a 20 ms round trip, with the round trip and what a remote link cost #92’s comparison, unless REMOTE=TRUE; every run logs the round trip and whether the override acted, and a refused run writes its log to its own file. 0330 also logs each key’s minutes by stage, which knw_compare_rules() returns as timing.
knowledge 0.4.14
Habitat rules and curves are tuned on the gradient of the link segment a site sits on (segment_gradient), the gradient link applies a rule to and the grade reads, instead of the whole Freshwater Atlas line’s; knw_tune_run(cols = ) names the columns, and every grade table, summary and curve records the gradient it was fitted on (curve_grad), which 0320, 0325, 0330 and the article scripts compare (#92). GRADIENT=fwa_gradient 0320 scores the Atlas line’s gradient as a sensitivity. Every key is retuned on #91’s evidence and both articles redrawn. link cuts a segment at a fish record, so 86 % of sites with fish start a segment of their own against 10 % without, and the results are on link 0.66.0’s segments, before link changed where it cuts streams.
knowledge 0.4.13
FDIS site visits are placed on the stream their own 1:50k watershed code names where one lies within 500 m (knw_fdis_place()), otherwise on the nearest stream within 200 m, with fresh::frs_point_snap() in place of fwa_indexpoint() (#91). The limits were set by measurement. The evidence keeps 45,529 visits against 40,552. 0310 takes SCHEMA and caches its walks. The tune tables and the articles still describe the #73 run until #92’s retune.
knowledge 0.4.12
The methodology article, “Choosing habitat thresholds from where fish were caught”, now runs from what a threshold is straight to the flow chart and the eleven steps; how rules are told apart and where the fish sampling records are lost follow the worked snap example, before the limits.
knowledge 0.4.11
A retune is fast enough to iterate on the database’s host (#87). Grading walks 500 segments per query (knw_link_walks(batch =)) and keeps every walk in knw.walk_cache (cache =), keyed on the flow load (knw_discharge_load()) and the walk’s code (knw_walk_md5()), so one key regrades in about 4 min with its walks cached, and a cold regrade of every key took 53 min, against 191. knw_compare_rules(workers =) spreads a key’s watershed groups over several connections, so one key recompares in 4.5 to 37 min (BT rearing 33, against 143); the full recompare is still about 2.9 h, bound by the database’s cores (#101). Every grade table records its walk_md5, which 0320, 0325, 0330 and data-raw/vignette_chinook.R compare, and 0320 and 0330 refuse a schema built by another run at connect (knw_build_record()). knw_discharge_load() refuses a flow comment that does not name the historical load the walk reads. research/database_builds.md covers what a run reads and what a second database needs.
knowledge 0.4.10
A methodology article, “Choosing habitat thresholds from where fish were caught” (vignettes/habitat_thresholds.Rmd), walks the pipeline from the FDIS records to the published thresholds in plain language: what a threshold and a tuned threshold are, how rules are scored and chosen (the standard error, and 1 against 1.96 SE), each of the eleven steps with its code and settings, where the fish sampling records are lost, and every setting that shapes a result, with where it came from and whether it has been tested (#93). The settings are inst/method/settings.csv, each naming where its value is held, and test-settings-register.R fails when the register and the code disagree. The chinook article now says how sites reach the network, where segment flows come from and which gradient was scored, and cites PCIC’s VIC-GL, Raven and the fwapg fork (#89).
knowledge 0.4.9
Every species’ threshold is retuned on rebuilt evidence (#73): FDIS visits on wet’s 2026-10-19 flows, the walk to a river ending at a double-line river’s bank with every hop’s gap counted (fresh 0.42.0’s rule), a gradient ladder to 30 % for every species, and a second pick read from link’s flow version of default. The scorer was reorganised after two defects surfaced on this evidence: the held-out intercept search now minimises the deviance directly, and knw_tune_smoke() gates every key with a minutes-long run before a grid is scored (#84). The fit absorbs the watershed-group intercepts and carries a normal prior (SD 2.5) on the data-set and era terms (#83), which made rainbow trout feasible again. The fitted curve takes the same prior, and grades are published only for a key whose curve predicts held-out groups better than no rule on the curve’s own visits: bull trout rearing, chinook, grayling and rainbow are graded; bull trout spawning, coho, steelhead and westslope cutthroat carry a rule and no grade. knw_compare_rules() sets each tuned rule beside default on link’s segments under its through-wetland structures, with a GeoPackage per key; the thresholds table gains near_river_scope and evidence_md5; every table a later run reads back records the gate, the evidence and the pick, and every reader refuses a mismatch. The chinook article and its data are regenerated on the new tables, and the eight research files are rewritten.
knowledge 0.4.8
The tuning’s “within one standard error of the best” is stated as what it is, a parsimony rule of thumb fixed before the run, with its sources, and the same choice at 1.96 standard errors is published beside it (#79). knw_tune_repick() replays the pick from a saved tune table at any width, and data/thresholds.csv gains rule_196 and holds_196 (still knowledge/thresholds/v1, not uploaded). At 1.96, five of eight species × stages keep default, though for four of them the stricter bar takes in most or all of the rules; chinook keeps its 0.75 m³/s flow floor at default’s 5.49 % with no near-a-river allowance. The chinook rearing article is shorter and plainer, with its sources checked against the papers, a bibliography that data-raw/vignette_bib.R writes from Zotero with the run’s software versions, the line-graph figure marking close calls at both bars, a chinook 1.96 table, and on the website an interactive Bulkley map on Esri’s World Topographic Map (knw_compare_mapgl()). The README credits fwapg, bcfishobs and bcfishpass to Simon Norris.
inst/CITATION reads a Zenodo concept DOI from DESCRIPTION’s Config/DOI when the package has one. It has none yet, so the citation is unchanged.
knowledge 0.4.7
The README is rewritten to the README convention (#75). It says what knowledge is and the thresholds it tunes for link’s habitat models, maps chinook rearing in the Bulkley under link’s default rule and the tuned rule (data-raw/readme_figure.R), shows how to install the package and reach the published FDIS data, and reads data/thresholds.csv in a short example. DESCRIPTION’s title and description say the same. The package gains a hex sticker, a citation (citation("knowledge"), and CITATION.cff kept current by a workflow) and inst/NOTICE.md, which names each outside data source with its terms; the terms of the PSCIS layers and the EcoCat- and bcfishpass-derived files are open in #76. knw_compare_map()’s title is documented as what it does: it names the interactive page and the message when nothing is drawn, and the static map draws none.
knowledge 0.4.6
The grade reads every segment at its own distance to a river (#65). knw_grade_species() used to walk only the segments the rule’s exception could admit and read the rest at the 50 km distance cap, which on the joint size and distance curve understates any segment near a river: every key was affected, and six of the eight walked nothing at all. It now walks every accessible segment with a gradient to the pick’s exception river (10 m³/s where it has none) and records dist_status. The reference rate that reads 1 is now taken over the stream branch, the edge types the curve was fitted on. Over every accessible segment, river-polygon segments would make up most of the top 5 % for chinook, coho, steelhead and cutthroat and raise their reference 1.4 to 4 times (chinook 0.14 to 0.52); for bull trout, grayling and rainbow the two references are within 7 %. knw_grade() gains ref_rows and knw_grade_species() gains ref_edge for this; its edge_stream, which limited the walk to the stream branch and which no caller set, goes with the walk’s old limits. habitat_grade/v1, never uploaded, is regraded before its first publication, and gains stream_branch: off the stream branch, on lines through lakes, wetlands and double-line rivers and on other connectors, the grade extrapolates the curve past its visits, and those lines hold 21–79 % of each key’s km graded 0.8 or more. The thresholds, default‘s km and the width rules’ km do not move; the flow-ruled tuned rules move by under 0.3 % because wet republished the segment flows the same day (NewGraphEnvironment/wet#103), and retuning on those flows is #73.
knowledge 0.4.5
knw_tune_run() gains grid_args, passed to knw_tune_grid(), and reads a cached score back only for the grid it was scored on (#67). scripts/0322 scored chinook rearing on gradient limits up to 30 %: the best rule moves from 9 % to 15 %, and every limit from 7 to 30 % has rules in the region, but choosing among the longer ladder loses to default under nested selection. CH_rear_v1 keeps 9 %; extending the ladder for every species is #68. The chinook rearing article says so.
knowledge 0.4.4
A pkgdown article, vignette("chinook_rearing"), explains in plain language how the chinook rearing threshold was chosen from the province’s electrofishing records and what it changes on link’s segments, with a Bulkley map, a zoom and three labelled example segments (#62). Its data are written by data-raw/vignette_chinook.R to inst/vignette-data/, and a test fails when a tracked source moves on. knw_compare_map() gains ncol, panels, mark, lwd, legend and labels; its default output is unchanged. The package now declares R >= 4.1.0.
Eight species × stage fit. default is outside the one-SE region for every species but grayling; chinook’s size floor settles at 0.75 m³/s; bull trout spawning fits but is weak. data/thresholds.csv and data/thresholds_groups.csv carry the result in link’s parameter names; 0290, 0294 and 0296 write their tables from the new build, and 0291–0298 read the published v2 visit tables (0299 PULL=TRUE VERSION=v2). Results: research/habitat_thresholds_tuned.md and one file per species.
knowledge 0.4.2
The README points at the published FDIS visit tables’ second release, knowledge/fdis/v2/, which keeps one copy of a survey loaded under several collection ids and carries the crosswalk from v1’s site_visit ids (#35).
knowledge 0.4.1
The package now sits at the repository root and installs with pak::pak("NewGraphEnvironment/knowledge"). DESCRIPTION gains URL (the new pkgdown site and the repository) and BugReports, and the knw_pscis_pull() example connects through the standard PG* environment variables.
knowledge 0.4.0
New evidence variable instream_flow: a flow need, the water a life stage needs while it is in the stream, kept apart from discharge, which is stream size (usually the MAD an intrinsic-potential model filters on). It takes m3/s, cfs, percent MAD, a percentile or a percent exceedance, and a metric naming the flow statistic it is held to (instantaneous, daily_mean, 7day_low, monthly_mean, monthly_median, unstated). It carries no parameter or model_id, and the *_mad_* parameters now sit on discharge rows only. A 7day_low row needs a note giving its return period.
depth and velocity read in the unit the source wrote: cm and ft for depth, cm/s and ft/s for velocity, as well as m and m/s. A channel’s bankfull depth is stream size and reads as m bankfull.
Rows that changed variable or unit, so a filter on the old values returns fewer rows: data row 121 (a migration low flow, now instream_flow and coded SALMON), rows 250, 257 and 264 (bankfull depths, now m bankfull), and rows 497 and 498 (redd velocities, now cm/s as quoted).
knowledge 0.3.1
New evidence variable groundwater: upwelling, hyporheic exchange and springs. It takes no unit, so a groundwater row is a statement with no value, until a source gives an upwelling measure.
knowledge 0.3.0
knw_literature_read() requires a metric column, after unit: which statistic a temperature value is (daily maximum, MWMT, 7DADM, MWAT, 7DEC, a weekly, monthly, August, summer or annual mean, a constant laboratory exposure, or unstated), or which window a degree-day total accumulates over (growing_season, to_hatch, to_emergence, annual). A table without the column is refused. A temperature-family value needs a metric that fits its variable; a none row may name the one it searched for; every other row leaves it blank.
New evidence variables degree_days (degC-day) and air_temperature (degC). Air temperature is its own variable because intrinsic-potential models use it as a proxy for water temperature. A cited degree_days row needs a note giving its base temperature.
A none row may carry bound, naming the limit that was searched for.
New kindlab: a laboratory or hatchery measurement (a lethal limit, CTmax, survival or growth at held temperatures, a development time in degree-days), which is not expert opinion.
A qual row is refused a value: it is a statement, and a value on it would be counted as evidence it is not.
knowledge 0.2.1
knw_literature_vocab() adds elevation to the variable vocabulary, in m or ft, so a source that fits an elevation term (a watershed-scale occurrence coefficient, say) can be recorded. Other units are refused.
knowledge 0.2.0
knw_ip_read() and knw_ip_vocab() read and validate the intrinsic-potential model registry: one row per model, species and life stage, recording how the model sizes streams (flow_metric, flow_source, flow_inputs, flow_surrogate), its valley-confinement variable and whether it was validated against fish, each claim with a verbatim quote.
knw_literature_read() requires a model_id column (its value may be blank), linking a threshold to the registry model that uses it. A table without the column is refused.
parameter is now a controlled vocabulary (the habitat-model parameter names, in knw_literature_vocab()), so a misspelt or invented parameter is refused rather than silently matching nothing.
New evidence variables valley_width_index (ratio) and drainage_area (km2, mi2); cfs for discharge and ft for channel width are read as the source wrote them.
knowledge 0.1.0
First version: PSCIS pulls by watershed group, GNIS collision checks, size classes, and readers for the life-history timing and literature evidence tables.
