Repository navigation
Deck and charts pipeline
In one sentence: everything under doc/ turns the sweep logs into deck_data.json, the charts, the deck lambda_sweep5.pptx and the CSVs JRC receives, without touching a solver — this page says which script does what and how to redo it after a re-sweep.
Where this sits: at the very end of the pipeline (GeoDMS → Arrow → Julia LP → sweep → rounding → scenarios → deck). Inputs: the sweep logs of Log lines and output files and, for the "today's network" slides, a GeoDMS descriptives table. The formulas live on Scenarios S1 S2 S3 and Metrics aggregation and ranking; this page is the machinery, and which script produced which figure on Pharmacy results September 2026. Units: λ =
w· €100 000, a placeholder (The lambda sweep and the Pareto frontier); every count is 1 km² cells with at least one pharmacy, not pharmacies (Log lines and output files). All numbers are the state at commite25416b(10 September 2026), documented on 15 September 2026.
A sweep takes minutes to hours per area (Running a sweep end to end); 42 areas × 2 travel-cost functions, re-swept several times over the summer of 2026, made a hand-edited deck impossible to keep honest. Since 0c8aabb (10 June 2026) the deck is generated "from the sweep logs, not hand-edited" (doc/README_deck.md); only three concept slides and the hand-drawn Netherlands maps slide survive from the May 2026 deck lambda_sweep4.pptx.
Nothing in doc/ solves anything: every script starts from the committed doc/deck_data.json, so a fresh checkout regenerates everything downstream in seconds. Only the parser cannot run (logs/ is git-ignored, on Maarten's machine), nor the two Arrow readers region_typology.jl and collect_descriptives.jl (Branches data and environment). The slide build has two stages because pptxgenjs (Node) can only create a .pptx; carrying over existing slides and re-ordering needs PowerPoint itself, driven from PowerShell through COM (the Windows interface for scripting a running application).
From doc/README_deck.md (prefix every Python call with PYTHONIOENCODING=utf-8 on Windows, because the scripts print λ and €):
python build_deck_data.py # logs/sweep_*.log (+ refine_*.log) -> deck_data.json
python build_charts.py # -> charts/<AREA>_<FN>.png, AGGREGATE_<FN>.png, logistic_compare.png, agg_data.json
python tangent_chart.py # -> charts/lambda_tangent.png (not listed in README_deck; the next step needs it)
node build_deck.mjs # -> region_summary.pptx (npm install once)
powershell -ExecutionPolicy Bypass -File merge_deck.ps1 # lambda_sweep4 concept slides + region_summary -> lambda_sweep5.pptx
python frontier_metrics.py interp3 # -> frontier_metrics_interp3.{json,csv}
METRIC_SUFFIX=interp3 python frontier_charts.py # -> pointcloud_*, rank_lambda_*, rank_area_*, rank_area_rel_*
python lambda_axis_chart.py # -> charts/lambda_axis_AGGREGATE.png, lambda_axis_Netherlands.png
powershell -ExecutionPolicy Bypass -File rebuild_analysis_slides.ps1 # appends the 10 analysis slides (from the repo root)
python export_issue48.py # -> issue48_crossings.csv, issue48_sweep_results.csv (on request)
flowchart LR
L[logs/sweep_*.log<br/>refine_*.log] --> P[build_deck_data.py]
P --> J[(deck_data.json)]
J --> C[build_charts.py] --> PNG1[area + AGGREGATE PNGs] & AGG[(agg_data.json)]
J --> T[tangent_chart.py] --> PNG4[lambda_tangent.png]
J & AGG & PNG1 & PNG4 & D[(pharmacy_descriptives*.csv)] --> M[build_deck.mjs] --> R[region_summary.pptx]
R & B[lambda_sweep4.pptx] --> MD[merge_deck.ps1] --> DECK[lambda_sweep5.pptx]
J --> F[frontier_metrics.py] --> FM[(frontier_metrics_interp3)]
FM & PT[(policy_typology.csv)] --> FC[frontier_charts.py] --> PNG2[rank + pointcloud PNGs]
J --> LA[lambda_axis_chart.py] --> PNG3[lambda_axis PNGs]
PNG2 & PNG3 --> RA[rebuild_analysis_slides.ps1] --> DECK
J --> E[export_issue48.py] --> CSV[(issue48_*.csv)]
Three area counts recur, all correct at once: 44 names in the parser's ORDER list, 42 areas in the committed JSON (last regenerated 3 September, de5b178, before Hungary and Finland joined ORDER in 8963d27 and 2aeb6c0), and 41 in the aggregate, because country-Poland overlaps its seven NUTS-1 parts and is dropped (Metrics aggregation and ranking).
The descriptives slides (p3–6) need two CSVs from GeoDMS, not logs: make_descriptive_table.bat (repo root) runs GeoDmsRun on /Analyses/Pharmacies/Descriptives/Table per study area and collect_descriptives.jl gathers the Arrow rows into the committed doc/pharmacy_descriptives.csv and _nuts1.csv. The bat's header "no network needed" is stale since df77729 (23 June 2026): catchments are road-based, so the Existing side's network and OD matrix must exist first (GeoDMS OD matrix and choice set; the indicators are defined on Metrics aggregation and ranking).
parse(path) scans one logs/sweep_<area>_<FN>.log three times (around lines 70–146):
- The baseline block: the first
facilities used (cells):,total travel cost(c):andmean travel time (min):lines. - The
Combined sweep, sorted by w:table: every line after thewheader with at least 12 whitespace-separated tokens; tokens 0, 2–7, 9 and 10 become the keysw,sum_y,relax,topp,greedy,multi,n_open,mean_tandfrac(meanings on Log lines and output files).multi, the multistart travel, is the drawn curve. A row whose rounding columns are-failsfloat()and is skipped on purpose: the LP-only rows of a refine-only run (Scenarios S1 S2 S3). - The
S1 —/S2 —blocks, wherestranded : topp a greedy b multi cbecomesst = [a, b, c]— the clients each rounding leaves with no open pharmacy in their choice set, priced at BIG (Glossary). Each scenario is matched to the frontier row with the nearestsum_y, soscen.wis a snapped grid point — the first of the three S1/S2 definitions on Scenarios S1 S2 S3, and the one the area-slide tables print.
main() walks ORDER and honours DECK_ONLY. The refine-log rule (around lines 171–179): if logs/refine_<area>_<FN>.log is newer than the sweep log and holds both S1 and S2 summaries, its rows are merged into the frontier and its scen and baseline replace the sweep's — how the bisection-pinned S1/S2 of issue #52 enter without a full re-sweep. The trap: rows solved in two runs, possibly on a different OD or candidate set, are drawn as one curve — the provenance artefact behind the kinks of issue #54 (The lambda sweep and the Pareto frontier; per-area provenance on Branches data and environment).
render(region, fn, fd) (around lines 87–169) draws one PNG per area and function. interp_s1_s2 (lines 38–59) places the S1/S2 markers by chord interpolation on the drawn curve, not at the snapped scen rows, because a snapped row can sit a grid step off today's count (cb21c6c, 22 July 2026; SE2's snapped S1 had 309 open for 452 cells — Decision log; the three definitions on Scenarios S1 S2 S3). The tables on the same slide still use the snapped rows (deck p9 acknowledges it; no further reason recorded).
window (lines 62–79) clips the x-range to 0.6–1.7 × the anchors S1, S2, today's count and the smallest sum_y whose travel is above today's (the last anchor, 3a70448, keeps S2 on the drawn curve; the factors have no recorded rationale). aggregate_fd(data, fn) (lines 228–274) builds the all-areas frontier and render_aggregate writes it to agg_data.json (rule on Metrics aggregation and ranking); render_logistic draws logistic_compare.png (30/15 versus 25/10 logistic, deck p64 — Travel cost functions).
Netherlands LINEAR only, W1, W2 = 0.1, 0.5 (line 11). On the lower-left envelope of (n_open, multi) it draws for each w the line of constant total cost through the point that minimises it — slope −λ, every other swept point above it (The lambda sweep and the Pareto frontier). The optima it prints ("3,220 open, 22.6 M" and "831 open, 74.7 M") are typed into tangentSlide in build_deck.mjs by hand. The slide answers REGIO review comments 41–42: show the tangency before the frontier.
Node + pptxgenjs (^4.0.1). Flags: --only <REGION> (one area slide, no concept or summary slides), --no-summary, --out <file> (default region_summary.pptx). The assembly (lines 814–836) emits every slide type in deck order, with the two lambdaTableSlides (country / NUTS-1) split by the hard-coded COUNTRY_SET (line 488); the merge script then moves the descriptives to p3–6, a REGIO-review request since f46bbf8 — show what exists before what we optimise (Decision log). Two functions carry the numbers people quote:
-
regionSlide(lines 38–92): a baseline strip (cells,mean_t, LIN/LOG cost), the two charts, and a tablescenario · w · n_open · frac · multi vs relax · stranded (multi)from the snappedscenrows;multi vs relax=(multi − relax)/relax, the multistart-to-LP gap (From LP fractions to real pharmacies);strandedisst[2], the multistart's count. -
lambdaTableSlide(lines 489–563):interp(rows, key, target, out)walks the w-sorted rows and, at the first adjacent pair whosekeybracketstarget, returnsoutlinearly interpolated — except thatout = "w"is interpolated in log w, then ×FMIN = 100000, because the w-grid is geometric (4013c91only calls it "the λ-table rule"). Chord definition and the Netherlands worked example: Scenarios S1 S2 S3.
summarySlide (lines 152–193) has two hard-coded cards, "+0–18 %" and "+0–4.7 %", and a table cell +x% (n) = the scen-row gap with st[2] in brackets.
Parameters -Base (lambda_sweep4.pptx), -Insert (region_summary.pptx), -Out (lambda_sweep5.pptx), -KeepFirst 4, -MapsSlide 6 (0 = skip). Through COM it keeps the base's first four slides, inserts the generated file after them, drops the base title slide, puts base slide 6 (the hand-made NL maps) right after the Netherlands slide, and Move-ByMarker moves each generated concept slide into place by a text fragment (title → 1, agenda → 2, descriptives → 3–6, … multistart → 12; case-sensitive -cmatch, because the agenda mentions the descriptives in lower case). Markers rather than indices, because the number of leading slides changed several times. Saves with SaveAs($Out, 24) (.pptx OpenXML); never calls $pp.Quit().
Opens -Deck (default doc\lambda_sweep5.pptx, resolved from the repo root), deletes every slide whose title starts with one of the five $prefixes, then appends the ten $charts entries: a title box, the picture scaled to 450 pt height with a hard-coded aspect ratio per chart family ($RANK 0.913, $CLOUD 1.4516, $LAXIS 2.8333), sometimes an italic caveat. Delete-then-append makes it safe to run repeatedly. Keep the file ASCII-only: PowerShell 5.1 reads a BOM-less file as ANSI, so a €, λ or middot in a note string comes out as mojibake (this happened once). It does call $pp.Quit().
Per area and function, metrics_for (around line 113) reads S1 and S2 off pareto_frontier (the third S1/S2 construction in the artefact matrix on Scenarios S1 S2 S3), then the improvement rectangle, the diagonal crossing and lambda_cross — interpolated linearly in count where every other λ interpolation in doc/ is logarithmic. Meaning, formulas and a worked toy: Metrics aggregation and ranking. Outputs frontier_metrics_<suffix>.json and .csv (columns on Log lines and output files); s1_bracketed / s2_bracketed say whether a value is interpolated rather than clamped.
interp1 / interp2 / interp3 differ in nothing but the file name (suffix = sys.argv[1], line 146): the same computation on the deck_data.json of 16 July (e75c73c), 22 July (cb21c6c) and 3 September (de5b178, after the landbody fix and NUTS exclusion). interp1 and interp2 are byte-identical; interp3 differs only in Portugal, ITG, SE2, SE3 and PL8, the five areas the scope rules moved (Scope rules and data gaps). The deck and the results page use interp3.
frontier_charts.py reads frontier_metrics_<METRIC_SUFFIX>.json — default interp1, so set METRIC_SUFFIX=interp3 or you redraw the July rankings — and policy_typology.csv; SKIP = {"Poland"} drops country-Poland. It writes, per function, pointcloud_<FN>.png, rank_lambda_<FN>.png, rank_area_<FN>.png and rank_area_rel_<FN>.png, bars coloured by formalisation level (Glossary); the header's "coloured by DEGURBA-style class" is stale since 559dad9 (17 July 2026).
policy_typology.py writes policy_typology.csv (44 rows): an archetype (T1–T4) and a formalisation level per country, transcribed by hand from slides 11–16 of the untracked, restricted-use JRC deck doc/Typology_Countries_to_share.pptx. region_typology.jl writes the older urban/rural key region_typology.csv (42 rows), which no script reads at e25416b (verified by grep). Class meanings and why the colour is formalisation: Metrics aggregation and ranking.
Reuses metrics_for, pareto_frontier and _interp_at, rebuilds the aggregate as pseudo-area AGGREGATE, and writes issue48_crossings.csv (86 rows = 42 areas × 2 functions + 2 aggregate rows) and issue48_sweep_results.csv (1 530 rows, one per area × function × w; columns on Log lines and output files). issue48_README.md still says 1 626 rows. Requested in issue #48; the 31 August delivery was declared superseded after the scope rules (issue48_superseded.md).
How to read the charts: The lambda sweep and the Pareto frontier. This section maps chart elements to fields and scripts.
Area chart (<AREA>_<FN>.png, deck p16–58 — p17 is the carried-over Netherlands maps slide — one LINEAR | LOGISTIC pair per area; results page §6):
| element | what it is | source field |
|---|---|---|
| x-axis "sum_y (LP-relaxed facility count)" | Σ yⱼ of the LP relaxation — a fractional count; the one chart that plots it rather than n_open
|
rows[].sum_y |
| left y "total travel cost (person·c(t))" | Σ popᵢ·c(tᵢ): person-minutes under LINEAR (axis from 0), dimensionless under LOGISTIC (axis stretched to the data, 8 % padding, so the tiny bound gap stays visible) |
rows[].relax, rows[].multi
|
| blue solid, round markers — "cost upperbound" | travel of the multistart integer solution: a feasible network, so an upper bound (From LP fractions to real pharmacies) | multi |
| grey dashed — "cost lowerbound", drawn on top | travel of the LP relaxation, the certified lower bound; on top so both stay visible where they coincide (71ed7bc, "per Maarten") |
relax |
| amber dashed, right axis "log₁₀(facility cost weight €)" | log₁₀(w · 100 000) — the only place λ appears on this chart | w |
| ★ "baseline" + dotted horizontal | today's network at (cells, cost) |
baseline |
| ◆ S1 + dotted vertical | chord-interpolated on the blue curve at x = today's count | interp_s1_s2 |
| ■ S2 | chord-interpolated on the blue curve where it crosses today's travel | interp_s1_s2 |
Aggregate chart (AGGREGATE_<FN>.png, deck p59, results page §2): the same render() on the summed rows of the 41 disjoint areas; only w-values every area solved are summed, so the slowest sweep sets the upper limit (LINEAR w ≤ 0.5, LOGISTIC w ≤ 0.02 at e25416b).
λ-axis chart (lambda_axis_<REGION>.png, deck p65–66, results page §2; 08de764, 10 September 2026): the same sweep along λ (log x), LINEAR | LOGISTIC panels — travel bounds and today's travel on the left axis, n_open, sum_y and today's count on the right (log, limits shared between the panels). The S1/S2 guides come from crossing(), the deck-table rule (log in w), not the snapped scen rows (header example: NL LINEAR snaps to 1 334 open at w = 0.3 where S1 has 1 615). Since 1ed57f2 each level line says what defines it and the S1 guide shows λ·N, the facility bill at that price.
Ranked lists and point cloud (rank_area_rel_, rank_area_, rank_lambda_ and pointcloud_<FN>.png, deck p67–74, results page §4): one bar per area (41), sorted on rect_rel, rect_area (log axis) or lambda_cross; the point cloud is a log–log scatter of locations against travel with ★ baseline, ◆ S1, ■ S2 and ● the diagonal crossing per area. Clamped values are not marked, so check the CSV flags before quoting a rank (all 84 rows are bracketed in interp3). Axes, sort keys, colour key and the current ranking: Metrics aggregation and ranking.
| results page section | artefact | script and rule |
|---|---|---|
| §1 today's network table |
pharmacy_descriptives.csv, pharm_* columns (deck p3) |
GeoDMS Descriptives/Table → collect_descriptives.jl; #empty = pharm_n_empty (catchment rule on Metrics aggregation and ranking) |
| §2 aggregate table (w, locations, Δ, multi vs relax, Δ travel) |
agg_data.json → aggregateSlide (p59) |
aggregate_fd; S1/S2 are the nearest union rows (snapped), so S1 reads 43 357, not 43 320, although labelled "S1 (same count)"; why snap here while the per-country table interpolates is not recorded |
| §2 "the count collapses from ~433 000 open locations at λ below €100 … to ~43 000 at S1" |
lambda_axis_AGGREGATE.png, _Netherlands.png
|
lambda_axis_chart.py; 433 828 is the aggregate's most-locations row (many in agg_data.json, w = 0.0001); the S1 figure is the crossing() guide |
| §3 per-country and NUTS-1 tables |
lambdaTableSlide (p60–61) |
interp in build_deck.mjs: Δ linear in the crossed quantity, λ log in w. frontier_metrics_interp3.csv, which the page cites, agrees on S1_y/S2_x to rounding but has no λ S1/λ S2 columns |
| §4 rankings (ITF 0.268 …; ITF 52 807 …) and point cloud |
rank_area_rel_<FN>.png, rank_lambda_LINEAR.png, pointcloud_LINEAR.png
|
frontier_metrics.py interp3 (rect_rel; lambda_cross, w linear in count) → frontier_charts.py
|
| §5 "+0–18 % / +0–4.7 %" |
summarySlide cards (p62) |
hard-coded text in build_deck.mjs (lines 165–167, repeated in optProblemSlide line 622); not recomputed |
| §5 "SE2: 56 … SE3: 130" |
summarySlide table |
scen.st[2] of the snapped S1 row — the multistart's stranded clients |
| §6 "1 615 cells, 3.48 min", "S1 at 1 334 open (w = 0.300), S2 at 1 180 (w = 0.341)" |
regionSlide strip and table |
baseline and the snapped scen rows — not the chord-interpolated S1 of the chart on the same slide, which by definition sits at 1 615 |
In short: the area-slide tables and p62 quote the snapped row; the chart markers, the p60–61 tables and the λ-axis guides quote the chord; frontier_metrics.py uses the Pareto envelope — the artefact matrix on Scenarios S1 S2 S3.
Area X was re-swept (a new logs/sweep_X_*.log or a newer refine_X_*.log). There is no incremental mode; the chain re-reads everything, cheaply:
-
python build_deck_data.py— needs all logs: an area without one is skipped silently and vanishes from the deck and the aggregate. Compare the printed per-area row counts with the previous run (NL LINEAR: 29 rows at e25416b). -
python build_charts.py— X's two charts change, and so doAGGREGATE_*.pngandagg_data.json. -
python tangent_chart.pyonly if X is the Netherlands (then re-type the numbers intangentSlide). -
node build_deck.mjs, thenmerge_deck.ps1. Quick look at X alone:node build_deck.mjs --only X --out preview.pptx. -
python frontier_metrics.py interp3,METRIC_SUFFIX=interp3 python frontier_charts.py,python lambda_axis_chart.py, thenrebuild_analysis_slides.ps1from the repo root with PowerPoint closed. -
python export_issue48.pyif JRC gets a fresh export;julia export_s1s2_locations.jl Xfor the open-location files (Scenarios S1 S2 S3). - By hand: copy the ten PNGs the wiki shows from
doc/charts/todoc/img/and the wiki clone'simages/(no script does this — Branches data and environment); update the numbers on Pharmacy results September 2026; re-read the hand-typed prose listed under Pitfalls.
For a new area, five lists need an entry first: ORDER/NICE in build_deck_data.py, REGION_ISO in policy_typology.py (then re-run it), COUNTRY_SET and the COUNTRIES set of descriptivesSlide in build_deck.mjs for a country, country_order in collect_descriptives.jl, STUDY_AREAS in make_descriptive_table.bat. At e25416b Hungary and Finland are in all of these except COUNTRY_SET (line 488) and COUNTRIES (around line 325) — 8963d27 put Hungary "in every per-area list" but missed those two — so once their logs arrive they would render as NUTS-1 rows. Whether the unpushed 216ac77 fixed this cannot be checked from this checkout.
From .gitignore: /logs (large, reproducible only by re-solving; on Maarten's machine), /doc/node_modules, /doc/charts (regenerated in seconds from the JSON), /doc/region_summary.pptx (an intermediate), /scratch, ~$* (PowerPoint lock files). Committed on purpose: deck_data.json (the one artefact that needs the logs), agg_data.json, the frontier_metrics_interp* files, the issue48_*.csv, the descriptives and typology CSVs, doc/img/ (the ten PNGs the wiki shows), and the decks lambda_sweep.pptx, lambda_sweep4.pptx (the base) and lambda_sweep5.pptx (the 13 MB result). doc/Typology_Countries_to_share.pptx is referenced by policy_typology.py but not tracked (restricted use).
Open first: doc/README_deck.md, then build_deck_data.py → parse (permalink), build_charts.py → interp_s1_s2 (L38) and aggregate_fd (L228), build_deck.mjs → lambdaTableSlide (L489) and the assembly (L814), frontier_metrics.py → metrics_for (L113), lambda_axis_chart.py → crossing (L59), merge_deck.ps1 → Move-ByMarker (L66). The log lines the parser matches are printed by lambda_sweep_simplex.jl → print_sweep_row (around line 221) and the scenario summary (around line 515); see Code walkthrough lambda_sweep_simplex. parse_compare.py at the repo root (env FUNC) is a console sanity table over the S1/S2 rows of every log, not part of the deck.
| name | where | default | what it changes | why you would change it | interactions |
|---|---|---|---|---|---|
DECK_ONLY (env, "A,B") |
build_deck_data.py |
unset = all of ORDER
|
restricts the JSON to a subset | preview while other sweeps run | the aggregate then sums only those areas — never publish one from a partial JSON |
ORDER, NICE
|
build_deck_data.py |
44 names | deck order, eyebrow and display name | new area | four other lists (above) |
suffix argument / METRIC_SUFFIX (env) |
frontier_metrics.py / frontier_charts.py
|
interp1 |
output file name / which metrics file is charted | always interp3 |
must match each other; the charts silently draw July numbers otherwise |
FMIN / FACILITY_MIN_COSTS = 100000 |
build_charts.py L103, frontier_metrics.py L24, lambda_axis_chart.py L49, build_deck.mjs L490, tangent_chart.py L11 (FMC) |
100 000 | the placeholder €/location behind every € label and lambda_cross
|
if the real fixed cost is ever calibrated |
hard-coded in five files, not read from settings.jl; only € labels move, not the frontier, the Δs or the rankings |
| positional areas | lambda_axis_chart.py |
AGGREGATE Netherlands |
which λ-axis charts are drawn | any region id |
rebuild_analysis_slides.ps1 only picks up the two defaults |
-
frontier_charts.pydefaults to interp1 and nothing warns: withoutMETRIC_SUFFIX=interp3you get the pre-scope-rule rankings (ITG and Portugal on top). -
Snapped rows in the committed deck.
deck_data.jsonat e25416b is pre-#52 for every area (regenerated 3 September; the #52 bisection landed 10 September): snapped S1 rows up to a grid step off (Netherlands LINEAR 1 334 for 1 615; in FRC, FRH, FRJ, FRK and SE2 LINEAR the S1 and S2 rows are the same point — Scenarios S1 S2 S3), and results page §6 quotes them. -
Five area slides print "undefined". The
scenrows of Portugal, ITG, SE2, SE3 and PL8 (the September re-sweeps) have non_open/fracfields, so their slide tables (p27, p46, p50, p51, p57) show the literal string "undefined"; the parser'sn_openregex (line 122) did not match those logs; why is not recorded and the logs are not in the repository. -
The results page is one data generation behind. Issue #53's closing comment (
216ac77, ~12 September, not pushed as of 15 September) describes a regenerated deck with 44 areas, the OECD Italian list and a different ranking; every number this page cites is the e25416b state (Branches data and environment). -
Hard-coded prose goes stale silently: the "+0–18 % / +0–4.7 %" cards (derivation not recorded; at the snapped S1/S2 the gap is under 3.5 % everywhere), the tangent-slide optima, the choice-set and scope-slide figures, the agenda page references, the aggregate footnote "13 countries + FR/IT/SE/PL", the "41 areas" in
lambda_axis_chart.py(line 167) and in therebuild_analysis_slides.ps1notes. With Hungary and Finland every one of these needs an edit.README_deck.mdlikewise still lists afrac_ycurve (as pre-4cee238frac_x) dropped in9bdae8d(6 July 2026, "overcomplicating") and a "log₁₀(w)" axis relabelled in71ed7bc; the code wins. -
lambda_crossis linear in w, every other λ interpolation is logarithmic. Intent or oversight is not recorded; the ranked list on the results page uses it. -
Duplicate w rows (a coarse grid point re-solved by the fine sweep; NL LINEAR has two w = 0.3 rows) are tolerated by every interpolation but appear twice in
issue48_sweep_results.csvand in the parser's row counts;agg_data.json'sn_wis built from a set and is unaffected. -
region_typology.csvis orphaned; whether it returns or can be deleted is undecided.issue48_README.mdis hand-maintained (its 1 626-row figure is stale). The descriptives' NUTS-1 CSV covers FR/IT/SE only; the Polish NUTS-1 rows the batch computes are in nocollect_descriptives.jllist (reason not recorded — Metrics aggregation and ranking). - The refine-log override also replaces the baseline (
base = rbase or base); a refine run with a differentCLIENT_WEIGHTor exclusion setting would silently move it (not observed; unverified whether intended).
- Both
.ps1scripts need PowerPoint; there is no LibreOffice path (README_deck.md's QA check renders slides to PNG withSlide.Export(…, 1600, 900), also through COM). Run them with-ExecutionPolicy Bypass. -
merge_deck.ps1never calls$pp.Quit()(COM may attach to a PowerPoint you already have open) and may leave a process running;rebuild_analysis_slides.ps1does Quit — close PowerPoint before running it or an open deck of yours goes with it. - Change a
figsizeinfrontier_charts.pyorlambda_axis_chart.pyand you must update$RANK,$CLOUDor$LAXIS, or the picture is distorted. - Slide positions come from marker text: rename a slide title in
build_deck.mjswithout updating itsMove-ByMarkerpattern and the slide stays where it was inserted; change a title in the$chartstable ofrebuild_analysis_slides.ps1without its$prefixesentry and the old slide is duplicated, not replaced.
- Pharmacy results September 2026 · Log lines and output files · Running a sweep end to end
- Scenarios S1 S2 S3 · Metrics aggregation and ranking · The lambda sweep and the Pareto frontier · From LP fractions to real pharmacies
- Scope rules and data gaps · Travel cost functions · GeoDMS OD matrix and choice set · Alternatives not pursued (the cap-ladder slide, p15)
- Code walkthrough lambda_sweep_simplex · Branches data and environment · Decision log · Glossary
NetworkModel EU documentation
Start here
Understanding the method
- The facility location model explained
- Travel cost functions
- The lambda sweep and the Pareto frontier
- From LP fractions to real pharmacies
- Scenarios S1 S2 S3
- Scope rules and data gaps
- Metrics aggregation and ranking
- Background theory
Working with the code
- GeoDMS OD matrix and choice set
- Code walkthrough settings
- Code walkthrough lp_run
- Code walkthrough lambda_sweep_simplex
- Running a sweep end to end
- Log lines and output files
- Deck and charts pipeline
- Installation of Julia
Reference
- Lambda sweep method
- Pharmacy service access
- Pharmacy results September 2026
- Decision log
- Alternatives not pursued
- FAQ
- Glossary
- Service allocation procedure
External