Skip to content

Log lines and output files

Jip Claassens edited this page Sep 15, 2026 · 1 revision

Log lines and output files

In one sentence: the reference for everything a sweep writes — the log line by line and column by column, the Arrow files, deck_data.json and the CSVs in doc/ — so that you can read a number off any of them and know its unit, its origin and its traps.

Where this sits: GeoDMS exports three Arrow files per side (Existing and New) per study area → lambda_sweep_simplex.jl reads them, prints one log per area and travel-cost function and writes per-point Arrow files → doc/build_deck_data.py parses the log into deck_data.json → charts, tables and CSVs derive from that JSON. Running it: Running a sweep end to end; meaning: Scenarios S1 S2 S3, Metrics aggregation and ranking; code: Code walkthrough lambda_sweep_simplex. As of September 2026 (commit e25416b).

Why the log is the primary artefact

The sweep prints to standard output; the batch scripts redirect it to logs\sweep_<area>_<FUNC>.log (refine_ for a refine-only run), overwritten on every run. Nothing downstream re-solves an LP — a point takes minutes to hours — so build_deck_data.py parses the log and every chart, table and CSV derives from it (doc/README_deck.md: "Sweep logs are the inputs; regenerate them only when re-running the optimisation (slow)"). The log's format is therefore an interface: the parser keys on line prefixes and token positions, and a cosmetic println change silently drops rows (Pitfalls).

Three voices are mixed: the driver's pln lines (println + flush, readable mid-run), Julia [ Info: … lines from the loader, and HiGHS's own iteration output (output_flag = true in build_lp_warmstart, "show simplex progress live"; no reason recorded — likely so that a stalled worker can be diagnosed from the log alone).

Anatomy of a sweep log

Line texts are from analyze_country (around lines 245–548). No current Netherlands log is on this machine: the header excerpt is from the May 2026 log in oldlogs/, the values from the baseline block on are the September ones in doc/deck_data.json; the eras do not reconcile ("If you open a pre-July log").

Header and Loading block

Country: Netherlands (warm-start simplex)
Loading ExistingPharmacies/Netherlands ...
  N=114540 OD rows, M=1617 facilities
Loading NewPharmacies/Netherlands ...
  N=1001117 OD rows, M=29793 candidate facilities

N counts OD rows — (client, candidate) pairs in the client's choice set (GeoDMS OD matrix and choice set); M the location ids left after region exclusion and, on the New side, candidate subsampling (Scope rules and data gaps). The Existing side serves only the baseline; the New side is the LP's candidate set, hence an N about nine times larger. [ Info: lines may precede these: <country>: excluding k NUTS3 region(s) … and load_from(<country>): protected P baseline candidates; kept K of M (factor=F).

Baseline block

Baseline (ExistingPharmacies, each client → nearest pharmacy; unreachable priced at BIG):
  facilities used (cells):     1615 (of 1617)
  raw pharmacy count:          1992 (external reference)
  total travel cost(c):        6.0676723e7
  facility cost @ €100000 ea: 1.615e8
  mean travel time (min):      3.4754
  total client population:     17365050
  unreachable clients:         0 cells / 0 residents priced at BIG=120.0 (coverage-consistent)
  baseline arrows → C:\LocalData\networkmodel_eu\ExistingPharmacies\Netherlands\lambda_sweep\LINEAR\baseline

This is today's network as one point, the baseline ★ (Glossary); its count and travel are the S1 and S2 targets (Scenarios S1 S2 S3 explains the choices behind baseline_metrics).

Line What it is (baseline_metrics, around line 167) Read it as
facilities used (cells) n_used: existing pharmacy cells that are road-nearest for at least one client; (of M) is the Existing-side M. Today's count as the model sees it, the S1 target; a cell that is nobody's nearest ("empty", Metrics aggregation and ranking) is not counted: NL 1 615 of 1 617.
raw pharmacy count RAW_PHARMACY_COUNTS[country], hard-coded; only "Netherlands" => 1992 exists. Informational; feeds n-raw.
total travel cost(c) Σ pop·c(t_nearest), plus BIG·pop for every client in the New-side OD but absent from the Existing-side OD. The S2 target, in the function's units (units table). Nearest is by raw t, not c(t) — no difference while c rises with t.
facility cost @ €100000 ea target_cells × FACILITY_MIN_COSTS. Placeholder bill; unused downstream.
mean travel time (min) time_total / total_pop, unreachable clients at 120 min. Minutes per resident — not the recipe of mean_t in the rows (units table).
total client population covered_pop + stranded_pop; excluded regions weigh 0. Denominator of the mean.
unreachable clients n_stranded cells / stranded_pop residents charged BIG = big_cost() (120 LINEAR, 1.0 LOGISTIC). Stranded (Glossary): a client cell with no reachable pharmacy. Zero-weight cells are skipped on both counts (d834a7c).

Before 206923c (July 2026) unreachable clients were silently dropped and the Danish and ITG (Isole) baselines sat below their own frontiers; pricing them at BIG makes ★ coverage-consistent with the LP (Scenarios S1 S2 S3). A large resident count here is a data gap until proven otherwise (Scope rules and data gaps).

Build and grid header

Building LP (once) ...
  built in 4.2 s
Common λ sweep (LINEAR): fixed grid 1-2-5/decade, 1e-4 … 5.0, warm-start dual simplex, no early stop.

The LP is built once; only the y coefficients change per point (Code walkthrough lp_run). The text is fixed: it still says "1-2-5/decade" when SWEEP_MULTS asked for a denser grid (The lambda sweep and the Pareto frontier).

The sweep table — reading one real row

Each grid point yields the HiGHS output (next section), a timing line and a row — here Netherlands LINEAR at w = 0.2, September numbers from deck_data.json:

w           λ (€)         sum_y       travel_relax    travel_topp     travel_greedy   travel_multi    n_open    fac_€         mean_t    frac_y    n-cells   n-raw
  solved w=0.2 in 74.9 s
0.2         20000.0       1899.83     4.1587045e7     4.2340372e7     4.1596596e7     4.1590157e7     1900      1.89983e8     2.3822    57        285       -92

At w = 0.2 a location costs λ = €20 000. The LP relaxation (Glossary; y allowed to be fractional, The facility location model explained) wants 1 899.83 locations' worth of openness (sum_y) and says travel cannot go below 41.59 M person-minutes at this price (travel_relax). Three roundings turn the fractions into a real set of 1 900 cells (n_open): top-p 42.34 M, greedy 41.60 M, multistart 41.59 M — only +0.007 % above the bound, so the relaxation is essentially integral here, as the 57 fractional candidates (frac_y) also say. n-cells = 1 900 − 1 615 = 285: S1 lies at a higher price; travel_multi is below the 60.68 M baseline, so S2 does too.

Token Column Printed as Unit How to read it
0 w 5 significant digits dimensionless The price knob. Grid points print exactly; bisection points rounded (0.34087) — their Arrow folder carries the full float.
1 λ (€) w × FACILITY_MIN_COSTS, 2 decimals € (placeholder) Price per open location.
2 sum_y Σ_j y*_j, 2 decimals locations (fractional) The relaxation's count. The S1 bracket test and the snapped S1 row use this, not n_open, because sum_y falls monotonically with w (The lambda sweep and the Pareto frontier) while rounded travel is monotone only "up to rounding noise" (comment around lines 320–321).
3 travel_relax 0 decimals; ≥ 10⁶ as 4.1587045e7 travel-cost units The LP lower bound (Glossary): served travel plus every client's unserved fraction priced at BIG (solve_at_w!, around lines 503–513). The grey dashed curve in the charts.
4 travel_topp idem; - on an LP-only row travel-cost units Coverage-honest travel of the top-p rounding (the p cells with the largest y*): an unreachable client is charged BIG, not dropped. Diagnostic.
5 travel_greedy idem travel-cost units Same for the greedy (CELF) rounding. Diagnostic.
6 travel_multi idem travel-cost units Same for multistart (From LP fractions to real pharmacies) — the frontier the deck plots, a real network and hence an upper bound; equals cost_c under the default ROUNDING=multistart.
7 n_open integer cells p = clamp(round(sum_y), 1, M); nothing is opened on an LP-only row.
8 fac_€ sum_y × 100 000, 0 decimals € (placeholder) Bill at the fractional count; not the facility term of the objective (units table).
9 mean_t 4 decimals; - on an LP-only row minutes per resident The selected rounding's assignment over the total population; stranded clients count 0 min (units table).
10 frac_y integer count Candidates with 10⁻⁶ < y* < 1 − 10⁻⁶; small = nearly integral, grows at high λ and under LOGISTIC.
11 n-cells integer cells n_open − target_cells; S1 is where it crosses zero.
12 n-raw integer, Netherlands only pharmacies n_open − 1992.
12/13 (LP only) literal — Light row from a refine-only run; rounding columns are -.

Tokens 3 and 6 bracket the integer optimum; multi / relax − 1 is the certified gap (41 590 157 / 41 587 045 − 1 = +0.007 %) — the LP ≤ integer ≤ multistart sandwich of From LP fractions to real pharmacies. The parser (build_deck_data.py, around lines 85–105) requires at least 12 tokens and reads tokens 0, 2–7, 9 and 10; λ, fac_€, n-cells and n-raw are not stored. A light row fails on float("-") and is skipped on purpose.

The timing line solved w=0.2 in 74.9 s is the wall time of solve_at_w!, LP plus roundings; HiGHS's HiGHS run time just above is the LP alone. Its w is the full float, matching the w=<float> Arrow folder. A light row says (LP only, no rounding); a failed point prints LP at w=1.0 FAILED: LP at w=1.0 did not solve (termination_status=TIME_LIMIT). and no row.

Bracket and bisection lines

When two consecutive successful grid points enclose a target, the driver bisects before continuing (issue #52; concept on Scenarios S1 S2 S3, code on Code walkthrough lambda_sweep_simplex):

S1 bisection over w in (0.2, 0.3): pin sum_y ≈ 1615 to ±3.23 (≤8 solves)
  solved w=0.2449489742783178 in … s
0.24495     24494.9       …
  S1 pinned at w=…: sum_y=… (target 1615)
  basis restored at w=0.3 (OPTIMAL, … s)

(Bracket, tolerance and midpoint √(0.2·0.3) are the NL LINEAR case computed from the code.) Always read the (target …) residual: bisect! says "pinned" after at most REFINE_ITERS solves whether or not it got within tolerance (Denmark LOGISTIC in #52: 9.5 % off, after a post-hoc bisection line — a bisection across a failed grid point). basis restored re-solves the bracket's upper endpoint so that the next grid point warm-starts (Glossary) as if there had been no detour; expect OPTIMAL. Bisection rows print twice, here and in the Combined table.

After the loop: S1 (sum_y=1615) not bracketed at w_max=5.0 (raise SWEEP_WMAX); Fine λ sweep over w in [lo, hi] retries failed grid points inside the brackets; … not pinned in the coarse loop (bracket spans a failed point?): post-hoc bisection is the fallback; both scenarios are pinned and w=2.0 timed out: … ending the sweep is the tail stop (why on The lambda sweep and the Pareto frontier).

Combined table and scenario summary

Combined sweep, sorted by w: repeats the header and every row — grid, bisection, fine and light — sorted by w; it is the only table the parser reads. Then:

S1 — closest to sum_y = 1615 (baseline cells):
  w = 0.3   λ = 30000.0 €
  sum_y         : 1334.42      (target 1615)
  n_open        : 1334         frac_y: 63
  travel_relax  : 5.5459326e7     (LP1 with all fractional x)
  travel_topp   : 6.0654253e7   greedy: 5.5494283e7   multi: 5.5481849e7
  stranded      : topp 33   greedy 0   multi 0
  travel_c (multistart): 5.5481849e7     (selected; baseline 6.0676723e7)
  travel change (ph2): -8.56%
  facility €    : 1.33442e8     (baseline 1.615e8)
  mean t (min)  : 3.1778    (baseline 3.4754)

(Values: NL LINEAR scen.S1 in the checked-in deck_data.json, the pre-#52 snapped row at w = 0.3, 17 % below today's count; its S2 at w = 0.34087 is a post-hoc bisection point — before #52 only S2 was bisected, after the fine sweep.)

S1 is the rounded row whose sum_y is closest to today's cells; S2 (closest to travel_c (ph2) = … (baseline)) the one whose cost_c is closest to today's travel — closest() over all rounded rows, bisection rows included, ties to the lowest w. stranded counts client cells with no open candidate under each rounding. The parser reads the block from sum_y to mean t (min), which closes it; "ph2" and "LP1" are May relics.

Two S1s. This summary is the snapped S1/S2, a solved row; the deck's tables and chart markers use chord interpolation (Glossary) between neighbouring rows, and both are called "S1" (artefact matrix: Scenarios S1 S2 S3). The S1 arrows → … (w=…) and S2 arrows lines that follow export the snapped rows; the S1/S2 folders are overwritten by every run, refine-only ones included.

Reading the interleaved HiGHS output

The only HiGHS output on this machine is oldlogs/lambda_sweep_netherlands_simplex.log (May 2026, HiGHS 1.13.1): a hard-coverage run — every client had to be assigned — and, inferred from its numbers, a QUADRATIC one ("If you open a pre-July log"). Message texts are unchanged; sizes and iteration counts are indicative only. Solver terms: Background theory.

LP has 1024025 rows; 1030910 cols; 3003351 nonzeros (first solve only, with Presolving model) checks against the Loading block: columns = N + M = 1 001 117 + 29 793 (one x per OD row, one y per candidate); rows = N link rows (x ≤ y) + 22 908 client cells (one coverage row each). Coefficient ranges → Cost [3e-02, 1e+05] is the spread of the objective coefficients; a very wide one triggers WARNING: Problem has some excessively large costs (Italy: 3e+06), and under soft coverage the negative x-coefficients (c(t) − BIG)·pop widen it further — the recorded reason interior point fails on soft LOGISTIC (4ec47a9, The lambda sweep and the Pareto frontier).

Objective value is the LP objective as coded: travel + λ·Σy under hard coverage (w = 1e-4: 146 475 + 10 × 22 718 = 373 655 ✓); under soft coverage travel_relax + λ × sum_y − BIG × total client population (NL LINEAR w = 0.2: ≈ −2.004e9; verify on a current log).

Warm steps. A healthy one opens with Solving LP with useful basis so presolve not used and Ph1: 0(0) at iteration 0 — the previous basis was accepted and is still feasible (w = 2e-4: 1 369 iterations, 1.6 s, against 45 870 and 10 s cold). A cliff solve opens with large Ph1: …; Du: … infeasibilities and rebuilds feasibility first (w = 0.5: 357 108 iterations, 2 086 s; w = 1.0: 13 061 s; why on The lambda sweep and the Pareto frontier); the same pattern appears whenever a bisection or retry starts from an aborted basis. At LP_TIME_LIMIT (3 600 s) the driver prints its FAILED … termination_status=TIME_LIMIT line; HiGHS's own wording for a time-out, and for the OTHER_ERROR of interior point on soft LOGISTIC, is in no log here.

Units and identifiers — one table

Quantity On disk / in GeoDMS In Julia and in the log Trap
Travel time t_ij in _od.arrow: float32 seconds (MinTime = 1 s), capped at 120 min t_ij_col = od[:t_ij] ./ 60 → Float32 minutes; traveltime.arrow is in minutes Never feed raw t_ij to c(t).
Travel cost — LINEAR: c(t) = t, Σ pop·c(t) in person-minutes. LOGISTIC: c(t) ∈ [0.076, 1], flat at 1 (= BIG) by about 60 min; sum dimensionless. BIG = 120 / 1.0 (Travel cost functions). Not comparable: NL baseline 6.07e7 vs 1.91e6.
w, λ, € — w dimensionless; λ = w × FACILITY_MIN_COSTS (100 000) sits in the same objective as travel, so its true unit is "travel-cost units per location"; € assumes the placeholder scale (The lambda sweep and the Pareto frontier). Rankings and native-unit deltas do not move with the placeholder; € labels do. Five scripts in doc/ hard-code 100 000.
Facility cost facility cost @ €100000 ea = cells × 100 000 fac_€ = sum_y × 100 000 (bill at the fractional count); issue48_sweep_results.csv facilitycost_eur = λ × n_open (the price actually paid, the facility term of the objective). A factor w apart — NL LINEAR w = 0.2: 1.89983e8 vs 38 000 000; why both exist is not recorded.
Cells / pharmacies / residents Descriptives: n_pharmacies (raw points), n_pharmacy_cells (1 km² cells), n_residents Every model count — facilities used, sum_y, n_open, stranded, S2_facilities — is in cells (NL: 1 992 pharmacies in 1 617 cells); population is total_pop residents (CLIENT_WEIGHT). "Locations" in the deck = cells; n-raw (NL only) is the one column in pharmacies.
Ids client_rel, facility_rel in _od; id in _i/_j: 0-based, _j ids local to each file Clients match across sides by id (both exports use ModelClient); locations never — baseline protection matches on rounded (x, y) in metres. Join a NewPharmacies/…/assignment.arrow to the New-side _j.arrow, never the Existing one.
mean_t — Rows: t·pop over assigned clients divided by the total population — a stranded client counts 0 min (around line 551). Baseline: stranded at 120 min (around line 193). Not comparable where stranding matters (ITF S2 LINEAR strands 208 cells); an oversight rather than a decision (inference). cost_c and travel_relax are consistent.
open flag; coordinates assignment.arrow open 0/1, read by GeoDMS as OpenEnum 0/1/2; x, y float64 EPSG:3035 metres the driver never writes 2 The exported location files carry only x, y.

If you open a pre-July log

oldlogs/lambda_sweep_netherlands_simplex.log (late May 2026, the first warm-start driver, 011cafb) has a different row:

w           λ (€)         sum_x       travel_relax    travel_c_ph2    n_open    fac_€         mean_t    frac_x    n-cells   n-raw
0.2         20000.0       1667.48     3.5373693e7     3.4606704e7     1667      1.667475e8    2.6403    343       54        -325

Three things before quoting it. First, the log does not name its travel-cost function; the numbers say TRAVEL_FUNC = QUADRATIC (then and still the default, settings.jl line 44; c(t) = 0.05t² + 0.5t), so its travel columns are quadratic cost units, not person-minutes: at w = 1e-4 mean_t 0.0168 min × 17 365 050 residents ≈ 291 700 person-minutes under LINEAR, yet travel_relax is 146 475 ≈ the quadratic cost of a 1-second trip × population (144 950). Whether the QUADRATIC era (the early-June NL sweep, 439b22e) is why Analyses.dms SweepResults/baseline still reads the QUADRATIC folder is not recorded (GeoDMS OD matrix and choice set). Second, travel_c_ph2 = 3.46e7 sits below the relaxation's 3.54e7, which a real network can never do: this "phase-2" figure was a plain nearest-open assignment to the top-p set summed over the clients that had an open candidate, leaving out those the set could not reach while the relaxation charged them (e1c426b). Third, never compare its numbers with current ones: cost function, coverage (hard → soft), client universe (ModelClient on both sides since 1 September), roundings and baseline consistency all changed since May 2026.

Old (May–June 2026) Current Changed by
sum_x, frac_x sum_y, frac_y 4cee238 (x = assignment, y = facility); the parser accepts both
Coarse λ sweep (… stop when travel_c > baseline); no unreachable clients line Common λ sweep (…) no early stop.; with it 206923c
No bisection (May); from b76f643 (7 June) a post-hoc S2 bisection over w in (lo, hi) to resolve cost_c ≈ baseline=… block after the fine sweep (≤ 6 solves, 0.5 %, S1 not bisected); the fine sweep re-solved grid points inside the bracket range interleaved S1/S2 bisection blocks, basis restored, light rows; the fine sweep skips solved w b76f643 → 29fcf7f, e25416b (#52)

The Arrow files

Inputs (written by GeoDMS, read by load_from)

Three files per side under C:\LocalData\networkmodel_eu\{ExistingPharmacies,NewPharmacies}\ (Analyses.dms → DistanceTableExport, ClientExport, FacilityExport, around lines 344–449; directive by directive on GeoDMS OD matrix and choice set).

File One row per Columns Notes
<area>_od.arrow (client, candidate) pair in the choice set client_rel, facility_rel (0-based indices into _i, _j), t_ij float32 seconds, d_ij km, travelcost_ij, Deterrence_ij Julia uses the first three. Declared Client_rel in the DMS, client_rel on disk, which Julia reads.
<area>_i.arrow client cell (populated 1 km² cell) id, pop, total_pop, x, y, NUTS (NUTS3 code or "") pop = total_pop since Client := 'population'. Declared nuts, written NUTS; Julia reads :NUTS (else it skips region exclusion with a warning) — the Portugal exclusion ran, so NUTS is what the file carries.
<area>_j.arrow pharmacy or candidate cell id, x, y, NUTS Ids are per-file local.

Outputs (written by the driver)

<base>\<area>\lambda_sweep\<TRAVEL_FUNC>\<label>\assignment.arrow   (id, open)
<base>\<area>\lambda_sweep\<TRAVEL_FUNC>\<label>\traveltime.arrow   (id, t_ij)

<base> is ExistingPharmacies for label = baseline, NewPharmacies for label = w=<w> (every rounded point, full float, e.g. w=0.3872983346207417) and for the stable S1/S2 labels, which exist because their w differs per area and the .dms needs a fixed name (df4febf; sweep_dir, write_*_arrows, around lines 23–69). assignment.arrow: id (that side's _j id) and open (0/1) for every cell of the side; for the baseline open = 1 marks the cells that are somebody's nearest. traveltime.arrow: id (client, ascending) and t_ij (Float32 minutes to the assigned open cell); a stranded client has no row, so GeoDMS's rjoin gives it a null. Nothing cleans the accumulating w=… folders; ReadSweepResults_T reads them back (GeoDMS OD matrix and choice set).

Exported open-location files (issue #45): export_s1s2_locations.jl joins S1/assignment.arrow (open == 1) to the New-side _j.arrow and writes <area>_<s1|s2>_<linear|logistic>_open.arrow with x, y (EPSG:3035), one row per open cell (Scenarios S1 S2 S3). Until #52 these came from the snapped rows; the pinned rows are in the unpushed 12 September regeneration (Branches data and environment).

deck_data.json

One entry per area in ORDER (44 names; 42 present at e25416b — Hungary and Finland had no logs):

{"region": "Netherlands", "title": "NETHERLANDS", "name": "Netherlands",
 "func": {"LINEAR": {"baseline": {"cells": 1615, "cost": 60676723.0, "mean_t": 3.4754},
                     "rows": [{"w": 0.2, "sum_y": 1899.83, "relax": 41587045.0, "topp": 42340372.0,
                               "greedy": 41596596.0, "multi": 41590157.0, "n_open": 1900,
                               "mean_t": 2.3822, "frac": 57}, …],
                     "scen": {"S1": {"sum_y": 1334.42, "n_open": 1334, "frac": 63, "st": [33, 0, 0],
                                     "multi": 55481849.0, "mean_t": 3.1778, "w": 0.3, …}, "S2": {…}}},
          "LOGISTIC": {…}}}

baseline = the three baseline lines; rows = the Combined table (keys relax, topp, greedy, multi, frac); scen = the snapped summary, st = stranded cells [topp, greedy, multi], w/relax/topp/greedy attached from the row nearest in sum_y. Duplicate w survive: NL LINEAR has two w = 0.3 rows next to the pre-#52 S2 bisection points 0.3873 (= √(0.3·0.5)) and 0.34087. A newer logs/refine_<area>_<FUNC>.log with both summaries has its rows merged in at unsolved w and its scen and baseline replace the sweep's — safe only when both logs came from the same LP (issue #54; rule and trap on Deck and charts pipeline). DECK_ONLY="A,B" restricts the output for preview decks; never aggregate from a partial JSON. doc/agg_data.json (build_charts.py) is the aggregate frontier, S1/S2 snapped to the nearest union rows, mean_t 0 because minutes are not summable across areas (Metrics aggregation and ranking).

The CSVs in doc/

File Rows Columns Notes
frontier_metrics_interp3.csv (+ .json) 84 (42 × 2 functions) region, func, B_x, B_y, S1_y, S2_x, rect_area, rect_rel, cross_x, cross_y, lambda_cross, s1_bracketed, s2_bracketed B_* today's cells and travel; S1_y, S2_x chord-interpolated on the Pareto envelope (Glossary) of (n_open, multi); rect_*, cross_* and lambda_cross (placeholder €) are defined on Metrics aggregation and ranking. w is interpolated linearly here, in log w elsewhere (reason not recorded; aggregate S1 €12 968 vs €12 984). interp1/interp2 are stale July snapshots; frontier_charts.py defaults to interp1 unless METRIC_SUFFIX=interp3.
issue48_sweep_results.csv 1 530 region, func, w, lambda_eur, n_open, sum_y_lp, travelcost_multistart, travelcost_lp_relaxation, facilitycost_eur, totalcost_eur, mean_traveltime_min, frac_open, baseline_facilities, baseline_travelcost, on_pareto_frontier Every deck_data row plus an AGGREGATE pseudo-region. facilitycost_eur = λ × n_open; totalcost_eur adds it to travelcost_multistart — meaningful only at the placeholder scale.
issue48_crossings.csv 86 region, func, B_facilities, B_travelcost, S1_travelcost, S2_facilities, corner_facilities, corner_travelcost, rect_area_raw, rect_rel, cross_facilities, cross_travelcost, lambda_cross_eur, w_cross, cross_facilitycost_eur, cross_totalcost_eur, s1_bracketed, s2_bracketed, n_frontier_pts The frontier_metrics quantities under descriptive names plus corner and crossing costs; the JRC delivery for #48 (issue48_README.md).
pharmacy_descriptives.csv / _nuts1.csv 17 countries / 25 rows study_area, n_pharmacies, n_pharmacy_cells, n_residents, residents_per_pharmacy, n_cells_multi, avg_pharm_in_multi, max_pharm_in_cell, then cell_n_empty, cell_min_nz, cell_min, cell_p10 … cell_p90, cell_max, cell_avg and the same ten pharm_* From GeoDMS Pharmacies/Descriptives/Table via collect_descriptives.jl; catchment rule and "empty" count in Metrics aggregation and ranking. n_pharmacy_cells − cell_n_empty usually equals the log's facilities used (NL 1 617 − 2 = 1 615).
policy_typology.csv 44 region, iso, country, archetype (1–4), archetype_name, formalisation, color Regulation typology from the JRC deck; colour key of the ranked lists.
region_typology.csv 42 region, pop, rural_share, town_share, urban_share, class DEGURBA-style shares from _i.arrow; no script reads it at e25416b.

In the code

Every environment variable named on this page is documented with default and reason in Code walkthrough settings and Code walkthrough lambda_sweep_simplex; DECK_ONLY and METRIC_SUFFIX in Deck and charts pipeline.

What Where
Baseline block lambda_sweep_simplex.jl → baseline_metrics (around 167), pln block in analyze_country (around 263–271) — permalink
Header and row print_sweep_header (around 212), print_sweep_row (around 221) — permalink
Row values lp_run.jl → solve_at_w! (around 460–560): travel_relax 503–513, light return 516, roundings 531–541, mean_t 551 — permalink; travel_of (around 137) for stranded
Bisection and pinned lines bisect! (around 343), refine_if_closed! (around 364), grid loop (around 387)
Combined table, summary, S1/S2 folders around 494–548; closest at 504
Arrow writers, inputs sweep_dir, write_*_arrows (around 23–69); load_from (around 71)
Units and constants settings.jl: FACILITY_MIN_COSTS (17), TRAVEL_FUNC default (44), c(t) (around 75), big_cost() (96–97)
Parser doc/build_deck_data.py → parse (around 70–146), refine merge in main (around 169–183)
GeoDMS exports and readers cfg/main/Analyses.dms: DistanceTableExport (~344), ClientExport (~415), FacilityExport (~436), ReadSweepResults_T (~219), Descriptives (~28)
Location export export_s1s2_locations.jl

Pitfalls and open questions

  • Two S1s. Snapped (summary block, scen, exported locations) versus chord-interpolated (deck tables, chart markers, frontier_metrics_*). Say which (Scenarios S1 S2 S3).
  • "Pinned" is not "within tolerance". Read the (target …) residual, especially after post-hoc bisection.
  • fac_€ is not facilitycost_eur — a factor w apart (units table).
  • mean_t asymmetry (baseline 120 min for stranded, rows 0 min); no issue filed.
  • The parser is positional and byte-exact. Fewer than 12 tokens and a row vanishes silently; the S1/S2 blocks are found by ^S1 — (em dash) in UTF-8, so a log written by PowerShell's own > (UTF-16) yields zero rows — redirect through cmd /c (Running a sweep end to end).
  • Stranded counts include zero-weight cells (travel_of counts every client cell with no open candidate, whatever its weight; frequency unverified).
  • Old logs are not comparable; the May log's QUADRATIC cost function is inferred, not stated ("If you open a pre-July log").
  • Nothing pharmacy-related is on this machine — logs and Arrow files live on Maarten's machine; the checked-in deck_data.json is the 3 September regeneration (de5b178), pre-#52, ESPON-Italy (Branches data and environment).

See also

Clone this wiki locally