Skip to content

Decision log

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

Decision log

In one sentence: every methodological decision that shaped the pharmacy model — what, when, by whom, why, and what number it moved — so you can tell a deliberate choice from an accident and know which results are comparable with which.

Where this sits: this page is the history behind the whole pipeline (GeoDMS → Arrow → Julia LP → λ-sweep → rounding → scenarios → deck). The neighbouring pages explain each stage as it is today; this one explains how it got there. Rejected ideas: Alternatives not pursued; branches and data: Branches data and environment.

How to read this log

Five eras tell the story; the table at the end lists every decision with its evidence: a commit on the ServiceAccess branch (dates are commit dates), a GitHub issue (#n) or a document in doc/ ("todo #n" and "todo A1" are items of doc/todo.md, the July recalculation checklist). Where no reason is written anywhere I say "not recorded"; where I infer one, I say so. "Moved" is the number the change shifted, when a source records one.

People, as the sources name them: Lewis and Bernhard Nöbauer (BN) (CFE/EDS, as the review doc labels BN) set the scenario framing; Ana Moreno Monroy (CFE/EDS) and Martijn Brons (DG REGIO) reviewed the methodology draft; Chris Jacobs-Crisioni (JRC) owns the data and the network configuration; Maarten Hilferink the Julia optimisation and the deck, Lola Dekhuijzen the school-era solvers, Jip Claassens the road network and OD matrix (all Object Vision).

Caveat, as of 15 September 2026: the branch on GitHub ends at e25416b (10 Sep). Issues #52, #53 and #44 cite later commits (216ac77, and in #52 four smaller ones) that exist only on Maarten's machine; #54 describes the same unpushed state. Everything attributed to them comes from the issue text (Branches data and environment).

Era 1 — the network model and the school problem (2023 – spring 2026)

The repository began in August 2023 as Jip's road-network model: a cleaned TomTom network per country, its largest strongly-connected component kept (Glossary) — a rule that later turned out to discard islands (Era 4). The optimisation half arrived in February 2026 for primary schools; two of its decisions still shape the pharmacy model.

The choice set. Maarten limited each client to its 5 nearest facilities (eaf133c); issue #22 gives the reason: searching from pupil locations to five schools has a stop criterion, the reverse direction has none — a compute choice, not a behavioural one. Today this is a radius: every candidate cell within the road time to a client's 5th-nearest existing pharmacy (71–86 candidates on average, 4267fe0), capped at Jip's 120-minute OD cutoff (535f796, no recorded rationale), which later became BIG, the price of an unserved client, by inheritance. Both sit in cfg/main/ModelParameters.dms (around lines 14–16). Because the radius is fixed by today's pharmacies it can bind at S2 (Era 5). Configuration: GeoDMS OD matrix and choice set; where it binds: Scope rules and data gaps.

The facility cost is linear. Chris fitted a power-law cost per pupil to 2023 school data (issue #38); the meeting judged it "may close schools too aggressively", and his three-piece linearisation fails because the model multiplies cost per pupil by load. Maarten's closing line — "We have continued with linear facility location costs: C_j := y_j·(c₀ + c_x·x_j)", a fixed cost c₀ per open school plus c_x per pupil served (x_j is the school's load here, not the later assignment variable) — is why the pharmacy model has a fixed price per location. Under the λ-sweep the per-client part drops out because (nearly) every resident is served in every solution (The facility location model explained).

The school objective — hard coverage plus a penalty proportional to under-enrolment (41f55bb) — was not carried over. No commit says why; the plain reason is that nobody has a minimum viable pharmacy size (inference). It survives in lp_run.jl → run_scenario and on Service allocation procedure.

Era 2 — pharmacies and the first sweeps (May – June 2026)

On 24 May Maarten added the pharmacy analyses, LOGISTIC at midpoint 30 / scale 15, the placeholders FACILITY_MIN_COSTS = 100000 and FACILITY_CLIENT_COSTS = 3333 (the rounded school numbers, never calibrated), and the first sweep. Why pharmacies replaced schools is not in the repository; doc/topics.md names only the mail thread of 13 May. The all-country pharmacy point files arrived on 12 June (0b36dcb, "shared by Taina Pacheco" — the commit does not say OECD); whether the data drove the switch is inference. The thread fixed the question the model still answers — S1, S2, S3 (issue #43, closed in July: "now implemented as S1 and S2, but with lambda_sweep"): a price λ per open location (in code the multiplier w, λ = w × €100 000) traces the frontier of best (locations, travel) pairs, on which S1 and S2 are read against ★, today's network (The lambda sweep and the Pareto frontier, Scenarios S1 S2 S3).

Three decisions in the first fortnight made the sweep usable, each answering a measurement:

  1. Bracket S1 on the LP's fractional count sum_y (9ee7ac6): bracketing needs a quantity that moves one way as λ rises, and the rounded count did not — under interior-point (IPM) solves many tiny y-fractions were counted as open, so it jumped from 2 296 to 3 894 between w = 0.01 and 0.1 on the Netherlands (Scenarios S1 S2 S3).
  2. Warm-started dual simplex, sequentially (011cafb): each re-solve takes seconds where a fresh interior-point solve of the old Netherlands sweep took roughly 700–4 000 s; "sequential is the point". A cold IPM solve still beats a cold simplex solve (a9ef6cc); the warm start is what wins (The lambda sweep and the Pareto frontier).
  3. Multistart rounding with coverage-honest scoring (e1c426b): once a stranded client was charged instead of dropped, top-p rounding was seen to abandon "hundreds-to-thousands of rural" cells; the multistart is provably no worse than its seeds (From LP fractions to real pharmacies).

June's detour was the cap ladder, Lewis's and BN's four-rung recipe: closures decided by minimum and maximum catchments (residents per pharmacy) in variants A–C, a price per location (the λ-sweep) only as fallback D. Maarten built the cap engine in a day, ran it on the Netherlands and de-emphasised it the next morning (6dc70f5, "per feedback" — whose is not recorded): catchments spread so widely that no defensible cap exists. Formally dropped on 4 September; rungs and NL numbers on Alternatives not pursued.

Era 3 — the June–July recalculation

By late June the pilots had exposed comparability defects; the 24-June thread plus Chris's methodology draft (28 June) fixed the inputs, and doc/todo.md is the checklist.

  • Candidates = cells ≥ 50 inhabitants ∪ pharmacy cells (7c6768f; MinPopSizeForPharmacyLocation := 50). Chris: keeps ~98 % of pharmacy locations, search space 7.5 % → 2.9 % of EU cells; Lewis: "the best option". Luxembourg's candidates fell 1 599 → 945 while its LINEAR S1 improved slightly (−7.98 % vs −7.66 % travel), and the 2.5× smaller LP made country-level Poland tractable (29 h). The same commit fixed a wiring bug: the 5-nearest limit had been counting school cells (GeoDMS OD matrix and choice set).
  • Clients = the whole population (033dd32; Client := 'population'), per methodology §2.1; both network sides had to be rebuilt because client cells are network nodes.
  • Logistic 25/10 (7c6768f): 30/15 was a first guess; Lewis wanted indifference below ~5–10 min and saturation by ~45. Old runs stay reproducible via LOGISTIC_MIDPOINT / LOGISTIC_SCALE (Travel cost functions).
  • Coverage-consistent baseline (206923c). baseline_metrics had dropped clients who cannot reach any existing pharmacy (Denmark ~11 % of residents, ITG ~15 %, SE2 ~5 %) while the LP had to serve them; now they are priced like scenario stranding, and Denmark's baseline travel roughly doubled.
  • Aggregated frontier (d0ae5ef): sums facilities and travel at each common λ over the 41 disjoint areas (country-Poland excluded because its seven NUTS-1 are in), exact because the objective separates by area (Metrics aggregation and ranking).
  • Common grid, no early stop (206923c): the sweep used to stop once travel exceeded today's, so the aggregate's range ended where the first area stopped (Belgium, w = 0.1). One fixed grid for every area proved intractable under hard coverage (NL 16.7 h at w = 5.0, Austria LOGISTIC 40 h of time-outs) and triggered three solver reversals in three days (The lambda sweep and the Pareto frontier).
  • Soft coverage (4ec47a9, agreed 10 July). Hard coverage was the root cause of the stalls and of the unbracketable S1: the LP could not open fewer facilities than needed to reach everybody, so for ITG, Norway, Portugal, PL8 and the Netherlands the frontier stopped at a floor above today's count. "At most one" pharmacy per client, with BIG for an unassigned one, lets the frontier run down past ★ so every area brackets S1 and S2; the warm start stays valid between grid points and high-λ solves became cheap (few facilities open, small basis), so simplex serves both functions up to w = 5.0 LINEAR / 0.5 LOGISTIC. Formulation and the "constant drops out" trick: The facility location model explained.
  • BIG fixed at c(120 min) (settings.jl → big_cost(), env BIG_TRAVELTIME_MIN): each region's own t_max, 42 min (Paris) to 120 (Norway), had priced stranding inconsistently and broken the aggregate (Travel cost functions).
  • Improvement metrics (e75c73c, 2a6ea08): the raw improvement rectangle "mostly ranks by region size", so rect_rel — fractional facility saving × fractional travel saving — replaced it; charts coloured by the pharmacy-regulation typology (559dad9) (Metrics aggregation and ranking).
  • PROTECT_BASELINE (f167da2): the candidate subsample used for slow areas had kept only 555 of FRI's 1 590 baseline cells, so S2 came out with more facilities than today. The subsample is also the seed of #54 (Scope rules and data gaps).
  • x/y renamed (4cee238, 24 July): nothing numerical changed, but every older artefact uses the opposite letters (The facility location model explained).

Era 4 — August–September: scope, review, pinning, data

Landbody networks. Chris's 79cb58b (31 Aug), prompted by issue #13 (every centre on Gran Canaria at the 120-min cap), replaced "largest SCC per study area" by "largest SCC per landbody" (Templates.dms around line 93); no data changed, the island links had simply been pruned. Maarten's split of ITG's baseline cost (a058fe3, #49) showed 84.6 % of it was BIG penalty on 983 169 unreachable residents; Portugal (62.6 %) and PL8 (44.4 %) were the others, so the ranked lists had been ranking coverage artefacts. After rebuilding all 42 areas only those three moved (ITG 22.30 → 4.878 min, PL8 unreachable → 0). The commit had also put baseline and LP on different client units, resolved by one ModelClient (f0acbc9). Numbers: Scope rules and data gaps; configuration: GeoDMS OD matrix and choice set.

The node fix (#50). Norway, SE2 and SE3 then failed at the alloc step, reading back the link set the network build had written. Maarten refuted his own node-id hypothesis by measurement (exactly one link had an undefined node) and found the cause: a junction-free section (a chain of links without junctions, here an island ring road) with no connector link to any OD point, so min/max over an empty group left First_ConnectorLink_rel undefined. Fix (7086928): count connector links explicitly and drop such sections (IsDisconnectedJFS) and any link with an undefined node unconditionally; Chris's component rule is untouched. Byte-identical link sets on Luxembourg and Portugal proved it a no-op elsewhere (GeoDMS OD matrix and choice set).

Region exclusion (#49 → #51). Portugal did not move: all 1 893 pharmacies lie on the mainland, every unreachable client on the Azores and Madeira — a data gap, not connectivity. Maarten proposed a landbody filter (Country_Split) and left it unapplied as a modelling decision; Chris endorsed it ("poor data coverage in these islands is a recurring theme"). Implemented as a NUTS rule (872b388; settings.jl → excluded_nuts_regions, NUTS_EXCLUDE_SHARE = 0.5): a region with more than half its inhabited cells absent from the existing OD is dropped from baseline and sweep. Only PT200 and PT300 trigger, both at 100 %; Portugal's baseline travel fell 62 % (Scope rules and data gaps).

The REGIO review (doc/deck_changes_from_review.md: 26 comments, Ana in July, Brons 6 Aug, BN 30 Aug), worked in on 4 September:

  • Trilemma dropped, "by how much" leads (comments 1–3, 5–6): all three rejected an efficiency–accessibility–demand "two out of three" framing — cost and accessibility trade off continuously under exogenous demand, and the draft's yes/no research questions ("can accessibility improve? can the count fall?") are trivially yes. BN's reading was adopted: the frontier answers "by how much" without any λ (b3c634b). Native units first: the S1/S2 tables lead with Δ person-minutes (or Δ logistic cost) and Δ locations; λ columns in grey.
  • €100 000 is a placeholder (20–22): Ana wanted no values, Brons said values drive the outcome, BN resolved it: λ weights only the fixed part of the cost, the sweep runs over the multiplier w, and €100 000 only converts w into a € label — doubling it would halve every printed λ and leave every solution, S1/S2 point and ranking where it is (The lambda sweep and the Pareto frontier).
  • Descriptives first (4), a tangency diagram (41–42), logistic named (26–28: a plain logistic, not log-logistic; c(0) ≈ 0.076, not zero at zero minutes), bounds on total cost only (37–38: the LP-lower / multistart-upper sandwich certifies total cost, not travel and count separately; From LP fractions to real pharmacies); age-weighted demand and candidate neighbours listed as "Not pursued" (ef05320). BN's proposal to rescale the logistic so that c(0) and c(60) coincide with the linear curve — which would also ease the λ-compression Brons noticed — is tabled: it changes the equity weighting the logistic was chosen for (Travel cost functions).

Cap ladder dropped formally the same day with the urban/non-urban split, pharmacist-based caps, flat-then-linear cost and logit sensitivity — confirmed with Maarten; each on Alternatives not pursued.

S1/S2 pinned (#52). The exported S1/S2 location files snapped to the nearest swept grid point; in FRC, FRH, FRJ, FRK and SE2 both scenarios were the same point (SE2's S1 had 309 open cells against 452, −31.6 %), and over 74 sweeps S1 was off by more than 10 % in 22. June's post-hoc S2 bisection had failed because it ran after the sweep had climbed through three 1-hour time-outs, and a warm start from that far basis timed out too (FRC: 75 s as a grid step). Now (29fcf7f, e25416b) S1 (on sum_y) and S2 (on multistart travel, 0.2 %) are bisected the moment their bracket closes, the sweep ends at the first time-out above it, and a refine-only mode re-pinned the 37 areas whose OD had not changed, walking the grid LP-only (a rounding costs 50–100 s per point, an LP 7 s). Result (216ac77): S1 median error 0.13 %, S2 max 0.74 %; only country-level Poland LINEAR keeps its grid point (S1 +1.6 %; solves near S1 take 3–15 h). Concepts: Scenarios S1 S2 S3; code: Code walkthrough lambda_sweep_simplex.

Italy, Hungary, Finland (#53, #44). On 7 September the OECD delivered the Italian Ministry of Health list (20 635 pharmacies, OECD-geolocated) replacing ESPON's 12 991, and Hungary (3 094; typology Type 1 / HIGH from slide 13 of the typology deck). Wired in on 10 September, it overturned the headline: ITF went from 1 951 to 4 955 pharmacies, ITG's mean travel from 4.9 to 2.0 min, and the Italian rect_rel collapsed — ITF 0.268 → 0.071, ITG 0.261 → 0.073 (ITG had been 0.384 in July, before the landbody fix); SE1, Norway, Lithuania and SE2 lead now. Finland's June failure was a GeoDMS 20.0 limitation; 20.19 reads the GeoParquet, so 2aeb6c0 reinstated it (797 pharmacies, S1 −13.5 % travel, S2 −24.7 % locations): 44 study areas as of 12 September. Why Italy went from the June parquet (20 642 pharmacies over the five NUTS-1 areas per 3a2688f; 20 637 in the country row) to ESPON in July is not recorded. The pushed deck, README and Pharmacy results September 2026 still carry the ESPON numbers; per-country provenance on Branches data and environment.

Era 5 — open today (September 2026)

Issue #54: a non-convex lower bound in FRI. The FRI LINEAR LP curve is not convex — impossible for one LP, whose sweep can only return points on the curve's lower convex envelope (The lambda sweep and the Pareto frontier). doc/build_deck_data.py merges FRI's July full sweep (6 765 of 17 141 candidates under LOCATION_SELECTION_FACTOR=3, IPM, pre-landbody OD) with the September refine points (17 133 candidates, dual simplex): two LPs on one curve, λ not even monotone along it. Per #54, FRI was the only subsampled area⁴; it is being fully re-swept. The check exposed 16 more areas (listed on Branches data and environment) whose July OD differs marginally from the 2 September rebuild (kinks ≤ 1.5 %); a clean frontier means ~32 runs, ~20 h — awaiting Chris's call.

⁴ doc/issue48_README.md says "FRI and PL8" (README says only "the largest areas"); #54 says FRI alone. PL8's 2–3 September re-sweep set no factor, so it is most likely full-set (unverified; the header of logs/sweep_PL8_LINEAR.log would settle it).

The roadmap, in priority order (ef05320, README "Status and roadmap", confirmed with Maarten 4 Sep), with the reasoning:

  1. Communicate λ intuitively (person-minutes per location) — running; every reviewer found λ opaque.
  2. Widen the candidate radius 5 → 10 nearest existing pharmacies, one area first. It binds at S2: 10.2 % of SE2's residents have exactly one open facility left in radius (4267fe0); in the September results Norway is at 18.6 % and ITF on the OECD list 1.8 % (#53). Where the radius, not geography, decides who can be served, the frontier is under-estimated (Scope rules and data gaps).
  3. RSSV candidate reduction (random-sampling spatial voting: solve many small random sub-problems, keep the candidates chosen often) to replace the stride subsample — after 2, because sub-problems on candidate subsets strand clients whose radius holds no sampled candidate, corrupting the votes.
  4. Exact soft-coverage p-median MIP for one area, to measure the multistart's distance from the true integer optimum (the integrality gap; Glossary); 5. territorial coverage constraints (e.g. a minimum number of pharmacies per NUTS region) as an equity lever; 6. DECIDE BN's logistic rescaling; 7. calibrate the real fixed cost; 8. capacity caps; 9. counterfactuals (−10 % facilities; replace X %; let the model choose which to close); 10. the #54 re-sweeps; 11. country-level Poland LINEAR S1 (needs the IPM path or the MIP). Terms on Background theory.

The log at a glance

Every term in this table is defined at first use in the eras above and on Glossary.

Date Decision Why Evidence Moved
2023-08 → 2024-05 TomTom network cleaned per country; largest SCC kept (Jip, Chris) Reproduce TomTom connectivity 961d2c6, c48740b, #10 The OD side
2026-02-23 Client reaches its 5 nearest facilities (Maarten) Stop criterion that way only eaf133c; #22 Choice-set radius today
2026-03 Piecewise proposed (Chris), continuous asked (Maarten), quadratic "for now" No jumps; OECD input awaited #27; 26e39c7 QUADRATIC still the default¹
2026-04-01 OD cutoff 300 → 120 min (Jip) Not recorded 535f796 (#15) Becomes BIG in July
2026-04-29 Power-law school cost rejected; linear kept Closes schools too aggressively #38; 658c886 Fixed price per location
2026-05-13 → Jun Mail thread: S1/S2/S3, A→D ladder, logistic, six indicators JRC framing doc/topics.md; #43 June agenda
2026-05-24 Switch to pharmacies; LOGISTIC 30/15; €100 000 + €3 333 placeholders; first sweep (Maarten) Not recorded; IPM beats simplex for a fresh solve 94f88db, 46f23e6, 40a3d55, a9ef6cc Pharmacy era
2026-05-25 Bracket S1 on sum_y, not the rounded count Rounded count not monotone 9ee7ac6 S1 rule ever since
2026-05-27 Warm-started dual simplex, 1-2-5 per decade, sequential Seconds instead of ~700–4 000 s 011cafb Production driver
2026-05-28 NUTS-1 study areas for FR, IT, SE, PL Countries too big as one LP 44eaf5c, babedf4 42 areas
2026-06-07 Multistart rounding, stranded clients charged; S2 bisection Top-p strands rural cells; S1 = S2 on grid e1c426b, b76f643 Gap to LP ~0–5 %
2026-06-12 All-country lists (Taina Pacheco); Finland removed (Chris) GeoParquet broke the union on 20.0 0b36dcb, c11fd38; #44 Finland out to 10 Sep
2026-06-22/23 Cap ladder built on NL, de-emphasised next morning No settable caps 7e7c031…6dc70f5 Alternatives not pursued
2026-06-24 Interpolated S1/S2 tables; candidates = cells ≥ 50 ∪ pharmacy cells (Chris, Lewis) "—" if unbracketed; LP 2.5× smaller 4013c91; todo A1, 7c6768f LU 1 599 → 945 candidates
2026-07-02 Clients = population; logistic 25/10; wiring bug fixed Methodology §2.1; Lewis's kinks 033dd32, 7c6768f LU OD 5 370 → 7 995 rows
2026-07-05 42 areas recalculated; aggregated frontier; Poland swept; Iceland dropped from the descriptives loop Separability; Iceland's descriptives hang d0ae5ef; todo #7, #10 PL S1 −18.8 % travel
2026-07-06 Coverage-consistent baseline; common grid, no early stop ★ priced like the frontier; aggregate needs every λ 206923c; todo #3, #4 DK baseline ≈ ×2
2026-07-07/09 Solver churn: IPM above 0.5 → always IPM → per function Hard-coverage stalls 86476bf, f93aea1, acd59a6 Superseded 16 Jul
2026-07-10 Italy from ESPON shapefile (12 991) Not recorded 6a208f0; #46 Superseded by #53
2026-07-16 Soft coverage; BIG = c(120 min) (decided 12 Jul²); simplex for both; w_max 5.0 / 0.5 Full-coverage floor above today's count 4ec47a9 Every area brackets S1, S2
2026-07-16/20 Rectangle, crossing, rect_rel; typology colouring Raw area ranks by size e75c73c, 2a6ea08, 559dad9 Ranked lists
2026-07-21 PROTECT_BASELINE in subsampling Subsample dropped baseline cells f167da2 FRI S2 1 772 → 1 288
2026-07-24 Notation x = assignment, y = facility Match ReVelle–Swain 4cee238 Notation only³
2026-08-31 Largest SCC per landbody (Chris); ModelClient (1 Sep) Islands lost their network (#13); units diverged 79cb58b, f0acbc9; #49 ITG 22.30 → 4.88 min
2026-09-02 #50: drop connector-less junction-free sections One link with an undefined node 7086928 NO/SE2/SE3 alloc
2026-09-02 NUTS region exclusion > 50 % unreachable (Maarten, Chris) Azores/Madeira have no records 872b388; #49, #51 PT baseline −62 %
2026-09-04 REGIO review in; ladder and five topics dropped 26 comments; confirmed with Maarten 4 Sep b3c634b, f46bbf8, 7f99138, ef05320 Deck framing
2026-09-10 S1/S2 bisected inside the sweep; refine-only; LP-only walk Snap off by up to 31 % 29fcf7f, e25416b; #52 S1 median error 5.9 % → 0.13 %
2026-09-10 OECD Italy + Hungary; Finland back Quality-checked lists; 20.19 reads GeoParquet 29fcf7f, 8963d27, 2aeb6c0; #53, #44 44 areas; Italian ranking collapses
2026-09 #54 open: FRI bound non-convex; 16 areas flagged Two LPs on one curve #54 Awaiting Chris

¹ settings.jl line 44 defaults TRAVEL_FUNC to QUADRATIC; every production batch sets LINEAR or LOGISTIC. ² The comment at settings.jl lines 91–95 says 12 July; git log -S BIG_TRAVELTIME_MIN first hits 4ec47a9 of 16 July: dated 12 July in the comment, committed together with soft coverage. ³ Old logs, archive/*.jl and Service allocation procedure use x = open.

In the code

Open, in this order: doc/todo.md, doc/topics.md, doc/deck_changes_from_review.md, doc/replan_connectivity_fix.md, doc/region_exclusion_for_deck.md, doc/issue*_*.md; then settings.jl, where every knob carries its decision as a comment (big_cost, excluded_nuts_regions); lp_run.jl (build_lp_warmstart for soft coverage, resolve_for_basis! for #52); lambda_sweep_simplex.jl (grid comment around line 293, refinement block around line 317); cfg/main/ModelParameters.dms lines 14–41. Walkthroughs: Code walkthrough settings, Code walkthrough lp_run, Code walkthrough lambda_sweep_simplex, GeoDMS OD matrix and choice set.

Pitfalls and open questions

  • Do not mix numbers across eras. Three regenerations changed every result: 2 July (candidates, clients, logistic), 3 September (landbody, exclusion) and 12 September (OECD Italy, Hungary, Finland, pinned S1/S2 — unpushed). Deck page numbers in commit messages drift with every insertion.
  • Never recorded: why pharmacies replaced schools; why 120 min; why 1-2-5 per decade and the multistart constants; why the first logistic was 30/15; why NUTS_EXCLUDE_SHARE is 0.5; why the exclusion was built per NUTS region rather than per landbody as Maarten proposed and Chris endorsed (likely because NUTS is the unit the descriptives and deck already use — inference); why Italy moved to ESPON; whose feedback shelved the ladder; whether StudyArea_default = 'ITG' (flipped by 79cb58b, still 'ITG') was kept on purpose.
  • Open: the choice-set radius, BN's rescaling, the #54 re-sweeps, and whether the wiki should track e25416b or the unpushed 216ac77.
  • The mail thread, the methodology draft and the review .docx are not in the repository; their content is second-hand through doc/topics.md, doc/todo.md and doc/deck_changes_from_review.md. The Iceland hang is on Running a sweep end to end.

See also

Start here

Understanding the method

Working with the code

Reference

External

Clone this wiki locally