OverviewOrientation

What drives desertion in a volunteer army: the question, the design, the evidence, how it is checked, and where it stands. · source: reference/OVERVIEW.md

Every number on this page is computed when the page is built, from data/final as of 2026-10-29 19:01; none is typed. To correct one, fix its source and rebuild.

At a glance ⇱ Interactive ↕ Hover for values

builtrunningprovisionalplannedHover a box for what it is; click to open its sheet. A dashed arrow is not yet joined.

RussiaItalyFranceAll

DesertionsPer 1,000 men in serviceEnrolmentsMen in service

Month of desertion or enrolment, where the record dates it to the month or the day. Men in service is the spell sample the hazard model will use — enrolment and exit month both known, 73.7 % of all records — and the rate is its deserters over the men at risk that month; months with fewer than 200 at risk are not drawn.

MaticeAmateur theatre 1880–1913LegionnairesDesertion

per inhabitantper Czech speaker

By 1910 political okres; the switch sets the denominator of the Matice, theatre and legionnaire maps — all 1910 inhabitants, or the Czech speakers among them — while desertion is per legionnaire either way. Provisional: built in data/interim/intensity/ and not yet joined to the persons (D-163). Grey is an okres without a value. The court and settlement views and the correlations are on Descriptives.

Podrobný seznam míst, IX. Čechy, printed page 256. The whole page, both editions and the symbol plate are on Digitising the 1910 register.

Another manAnother deserter

One of 300 complete records drawn at a fixed seed, half of them deserters. No name and no place below the district; dates to the month. The rank change is on the common scale of the rank ladders.

The design: one man, two desertions

Every legionnaire had already deserted once, from the Austro-Hungarian army, before he joined the Legions. A later desertion from the Legions (exit code D) shows the same man making two loyalty decisions in sequence, under different incentives: the first against a multinational empire that had conscripted him, the second against a national army he had volunteered for. Costa & Kahn (2003), the template for this study, observe one decision per Civil War soldier. Here the first decision is on record for every man, so the second can be read as revealed preference against it.

The record describes both sides of that choice. For 100,531 men it names the Habsburg unit he left, and for 91,561 it gives his rank in both armies, so the change in standing between the two decisions — the rank reset on entering the Legion — is measured man by man.

The headline fact

Desertion from the Legions was concentrated in Russia. Of the 87,734 legionnaires who served there, 4.1 % deserted, against 0.23 % in Italy and 0.05 % in France; Russia accounts for 97.0 % of all deserters in the record. The timing is as sharp: 86.1 % of the Russian desertions dated to the month or day fall after the Bolshevik coup of 7 November 1917. Whether that is the coup itself, the Legion's position in Russia after it, or the men who served there is the question the hazard model has to separate.

The data

• 132,993 legionnaire service records, scraped from the Legion register (csol.cz): date and place of birth, schooling, occupation, religion, political affiliation, the Habsburg unit and rank at capture, the Legion unit and rank at enrolment, and the date and manner of leaving.
• 186,192 fallen Habsburg soldiers from the Military Historical Archive (VÚA), the comparison population. 1,340 individuals appear in both files, each pair confirmed by an exact date of birth (1,201) or by an archive note naming the Legion (139); weaker coincidences of name and year are dropped, not down-weighted.
• Birthplaces in 1910 districts. 84.6 % of legionnaires are placed in a 1910 district — 158 political okresy of the Czech lands or 192 districts and towns of Upper Hungary — through settlement crosswalks of 18,916 Czech and 3,511 Slovak places, so each man carries his home district's 1910 census.
• Occupations in HISCO. The Czech job titles are coded to the historical international classification and its social classes (HISCLASS) through a Czech dictionary written for this project; 98.5 % of recorded occupations are coded.
• The 1910 place register. The census's printed register of every settlement gives its population, language and religion, and in symbols under each table the institutions it had. It is being read by machine against a hand-made gold transcription. The certified reading is running; its tables enter data/final when it finishes, and nothing downstream uses the register before then.
• Czech nationalism before the war, place by place. Two measures of the treatment: amateur theatre, from the national database of Czech companies (7,368 places, 1780–1913), and the member lists of the two Czech national book societies, the Matice česká and the Matice moravská (11,128 entries read from the scans). Both are built and mapped; neither is yet joined to the persons.

How it is checked

• Against the reference work. Fidler (2025), the Encyklopedie československých legií, gives every Legion regiment's theatre. The theatre assigned here agrees on 99.6 % of the 71,854 Czechoslovak-Legion records it can check. It also caught two defects, both fixed: generals read as majors, and remarks about a rank scored as that rank.
• Against other coders. OccCANINE, a neural coder trained on historical occupations and given the same German glosses, agrees on the social class of 89.1 % of coded records and on the HISCO major group of 96.8 %. The Czech statistical office's CZ-ISCO, which pairs a Czech and an English label under one code, agrees with the Czech reading on 99.1 % of the records it can compare.
• Against the census. The commonest occupation, rolník, is coded as an own-account farmer rather than a hired hand. Across 100 okresy, the legionnaires' own-account share within agriculture moves with the census share of Selbständige (Spearman +0.40).
• Against an outside transcription. Ostafin et al. (2020) typed every Silesian commune's 1910 census from the same printed volume. The census used here is identical in 12 of 12 political districts, and the machine reading of the register equals their population figure on 98.6 % of 441 communes.
• On every build. 126 of 127 invariants pass on the current data. Each corresponds to a bug that was once real, and notebook 02 runs them as its last cell, so a build that breaks the data fails.
• In writing. 188 judgement calls in a dated decision log, each with its status and a link to the argument, and 3 full code-and-logic audits with their findings and fixes.

Where it stands

Data construction is done and checked; the analysis has not begun. The next step is the person-period spell file: one row per legionnaire per period from enrolment to exit, the input the estimation needs. On it comes a discrete-time competing-risks hazard model following Costa & Kahn (2003), in which desertion competes with death, discharge and demobilisation. A static logit would condition on survival it does not model, and with a single spell per soldier individual fixed effects are not available. The first regressor of interest is the rank reset. Its main threat is that enrolment officers demoted men who had already shown signs of disloyalty; late-enrolment restrictions and balance tests are planned against it.

Two pieces remain open. The Habsburg order of battle of 1914, which ties a soldier's regiment to its recruiting district and its ethnic make-up and so carries the Costa–Kahn mechanism back to the first desertion, is built but not yet joined to the persons. The education ladder is checked against the Austrian school statistics, which cannot reach the Hungarian half of the monarchy. The Roadmap lists every open item.

Reading further

Three routes through the sheets, for the three questions readers arrive with. Click any box.

Every sheet

Running it yourself

python -m venv .venv && . .venv/Scripts/activate   # Windows
pip install -r requirements.txt

Then run the stage notebooks in order from notebooks/ (they resolve the project root as Path("..")):

1. 01_data_collection.ipynb — scrapes both person databases into data/raw/ (long-running; resumable) and runs post-scrape QA.
2. 02_data_cleaning.ipynb — the whole cleaning pipeline: person parsing (§1–§2), the 1910 geographic backbones + census panels + settlement crosswalks (§3), cross-dataset linkage + place standardisation (§4), wide analysis files (§4b), data dictionary (§5). Writes everything under data/final/.
3. 03_visualisation.ipynb — descriptive figures and LaTeX tables from data/final/ only, into output/.

The person records are personal data and are not in the repository; processing rests on §90 zákon č. 121/2000 Sb. (the scientific-research exception) and GDPR Art. 89.

About this document

This page is generated from the repository by utils/docs/build_docs.py. The markdown files under reference/ are the canonical prose and every figure on the Descriptives and Data QA sheets is recomputed from data/final at build time — nothing here is transcribed by hand. To correct something, edit its source and rebuild; do not edit the HTML.

Chips sit at the end of a figure's own heading, one per affordance, and never anywhere else. ⇱ Interactive means the block has controls — click, toggle, drag or filter it. ↕ Hover for values means its marks carry tooltips; a block that does both shows both. The red dot in the sidebar is derived from the chips a sheet actually rendered, so the two cannot disagree. Long tables announce themselves with a visible search box instead. Everything else is a plain table or figure.

Module index

Every importable module, its version, and its purpose — read straight from the source files. Pipeline explains what each one does in context.

155 rows

DescriptivesOrientation

Live figures and tables computed from data/final.

Datasets

Outcome — exit from the Legions

D (deserted) = 3,717 is the primary outcome. Records with no exit code carry no outcome information and must be treated as censored, not as non-desertion.

Desertion by theatre

Rank reset (headline regressor) ⇱ Interactive ↕ Hover for values

Nationalism, the Legion and desertion, side by side

The two nationalism measures, the Legion and desertion from it, each as a rate per 1,000 on the same 1910 units, side by side. Provisional (D-163): each definition is the simplest one that can be drawn, and moves when its measure is ruled on.

In plain terms

Four maps of the Czech lands, one per measure, at three levels: the political okres, the court district and the settlement. A switch sets the denominator of the first three: all 1910 inhabitants, or the Czech speakers among them. The question they put is whether the places that paid for Czech nationalism before 1914 are the places that sent men to the Legion, and whether those men stayed.

The four measures

Czech speakers are the 1910 census's Umgangssprache; a rate per Czech speaker is left out where fewer than 1,000 speak Czech, so the districts with almost no Czech speakers are grey under that switch. The per-Czech Matice rate is shrunk and standardised over those remaining units only. The two denominators answer different questions: per inhabitant, how Czech-national a place was; per Czech speaker, how active its Czechs were, which a few Czech speakers in a German district can make high.

The Matice and theatre sheets say how their counts are built. The okres is the census-pooled okres (148 units, D-18); the court is the census court (330, the nine statutory-city courts pooled with their rural twin).

What the levels can and cannot carry

• Court. Only about 57 % of the legionnaires born in the Czech lands name a settlement or court; the rest are known to their okres. A court's legionnaires are scaled up by its okres's ratio of all births to located ones, which assumes the unlocated births fall among the okres's courts as the located ones do. The desertion rate is a ratio of located births and is not scaled.
• Settlement. The repo has no 1910 population per settlement yet (the 1910 register's full run will carry it), so the place view shows counts, not rates. Desertion could be a rate there, since its denominator is the legionnaires themselves, but shrinkage gives every settlement the same rate: Marshall's between-place variance comes out at zero, so across settlements desertion varies no more than chance would make it. The place view draws deserters as a count instead.
• Scale. Each panel has its own colour scale, clipped at its 97th percentile. The maps show where each measure is high, not how the measures compare in size.

Still open

• Rulings. The theatre window and bound, the Matice rulings (see the Matice sheet), and the per-Czech-speaker denominator for legionnaires born in German-majority districts, where a few Czech speakers make a high rate.
• Settlement population. It is needed for rates at settlement level.
• Slovakia. The Matice and theatre measures cover the Czech lands only, so the maps stop there; Slovak births are left out.

Run it (offline, under a minute each):

python utils/reference_tools/places/run_intensities.py
python utils/reference_tools/places/map_intensities.py

The maps ⇱ Interactive ↕ Hover for values

Unlike the rest of this sheet these come from data/interim/intensity/ (run_intensities.py), and every definition is provisional (D-163). One map per measure, the same units in each. Hover a district or a place and it is outlined in every panel, with every number in the tip. The second switch sets the denominator of the Matice, theatre and legionnaire maps: all 1910 inhabitants, or the Czech speakers among them (Umgangssprache); desertion is per legionnaire either way. Each panel has its own colour scale, from zero (or, for the Matice z and desertion, the 3rd percentile) to the 97th percentile; grey is a unit with no value — under 1,000 Czech speakers for a rate per Czech speaker, or under 20 legionnaires for desertion. 101,274 legionnaires were born in the Czech lands; 57,994 of them name a court or a settlement, so the court view scales each court up by its okres's share of births known only to the okres. The settlements have no 1910 population yet, so the place view shows counts as dots (area ∝ count, an outlined dot is capped), deserters among them: shrunk, every settlement gets the same desertion rate, since across settlements it varies no more than chance would make it.

How the maps move together

Each cell is the Spearman rank correlation ρ between two of the mapped rates, taken across districts: rank the districts by the row's rate and by the column's, and ρ says how far the two orders agree — +1 the same order, 0 no relation, −1 the reverse. Blue is positive, red negative, darker is stronger. The matrix is symmetric, so each pair appears twice, and the diagonal is 1. A cell uses only the districts that have both rates — a rate per Czech speaker is missing where fewer than 1,000 speak Czech, desertion where fewer than 20 legionnaires were born — so cells can rest on different sets of districts; hover a cell for its count. Matice is the within-territory z, the rest are the rates as drawn on the maps.

Across the 148 political okresy. Units with a value per Czech speaker: matice 123, theatre 123, legionnaires 123; with a desertion rate: 146.

Across the 330 court districts. Units with a value per Czech speaker: matice 241, theatre 241, legionnaires 241; with a desertion rate: 247.

Place matching — birth

Tiers are ordered strongest → weakest; muni and fuzzy are the soft ones a strict specification can drop. A record is placed in at most one 1910 frame.

Cross-dataset linkage

1,340 confirmed overlaps over 1,340 distinct VÚA records; 319 satisfy both criteria. The rule is precision-first: weaker name/year coincidences are dropped, not down-tiered.

Coverage of key variables

Figures

Generated by notebooks/03_visualisation.ipynb; PDF and beamer variants of each live in output/figures/.

Legionnaires

Habsburg fallen

figures

Data QAOrientation

Invariant checks on the current build.

Invariants 126/127 passing

Defined once in utils/validation/data_checks.py and evaluated against the current data/final every time this page is built. Notebook 02 calls the same list as its final cell, so a run that produced broken data fails rather than reporting success. Each check corresponds to a bug that was real at some point in this project — see the audits of 2026-07-16 and 2026-09-22.
FAIL is a hard invariant that should never be violated; WARN is a soft one (it can drift for external reasons) that reports but does not block.

Roadmap / to-doOrientation

What's done and what remains. · source: reference/ROADMAP.md

At a glance

✅ What is already built

The table above is the whole of it. This sheet looks forward and does not narrate finished work: for what changed and why it mattered see What's new, for the commit record Commits & versions, and for the calls taken along the way Decisions.

🔨 Open — variable construction

⚠ Known issues — logged in the 2026-08-02 review of the occupation work

Found by reviewing the occupation change before it was committed, and deliberately committed unfixed so the record and the code stay in step. Every figure below was recomputed from data/final/persons/legionnaires_clean.parquet and occupation_hisco_curation.csv.

📋 Open — data structures for the analysis

• Person-period spell file. The analysis will need a longitudinal structure — one row per person × period — that tracks who is still at risk in each period and how each spell ends (desertion, death, discharge, demobilisation). This is the main blocker between the current cross-sectional data and any modelling. Note ~30.8k records carry no exit code and are censored, not "missing". And 3,865 exit dates are known only to the month or the year (exit_date_precision, D-61): at monthly resolution they are interval-censored, not dated to the 1st.
• Unit panel. Aggregate to unit_key for unit fixed effects, clustering and size denominators. Mind the lineage caveat: units were renamed and the same name denotes different units at different dates (see the External validation sheet), so a unit panel needs the record date, not just the name.
• 1910 register → municipal tables — built 2026-10-29 (D-148), waiting for the full run. settlements_1910 has a key (settlement_id); notebook 15 §18 anchors every register municipality and part to it and writes data/final/register/ (places, institutions, cooperatives, one row per municipality), keys asserted in the QA gate. Built today from cz_certification as a test. Next: re-run §18 on the full run of each edition; German Böhmen goes through the Czech volume since D-150 (municipalities 66.7 % → 84.3 % on the gold), German parts and Mähren/Schlesien still rest on their own names; final_abbreviations.csv for the text and cooperative counts; the D-55 correction before any regression. Columns and caveats: REGISTER_VARIABLES.md.
• 1910 register text abbreviations — combinations decoded 2026-10-29 (D-156). The print builds institutions out of its own abbreviation parts (hosp. pokr. k., Waisenh.); both printed lists are in abbreviation_components.csv, the code table grew by 148 rows, and notebook 15 §8b lists what no code reaches. Next: read that census on the full run and add what it prints; replay the frozen gold passes from their crop caches (Heavyhorse6) before scoring abbreviations; rule on the German GB. (the list prints Gerichtsbezirk, the table reads a court).
• German table reader — built 2026-10-29 (D-160), not yet read by a model. Böhmen, Mähren and Schlesien's table is cut into blocks off the image (no OCR) and read by the Czech table's two readers; the frozen route reads it after each German edition's notes, and the register names German places off it. Next: the first paid read on Heavyhorse6 (Böhmen gold, then Mähren/Schlesien calibration); score Böhmen against the German gold (rf.score_table) and look at the M/S rows' sums.
• The route's pre-freeze review — fixed 2026-10-30 (D-164 to D-167), not yet run. Six reviews of notebook 15, 47 findings ruled by the author: the route, the ledger, the Opus calls, the German table, the readers' caches, the register build and the sealed scorer, each fix with its test. Next, on Heavyhorse6: the Gemini batch check (7e) once, then "bohemia" / "gold" with REREAD — every pass from the cache where its prompt is unchanged; the Czech gold's Opus calls whose example cards changed, and its place-type calls, are asked again — then "de-bohemia" / "certification". The scorer refuses a run not read on its gold's frozen route (D-165), so nothing may change between the gold re-run and the last certification run.
• German half certified 2026-10-31 (D-164's route): F1 0.957 [0.927, 0.982], pooled over the 8 sealed pages (Böhmen 0.976, Mähren 0.984, Schlesien 0.909); 0.966 against the sheet the author's rulings corrected. 40 of its 54 errors were two look-alikes, now settled by where the print sets them (D-169: the cooperative mark before a colon; a poorhouse in the castle's slot), replayed at 10 fixes and 0 breaks. Next, on Heavyhorse6: the gold re-read with the rules (REREAD), then the full run; the fixed route's own measure on the sealed pages is in-sample (D-170). Open: one ruling, Schlesien 0038 Kuttelberg 11,0.

• German half certified a second time 2026-10-31 (D-170), in-sample: F1 0.995 [0.991, 0.999] (618 right, 1 false, 5 missed; entries exact 128/133; Böhmen 0.998, Mähren 0.998, Schlesien 0.990). The rules were found on these pages, so this is never quoted in place of round 1 or the Czech 0.996. Covariate symbols, blind round 1 / round 2: school 103/104 → 104/104, fire brigade 92/93 → 95/95, cooperative 69/71 → 71/71, midwife 60/62 → 62/62, post 34/35 → 35/35; false claims 1, 3, 0, 1, 1 → 0 each. Castle 6/14 → 13/14, poorhouse 7 false → 0. Text abbreviations on the sealed German pages (score_sealed_abbreviations.py): round 1 144/144 hit, 1 false, F1 0.997; round 2 139/144, 1 false, F1 0.979 (five cooperative letters lost on the re-read); Czech waits for the full run. The full tables are in LEXIKON_1910_ROUTE.md ("Everything certified, in one place"). DINOv3 still routes nothing for German (judges 0.115 of claims against 0.863 Czech); the rescale test (run_scale_e3_test.py, line height 66) has not completed. Next: the full run on this route.

• Before the full run — 2026-11-01 (D-173 to D-179). The gold re-read and re-frozen on Heavyhorse6 (14:31); the tree on disk checked identical to that freeze, every route module and input, and 754 tests pass. Four reserve pages (Mähren 71 and 139, Schlesien 28 and 48) are sealed for the next certification (D-173); the entry-level validation sample is not drawn (D-174). German DINOv3 dropped (D-175 to D-177): with a box certainly on the glyph and German exemplars in its bank it names German pictograms well, but nothing finds them on the German print — the lighter ink breaks the strokes and the glyphs touch their neighbours, so the candidate step loses 9–29 % of the author's 203 clicked glyphs and the placement puts the box on the right glyph for at most 27 % of claims; given the reader's own boxes, cut tight, it still kept only 0.856 of true claims against a bar of 0.936 written first, filing one German pictogram in seven with the letters around it. In the route a DINOv3 doubt sends a claim every reader agreed on to Opus, so that is thousands of extra Opus questions on true claims to catch agreed false claims running at about 1 in 620 — at most about 0.1–0.15 pp of precision, and nothing for misses, 5 of the 6 German round-2 errors. The German route stays as certified, DINOv3 recorded and routing nothing. Every page was then cut offline before anything was spent: the 17 statutory-city pages are skipped and their institutions read from the prose (D-178; both stages read, 5 groups ruled by the author, 1 open), and German pages with faint rules are cut by position and the volume's template (D-179; page audit 0 cut fails, 57 through a fallback). Register 1.6.0 types every school on the legionnaires' education ladder (n_school_<type>, school_top_rung); 1.7.0 keeps every claim's way back to its glyphs (entry, marks → the run's ledger). The run is held to completeness by outcome (D-180): a reader's request with no reply or an Opus call with no answer — an exhausted balance included — stops the ledger until it is filled or named in ACCEPT_FAILED. Still open on the person side: legionnaires reach the register at court or okres level only (no birth settlement_id), to be fixed after the run. Next: the Schlesien full run on Heavyhorse6 as a smoke test (37 pages, none through a fallback), then Mähren (its 9 fallback pages), then Czech and German Böhmen; notebook 15 §18 then builds the register.

📋 Open — analysis

The modelling stage has not started and the approach is still being worked out; this page deliberately does not commit to a specification. What is already framed:

• Question: what drives a second desertion — the revealed-preference core of the design.
• Covariate of interest: relative rank deprivation on Legion entry (rank_reset), read on combat ranks (~legion_noncombat) and examined by theatre (Russia carries essentially all the desertion).
• Identification threat (Ottinger): enrollment officers may have demoted soldiers showing prior disloyalty — to be addressed with late-enrollment restrictions and balance tests.
• Occupation is missing in a way that tracks the outcome (found 2026-08-27, open). In Russia, occupation is absent on 8.87 % of Code-D deserters against 0.31 % of the demobilised, and the gap holds inside every enrolment cohort (1914: 27.3 % vs 1.0 %; 1916: 23.6 % vs 0.3 %; 1918: 7.2 % vs 0.2 %). It is not archival coverage — the two groups hold the personal card at an identical 87.11 %, and conditioning on the card the gap is 7.96 % vs 0.16 %, so the document is present and the occupation line inside it is blank. Consequences: (a) complete-case analysis on occupation silently selects on the dependent variable — carry missing as its own category; (b) report results with and without the occupation control; (c) the archival question — whether the field was entered at enrolment or completed over the course of service — is now worth putting to VÚA directly, because the answer decides whether occupation can be treated as pre-determined at all. The provenance flags (D-44) describe this differential; they do not control it away.

⚠ Left open by the 2026-09-22 audit

Everything the audit found is fixed, except what it chose not to decide:

• education_complete (D-53) — still conflates years attended with classes completed; the recommended fix (<NA> on a bare ob. N) changes an emitted column and waits for a decision.
• Probable duplicate persons — 625 csol and 281 VÚA rows share surname, given name, full date of birth and birth municipality with another row of the same file. dup_group flags them; whether to collapse any is an analysis choice, not a cleaning one.
• Five Wikidata points shared inside one 1910 okres (11 settlements) — harmless for every okres-level variable, wrong for anything finer than the okres; kept rather than guessed.
• Four settlement timelines still carry another entry's text: three where the register itself omits the comma between two clauses, one narrative. Below the level at which a rule would not also break a correct entry.
• Anything computed on the old linkage — the overlap is 1,340 men, not 2,239, and the pairs that left are mostly the note-confirmed ones (463 → 139). Every comparison-design number from before 2026-09-22 has to be redone.

⏸ Deferred / out of scope

• Occupation / education refinements — as above, when the analysis needs them.
• The Slovak half of the party scheme. affiliations.py 2.0.0 rests on Akademický atlas plate IV.37, which is v českých zemích and so contains no Slovak party. SL (108 records, the largest unmatched token), ludovák/ludová, and the distinction between Slovenská liga and Hlinka's Slovenská ľudová strana are unresolved and recorded as refusals (D-50). Needs a Slovak party register — the natural counterpart to IV.37. 2026-11-03: the records settle SL without one. It is the Slovenská liga v Amerike (Slovak League of America), an émigré organisation: 188 of the 214 records naming SL or a liga mention the United States, and 139 enrolled in the Czechoslovak army in France, where the American volunteers went. The same check found a defect: ČNS, coded nar_soc (Č8), is on 220 records, 189 of which mention the United States, with applications filed in Cleveland, Chicago, New York and Omaha. That is the České národní sdružení (Bohemian National Alliance), not the National Social party. Fixed the same day (D-187, affiliations 2.3.0): SL joins slov_liga (63 → 211) and ČNS is its own dummy pol_cns (222), out of pol_nar_soc (6,541 → 6,319). Notebooks 02 and 03 must be re-run for data/final and the affiliation figures to carry it. A Slovak register is now needed only for ľudovák (15 records).
• Habsburg military-court inventories (VÚA) — examined 2026-10-12; closed for good 2026-11-03 (the author: no way of doing this). Three name-level inventories were parsed (utils/parsing/vua_courts.py: 2,330 case files, 3,034 accused, 899 soldiers) and the soldiers linked to the legionnaire file on name + birth year + Habsburg unit (utils/reference_tools/units/run_vua_courts_pilot.py). The link is barely above coincidence — 96 soldiers with a candidate against 83 and 77 under a shifted-birth-year placebo, 8 picks corroborated by unit or birthplace — so it gives no person-level measure of prior disloyalty; and at district level the 1,896 civilians with a birthplace spread over 1,241 places, too thin for a covariate. Not pursued; the code and the DATA_SOURCES.md #23 entry stay as the record.
• Amateur-theatre database (DbČAD) — scraped in full 2026-10-12; being built into a Czech-nationalism treatment since 2026-10-28. The Databáze českého amatérského divadla (NIK) was tried as a settlement-level, dated measure of pre-1914 Czech associational life in a legionnaire's birthplace. Every one of its 7,368 place entries is cached and parsed (utils/scraping/scraper_amaterskedivadlo.py → data/raw/theatre/places.csv and ensembles.csv): 34,569 ensemble rows with start and end years, a fuzzy 191x kept as a decade and never coerced to a year; and per place the první zpráva (first surviving report — not a founding date) together with the three free-text blocks kept whole for later reading — the chronicle (7,340 places), the bibliography (6,391) and the archive citations (5,359), one entry per line. 93.6 % of the places tie one-to-one to a row of settlements_1910. What the coverage tabulation shows (utils/reference_tools/places/run_amateur_theatre_pilot.py --all; every figure here is its printed output of 2026-10-13): an ensemble certainly starting before 1914 in 14.5 % of Bohemian and 21.0 % of Moravian 1910 settlements, but the gap is composition — inside okresy at least 90 % Czech-speaking it is 30.0 % against 28.2 % of ordinary communes. Three things stand between this and a covariate. The database records Czech theatre only (17 of 34,569 ensemble names look German), so a German-speaking village is uncovered, not zero, and language is known only at okres level until the 1910 register is merged. Coverage rises steeply with size — 69 % of judicial-district seats, 21 % of other communes, under 3 % of hamlets — and the repo has no settlement-level 1910 population, so size is a three-class proxy. And each okres was compiled from its own sources: first-report years heap on one year per okres (Jihlava 1929 × 21, Hradec Králové 1903 × 17, Kutná Hora 1949 × 21, all citing one district report), which argues for comparisons within an okres. Nothing is joined to the person files and nothing feeds data/final. When resumed: read the chronicle and archive lines for dates the ensemble list lacks, find a size measure, and settle language at settlement level. The chronicle is read (2026-10-27, amateur_theatre 0.2.0, D-137): chronicle_lines() turns every place's chronicle into one row per line, keyed (geo_id, line_no) — a dated event, an undated note, a periodical citation, the database's admin note on the present okres, a heading (a Loutkové divadlo heading opens a puppet section) or a separator — with the leading date as start and end intervals, the associations the line names (15 labels: the pilot's five plus the free-standing dramatic society, reading, singing and educational societies, Omladina, workers', Catholic, the national-defence unions, school and church) and a puppet flag. Figures from chronicle_report() on the full cache, 2026-10-27: 38,940 dated events in 7,115 places, 8,409 of them certainly before 1914; 2,966 places have a pre-1914 non-puppet event, and 256 of those have no ensemble certainly before 1914 — the chronicle adds organisations and dates more than places. The pilot writes the table to data/interim/theatre/chronicle_lines.csv; nothing goes to data/final. The database's own abbreviation key is read in as well (2026-10-27, amateur_theatre 0.3.0, D-138): saved as data/raw/theatre/abbreviations.csv (309 entries + 43 source sigla), it raises the pre-1914 events naming a free-standing dramatic society from 1,111 to 1,434 and leaves 3,130 of 8,214 naming none; 33 rest on the two-way ČOS and are flagged, 92 name ÚMDOČ. Still open: the archive and bibliography blocks, and the settlement key the person files lack (birth_settlement_id), without which no person reaches a theatre place except through birth_qid. Source, citation and rights: DATA_SOURCES.md #22. The treatment is being built (2026-10-28, amateur_theatre 0.4.0, D-140): the end product is a place × year activity panel, 1800–1913, with certain and possible bounds: companies active, performances, distinct plays, companies performing, attendance, foundings, national-cause proceeds; German, bilingual and puppet theatre in columns of their own. Settlement, court and okres tables are built from it, and nothing is joined to persons yet. Step 1 is done: chronicle_clauses() cuts each note where a new date opens (data/interim/theatre/chronicle_clauses.csv). That makes 8,656 pre-1914 non-puppet clauses in 2,974 places, up from 8,219 lines, and each one is tagged with the event types its words point to (EVENT_CUES, drafts and not codes). Step 2 is done (D-141): the author ruled each type and twelve design calls on the codebook page. Step 3, first pass, is done with rules: code_clauses(), then evidence_table() (the ensemble list, the clauses and the first reports in one interval form, with one company_id per company), then place_year_panel() (7,368 places x 1780-1913, each measure certain and possible), then area_year_tables() (summed to 1910 settlement, court and okres). All of it goes to data/interim/theatre/. Czech activity is certain in 3,063 places. Next is a model pass on what rules cannot settle (Haiku 5.5 against Sonnet 5.5 on a ~300-clause gold sample; prepared with spending off and run on Heavyhorse6), and the evidence table and panel. Step C is in (2026-10-28, amateur_theatre 0.6.0, D-145): the 682 ensembles that start before 1914 and print no end were scraped on Heavyhorse6 (data/raw/theatre/ensemble_pages.csv). The pages almost never give an end (1 year), but 588 list the decades active and 284 carry a chronicle of their own, 97 % of it absent from the place chronicle. By the author's rulings, a listed decade is possible activity, and a page that lists decades replaces the open end. A printed end closes the span, certain. The page's chronicle is coded like a place's and tied to its ensemble. Certain places 3,063 → 3,068; certain place-years 52,700 → 53,403; possible place-years 93,113 → 91,651. data/interim/theatre/ is built with it (2026-10-28, --all --reuse-tables). The rest of step C is in (2026-10-29): the pages of all 7,103 pre-1914 non-puppet companies are cached (the other 6,421 fetched from the laptop, 0 failed), and the runner has rebuilt data/interim/theatre/ on them. Every one of the new companies already had an end in the list, so they add what the companies did more than when: certain places 3,068 → 3,077 and certain place-years 53,403 → 53,696, but certain performances 2,917 → 11,322 (in 915 → 1,680 places), distinct plays 769 → 1,625, foundings 3,574 → 5,981. The ensemble chronicles now give 59,864 clauses, 13,193 of them left to the model pass. The full method is on the Theatre measure sheet.
• Matice česká founder register — downloaded and read into tables 2026-10-29. The society's yearly list of its founders (zakladatelé), issues 1873 and 1890, in library/associations/ (utils/scraping/run_matice_downloads.py; source, rights and caveats in DATA_SOURCES.md #28). Each entry gives name, occupation, residence, entry year and membership number; the 1890 issue adds the deceased and the corporate founders, so it reaches back to 1831. A candidate okres-level measure of committed Czech nationalism beside the theatre measure; the design (unit, denominator, local vs posted founders, Moravia) is still to be discussed.
• Matice moravská member lists — located, downloaded and read into tables 2026-10-29. The Moravian counterpart, printed inside Časopis Matice moravské: full rosters in roč. 15 (1891) and roč. 33 (1909, with a tabulation by occupation and place), partial lists for 1876–1878, the 1891 additions and the local trustees (důvěrníci). 1878 is held as an Internet Archive scan; the rest are jobs mm1876 … mm1909 of run_matice_downloads.py v1.1.0, run 2026-10-29 (7 items, 56 pages, every one with scan, OCR and ALTO). A small, educated membership, so an élite measure; source and caveats in DATA_SOURCES.md #29.
• Both registers are read by rule (2026-10-29, matice 0.1.0, D-158). utils/reference_tools/places/run_matice_pilot.py writes data/interim/matice/matice_entries.csv, one row per printed entry with its own text beside name, occupation, residence (as printed), entry year, membership number, year of death and amount paid, and matice_totals.csv, each list against the total its society printed: Matice česká persons 99.2 % (1873) and 99.7 % (1890), deceased 97.8 / 98.9 %, corporate 106 / 96 %; Matice moravská 1909 97–100 %, the 1891 run-on roster 92–94 %. The membership number links 3,125 Matice česká founders across the two issues, and gives 2,350 of 2,396 deceased founders an entry year. The 1876 and 1877 lists' (Pokračování budoucně) turned out to continue in what is held: the 1876 contributors in the 1877 list ("a) Za r. 1876"), the 1877 ones in the 1878 Internet Archive list ("Za rok 1877"), which is now read from its PDF text layer (310 entries, each dated by paid_for_year).
• The registers are an okres measure now (2026-10-29, matice 0.2.0 + locative 0.1.0, D-161, provisional). Residences are matched to 1910 settlements by declining every 1910 name into its locative and genitive forms (utils/parsing/locative.py): of Matice česká entries 49 % reach a settlement and 17 % Prague, 10 % are refused as ambiguous, 7 % lie abroad, 11 % are unmatched; of Matice moravská entries 73 % and 5 %. data/interim/matice/matice_okres.csv holds, per census okres, the founders and members placed there, rates per 1,000 inhabitants and Czech speakers, shrunk rates, posted and local counts apart, and matice_own_z — each land's own society, standardised within the land. Against the theatre measure (Czech-majority okresy, 1880–1890) it correlates 0.25 in Bohemia and 0.47 in Moravia. To rule: denominator, whether posted founders count. Ruled 2026-11-02: Moravia and Silesia are measured by the Matice moravská roster of 1 March 1909 (D-182; Moravian theatre correlation 0.49 → 0.37 against the 1890 list, the two lists rank okresy at Spearman 0.84). Matice česká has no founder list after 1891, so Bohemia stays on its 1831–1890 founders. Nothing reaches data/final or the persons yet. Name variants (locative 0.2.0): a settlement is found by its present Czech name, its 1910 name, its older Czech names (settlement_aliases_1910), a short list of historic names (Holomouc, Král. Vinohrady), old spellings (Oustí), abbreviations (Jindř. Hradci), and its German name both as written and as a Czech list spells it (Švarcbach, Šildperk); an OCR misspelling is matched only at rapidfuzz ≥ 96 with a clear margin, a cutoff set by reading every match at 92 (13 entries). Places abroad, Galicia and Upper Hungary included, are caught before matching. Since locative 0.3.0 the commonest OCR misreadings are repaired before matching (cli→ch, x→h, initial Y→V, a word cut in two, stray marks), and a qualifier is itself located: a settlement within 30 km of the place it names is preferred among homonyms, and a misspelling scoring 85–96 is accepted only when exactly one such nearby settlement fits (Pia šech u Plzně → Plasy). The parser also cuts entries the OCR ran together (… v Hranicích Štolba Josef, …). The 954 entries still unmatched after that (all but the 1878 PDF's) were re-read from line crops of the scans by a vision model (D-162, reference/places/matice_residence_readings.csv), and nineteenth-century spellings are tried as variants (Žehouni, Hořejní, Mezeřiči). Unmatched entries now: Matice česká 7.0 %, Matice moravská 4.3 % (694 in all, from 1,106). What remains is mostly small places the 1910 table does not hold and places abroad with no marker. Court districts and maps: data/interim/matice/matice_court.csv carries the measure per court district (court census 1910 as denominator, the nine statutory-city courts pooled as the census pools them), and utils/reference_tools/places/map_matice.py draws it by okres, by court and by municipality to output/figures/matice/.
• Intensity maps, side by side (2026-10-30, D-163, provisional). utils/reference_tools/places/run_intensities.py writes data/interim/intensity/intensity_{okres,court,settlement}.csv: Matice (matice_own_z), theatre companies active per 1,000 (1880–1913, certain), legionnaires per 1,000 inhabitants and per 1,000 Czech speakers, and desertions per 1,000 legionnaires (shrunk). map_intensities.py draws them to output/figures/intensity/, and the docs site's Descriptives sheet carries them as linked panels, with the definitions from INTENSITY_MAPS.md above them. Limits: 57 % of Czech-lands births name a court (court counts scaled to the okres), and settlements have no 1910 population, so the place view shows counts. Next: settlement populations from the 1910 register; the rulings on each definition.
• Matice česká joiners, 1886–1914 — found, downloaded and read 2026-11-03 (matice 0.4.0, D-184, provisional). Every issue of the Časopis Musea království Českého (Národní muzeum, Data sources #31) closes with the month's payments into the Matice, so the founders who joined after the register stops (1891) are named after all. run_matice_downloads.py 1.2.0 finds the lists by full-text search (--discover-payments → reference/places/matice_payment_pages.csv, 94 slices, 186 pages) and fetches them (--set payments); matice.parse_payments reads them month by month against each month's printed total, payment_persons gathers a payer's instalments, and the joiners of 1900–1909 are now Bohemia's measure, both lands standing at about 1909. The pilot writes matice_payments.csv, matice_payment_months.csv and matice_payers.csv beside the other tables and prints the checks. A joiner is dated by the first payment (ruled 2026-11-03). Placing improved the same day (locative 0.4.0: abbreviations that lost their last full stop, one-letter rivers, Prague streets, neuter plurals; matice.borrow_places): payments placed 82.7 % → 84.4 % (Prague included), the measure itself almost unchanged (Spearman 0.995). The combined measure is a log density since D-185, following Squicciarini & Voigtländer (2015): ln(1 + members per 100,000), z within territory, Prague (z 3.10, 8.26 in levels) a dummy in every regression, the level z and an any-member dummy as robustness. A small audit of the Matice and theatre code the same day fixed seven things (matice 0.6.0, amateur_theatre 0.6.1): founders keyed by number and surname (the OCR gave two people one number: 3,037 → 3,296 founders placed in Bohemia), institutions named anywhere kept out of the persons (Rokycany's gymnasium had been three joiners), a hyphen after an amount, OCR'd titles and totals, a stray bar read as a digit, each payer's place and occupation taken from whole payments, the name order read per month (1890 register check 50 % → 68 %), the theatre check set against each measure's own years, and the theatre tables rebuilt and keyed by settlement_id. The measure moved little (Spearman 0.996). Open: the window; that the joiners do not track the theatre measure inside Czech-majority Bohemia; and a look by eye at the months furthest from their printed total. Both denominators are carried (D-186, ruled 2026-11-03): per inhabitant (matice_own_z) and per Czech speaker (matice_own_z_czech, units with at least 1,000 Czech speakers), now built in matice 0.7.0; the intensity tables' matice_per_1000, which had been per million, is per 1,000. The window is 1895–1909 (D-188, ruled 2026-11-03): a decade was too few people (median two joiners per okres), and the wider window ranks okresy closer to their joining outside it (Spearman 0.83 against 0.79); 973 joiners placed in Bohemia, 25 okresy without one (35 before), matice_own_z 0.93 with the 1900–1909 version. Still open: whether posted members count (329 of the 973 Bohemian joiners are priests, teachers and officials, and 323 students).
• Side study: Habsburg censorship and the disruption of Czech amateur theatre, 1780–1914 — found 2026-10-28; dropped 2026-11-03 (the author: no longer relevant). The DbČAD chronicle (data/interim/theatre/chronicle_lines.csv) records refusals and bans of named performances with place and year, and nearly all of them hit local players — students, burghers, clergy, from the 1860s the amateur societies — not travelling companies. Three regimes show: church and state bans on passion and religious plays in the 18th century; the Bohemian governor's office (gubernium) refusing Czech student and burgher performances in the 1810s–1840s, sometimes over a regional office's recommendation; and district offices (hejtmanství) refusing permits for single named plays, or punishing an unpermitted one, in the 1860s–1880s — censorship at the level of the play, with the stated reason sometimes given. A keyword pass found 46 pre-1914 ban notes in 41 places, about a third of them WWII bans or noise, so nearer 30 real ones; beside them about 50 fires of the venue (inn, town hall, Sokol hall) that stopped theatre, a shock from outside the society, and about 40 suspensions (Bach's absolutism, WWI). Small, but dated and placed, and it links to Purš's repression of the tábory (D-37) and to the repression-and-loyalty mechanism of Rozenas, Talibova & Zhukov (2024). The event codebook of the theatre work codes these types; the study would need the archive block's governor's-office files and the cited SÚA Podání register to go further. Counts from the exploratory scripts of 2026-10-28, not from a build.
• PhD-pipeline paper — a separate interwar-electoral project sits outside this thesis.

Edit this file (reference/ROADMAP.md) as work progresses and rebuild the docs; it is the canonical source for this page.

What's newOrientation

Site updates + the live-page link. · source: reference/SITE_CHANGELOG.md

Live page: tomashaufer3.github.io/legions-docs — unlisted (a noindex tag keeps it out of search); share by link.

2026-11-03

• The Overview's map has the denominator switch. As on Descriptives, the Matice, theatre and legionnaire maps can be read per inhabitant (the default) or per Czech speaker; desertion stays per legionnaire. The map now draws the Descriptives panels themselves (_IX_PANELS), so its columns and labels cannot drift from theirs.

• The Overview opens with five interactive panels. Under the summary card, one tabbed At a glance block: the pipeline, every source, build step, check and the planned analysis as one diagram with its live count and status (hover for what it is, click for its sheet); desertion over time, by month and theatre, as desertions, enrolments, men in service or the rate per 1,000 men in service — the spell sample the hazard model will use, enrolment and exit month both known — with the February revolution, Zborov, the coup and the armistice marked; the map, Matice, theatre, legionnaires and desertion per inhabitant by 1910 okres (provisional, from data/interim/intensity); one page of the 1910 register, each zone glossed on hover; and one man, two decisions, a complete record drawn from a fixed-seed sample of 300, half of them deserters, without name or municipality and dated to the month.

• The Overview makes the case for the project. It was the repository README: data table, publishing notes and repository tree, with its numbers typed. It now runs question → the double-desertion design → the headline fact (the Russian monthly series around the Bolshevik coup, and the desertion rate by theatre drawn from the build) → the data → how it is checked → where it stands, then the reading routes, every sheet, how to run it and the module index. Its prose is reference/OVERVIEW.md, whose {{token}} slots overview_facts() fills at build time from data/final, the curation tables, the Ostafin check's output and the QA gate; a token the builder does not compute stops the build. Typing the numbers had already gone wrong once: the README's "under 0.5 % in every other theatre" is contradicted by the Serbian/Balkan formation, which the theatre bars now show with its count. The OccCANINE figure is computed by agreement() in run_occcanine_check.py 1.1.0, shared by the check and the page, on the current curation.

• Bohemia is measured by the Matice česká joiners of 1900–1909, on a log scale. The founder register stops in 1891, but every issue of the Časopis Musea lists the month's payments into the society; Matice, Data sources #31 and the intensity maps on Descriptives now take Bohemia from those lists: a joiner is a person whose first payment falls in 1900–1909 (D-184). The combined measure is ln(1 + members per 100,000), standardised within each society's territory, as Squicciarini and Voigtländer (2015) build Encyclopédie subscriber density, with a Prague dummy for every regression (D-185); Prague drops from +8 to about +3 on the map. The theatre check sets each Matice measure against theatre of its own years. A small audit of the Matice and theatre code the same day fixed founders the OCR gave one membership number, schools and offices counted as joiners, and the theatre tables' settlement key (Roadmap).

• Matice moravská is its 1909 roster. On Matice and the intensity maps on Descriptives, Moravia and Silesia are now measured by the members of 1 March 1909, the list nearest the 1910 census, not those of 1890 (D-182); the map circles, hover labels and the combined matice_own_z follow. The theatre check keeps the 1890 rate as a row for comparison. Bohemia stays on the Matice česká founders of 1831–1890, since the society printed no founder list after 1891.

2026-11-03

• The register's German table has an outside check (D-183). Ostafin et al. (2020) typed every Silesian commune's 1910 census from the same volume the full run read. The Validation register and the reading route now carry the comparison: the census we use is identical in all 12 political districts, and the full run's table equals their transcription on 98.6 % of 441 communes. The same comparison shows why the ČSÚ lexicon 1869–2005 does not reconcile: its 1910 figures sit on 2005 parts. The source is on Data sources and in the references.

2026-11-02

• The correlation tables are full matrices that say what they are. How the maps move together on Descriptives printed a lower triangle with numbered columns; it is now the full symmetric matrix, its columns named and grouped by denominator (per inhabitant, per Czech speaker, per legionnaire), shaded by sign and size, with each cell's number of districts on hover — a cell uses only the districts that have both rates.
• Ten more requirements in the legionnaire sample builder. Build your own sample adds the hazard model's spell (enrolment date known, exit date known, no exit before enrolment) and what it conditions on: Czechoslovak Legion only (no foreign armies), combat rank at entry, born in the Czech lands, birth year known, occupation class coded, education rung coded, and no probable duplicates. The link and the pandas snippet carry them.
• Four intensity maps, and a denominator switch. On Descriptives the maps are now Matice, theatre, legionnaires and desertion; the second legionnaire panel (per Czech speaker) became a switch that sets the denominator of the first three, all 1910 inhabitants or the Czech speakers among them. run_intensities.py 0.2.0 adds the Matice and theatre rates per Czech speaker (the Matice one shrunk and standardised as per inhabitant); desertion stays per legionnaire, and the settlement view, which has no population, keeps counts. The correlation tables carry both denominators.
• The settlement maps can be read. On Descriptives the place view of the intensity maps drew every settlement as a dot of radius 3.5–22 on a panel about 490 units across, so nearly 5,000 overlapping dots covered the map; they are now scaled as on the Matice map (0.9–7) and translucent. The desertion panel was black: shrunk, every settlement gets the same desertion rate (between-place variance comes out at zero), so its colour scale had no width and every dot was given an invalid colour. It now draws deserters as a count.
• Every script sits in its stage. Nine parsing modules — among them occupations, religion, provenance, ruian and the theatre model pass — had been filed under Output & support because the role table missed them; they now sit under Person parsing or Geography & census.
• Pipeline names every module again. The guide still described three notebooks and the state of 16 July, and named about a third of the modules in utils/. It now covers notebooks 15 and 16 and notebook 02's §3e, and adds three sections: the downloaders beside the two scrapers, the 1910 place register (the runner, the page modules, the readers and checks, the ledger and register build, the gold and the score), and the reference tools and pilots. Each module gets at least one line; its own sheet still holds the detail.

2026-11-01 (the register before its full run)

• The run cannot finish with replies missing (D-180). The reading route's Notebook 15 — how it runs table gains ACCEPT_FAILED and says what 7d, 12 and 14b now check: every queued request has its reply and every Opus call its answer, whatever stopped one, or the ledger is not built. 1910 register — the municipal data documents the two new columns that lead from any register claim back to its glyphs' history in the run's ledger, entry and marks.

• The statutory cities: 26 open groups down to 1. The reading route and the municipal-data sheet say how: the judge's rows are now accepted with a count printed after se/mit/( and a code given as its expansion (20 groups, from the cached verdicts), and the author's ruling settles five (a council with language sections is the council and a row for its sections). Prague now agrees across the two editions on transport and professions too.

• 1910 register — the reading route: the route chart (The route a mark travels) now shows all four editions — German Moravia and Silesia join Czech and German Bohemia — and says of Mähren and Schlesien that they have no gold of their own and are held to the Böhmen gold's freeze. The step texts carry the decisions since 30 October: the statutory-city pages skipped and read from their prose (D-178), the faint-rule cut on German pages (D-179), the look-alike rules from the print's syntax (D-169) and the German part-0 question (D-171), and the entry-level sample not drawn (D-174). Step 3 now states why DINOv3 routes nothing on the German volumes (D-175–D-177): it names a correctly boxed German glyph well, but nothing finds the glyphs on the lighter German print, and with the reader's own boxes it still flagged one true claim in seven, each of which would have gone to Opus.

• The chart counts Opus again. Since the claim-level rulings, a claim's rule joins its marks' rules (claim ruled: opus + opus, Czech anchor, claim ruled: opus), and the chart knew only the bare opus: most of Opus's settlements fell under "no band" (on German Silesia, Opus showed 0). A claim is now put under the strongest step that settled it, and the D-169 position rule is a fifth exit, off step 3: settled by where the print sets it. Every rule string on disk lands in a band.
• The German certification by round. The chart and the Is the sample large enough? tables label round 1 blind, the record and round 2 in sample (D-172); the site had been showing round 2 as the certification. Those tables had also been missing since the paragraph they hang from was renamed; the build now warns if that happens again.
• The sheet's certification section is corrected where the pre-run review found it overstated: eight German sealed pages, not four; the full run's readings of the sealed pages are a test–retest, not an independent measure; exact bounds on the error counts lead, and errors that cluster by geography are named.
• Roadmap, Decisions and 1910 register — the municipal data carry the state before the full run: D-172 to D-179, the statutory cities' institutions (register_statutory, printed totals flagged) and the school types on the education ladder (register 1.6.0).

2026-10-30 (intensity maps, side by side)

• Descriptives gets a section, Nationalism, the Legion and desertion, side by side. It shows Matice, amateur theatre, legionnaires per 1,000 inhabitants and per 1,000 Czech speakers, and desertion per 1,000 legionnaires as five linked maps, which can be switched between political okres, court district and settlement. Hovering a unit outlines it in every panel and lists all five numbers. The definitions and what each level can carry (reference/INTENSITY_MAPS.md) sit above the maps, and a rank-correlation table follows. Built from data/interim/intensity/, provisional (D-163).
• Matice map dots: the settlement circles now stop growing at the 99th percentile. Prague, Brno and the other largest places get an outlined dot instead of a disc covering half a land. The white halo is gone, and the dots are opaque enough to read over the shading.

2026-10-30 (the Matice measure gets a sheet)

• Matice measure, a new How it works sheet: the two member registers, how the entries are read from the scans, how a residence finds its 1910 settlement, the measure, and what is open. Its tables are built from data/interim/matice/ at each build: entries found against every printed total, where the residences landed, and the check against the theatre measure.
• An interactive map on it: every 1910 court district shaded by its political okres or by its own value — the combined measure or either society's shrunk rate per 1,000 — with the settlements' members as circles that can be switched off, and the numbers on hover.

2026-10-29 (the Matice česká and Matice moravská member registers)

• Data sources #28: the Matice česká founder register, issues 1873 and 1890, downloaded into library/associations/ as a candidate measure of committed Czech nationalism. The row says what one entry holds, the three lists of the 1890 issue and the three caveats (residence at the list date, appointed clergy and officials, two issues only). Not parsed yet.
• Data sources #29: the Matice moravská member lists printed in Časopis Matice moravské — the full rosters of 1891 and 1909 (the latter tabulated by occupation and place), the partial lists of 1876–1878, the 1891 additions and the local trustees — located from every public volume's contents and downloaded beside #28.
• Both registers are now read into one row per printed entry (utils/parsing/matice.py, D-158), each list held to the total its society printed; the module and its runner have script sheets.
• Their residences are matched to 1910 settlements (utils/parsing/locative.py, a new script sheet) and summed into a provisional okres measure (D-161), with its check against the theatre measure on the Roadmap. The matcher also reads German names and guarded OCR misspellings, the measure is carried down to court districts, and a new script sheet, map_matice.py, draws it by okres, court and municipality. Residences the OCR garbled are re-read from the scans and kept as a curation file (D-162).
• The downloaders in utils/scraping/ are now all filed under Scraping on the script sheets; the Kramerius modules and the run_*_downloads runners had been under Output & support.

2026-10-29 (the register becomes municipal data)

• 1910 register — the municipal data, a new How it works sheet: the four tables in data/final/register/ (places, institutions, cooperatives, one row per municipality), every column, how a claim becomes a variable, and the caveats. Its numbers are a block the build rewrites (D-148).
• Brzice carried past the frozen run on Digitising the 1910 register, in a new section under What one village had: the municipality and its four osady anchored to settlements_1910 (with the runner-up and why Grund is found through its 1910 name), the four claims, Brzice's row in the municipal table, and Brzitz found through the Czech volume. The last step shows the German alphabetical order the Czech table predicts, the gap the gold page cannot close (Aujezdetz is Újezdec, on a later Czech page) and the leave-one-out test of the rule (D-150). Built by utils/lexikon_1910/build_register_walkthrough.py from the gold runs.
• The figure's after it box no longer says the editions are matched by name only.

2026-10-28 (the theatre measure gets a sheet)

• Amateur-theatre measure, a new How it works sheet. The measure had grown across five decisions (D-137 → D-145) with no page of its own. The sheet sets out what the database gives, the six steps from chronicle text to the 1910 area tables, what certain and possible mean record by record, every ruling, the panel's columns, the before/after counts of D-145 (the ensembles' own pages) and what is still open.
• Three interactive figures on it, read from what the theatre runner writes (data/interim/theatre/) and absent until it has run:
• Czech theatre over time: places, companies or performances per year, 1780–1913, the certain line inside the possible band, with 1848 and 1867 marked.
• Where Czech theatre was: every 1910 court district shaded by the share of its settlements with Czech theatre, with a year slider and play, certain or possible.
• From records to the panel: for a few places picked by rule, each record's certain and possible years and what the place gets in the panel. Facts only; the chronicle's prose stays in data/.
• The sheet opens in plain terms: what the database tells us about each company (when, what, where) and what we make of it, before the technical account. Where it stands has the build on all 7,103 ensemble pages (2026-10-29) beside the 682-page one, and Open flags how one-off events take an open record's possible years.

2026-10-25, late night (the page loads in a quarter of its weight)

• The 117 script sheets load when opened. They were 32 of the page's 41 MB — mostly the line-numbered source — and every visit parsed all of them to show one. Each is now its own file under scripts/, fetched the first time it is opened: the page is 8.7 MB. A link to a function's anchor still lands (the page knows which sheet holds it), and the search box still searches the source: their text is fetched once, on the first search. Opened straight from disk rather than through the site or a local server, a script sheet says why it cannot load.
• Mähren and Schlesien have their own sealed pages (D-133) — the decision and the draw are on the Decisions sheet; the roadmap says what the German route still cannot see.

2026-10-25, night (is the sample large enough; the German route anchored on the Czech)

• Is the certification sample large enough? The certification section now answers it, with tables built from the score file rather than typed: for the headline, yes — even counting claims as independent, the error per printed claim is bounded at about 2 % (4 of 505); for each symbol, no — only the thirteen symbols with ten or more instances are bounded at all, and a symbol needs about 36 instances read without a miss for its sensitivity's lower bound to reach 0.90. That is the design target for the entry-level sample.
• The route chart reads every run on disk. Each edition × page set now shows its own run, so the Certification pill shows the sealed Czech run's own counts instead of going blank.
• Step 7 of the route changed (D-129). The German edition of Bohemia is no longer kept apart: every German claim the certified Czech inventory disputes goes to Opus, on the German print alone. Replayed on the gold, all 19 claims only the Czech side held were on the German gold. With D-130 to D-132 — the offices section shown for every part, the place types routed, and a German heading that fills its line found — this is what the German gold re-run will measure.

2026-10-25, last (two guards against a key that is only a label)

• A run says when a section's heading was missed (D-128). An offices section whose heading the page image does not find runs on under the label of the section before it, and a claim reaches the inventory only from a crop named offices — so those entries are not misfiled, they are dropped. Every court with osady on a page and no offices section beside them is now flagged next to the claim count. Grouped by the printed slug the check cried wolf twice; grouped by the 1910 court it flags one thing over the three runs on disk, and it is the real one.
• The inventory carries a key, not just a label. entry is page/court/municipality/settlement, with the printed district kept beside it as what it is. The slug the text layer gives a heading is damaged often enough to be useless as a join — mechanice for Nechanice, novaraka for Nová Paka — and keyed on it the first scoring of the sealed half lost 170 of 505 claims that had all been read.

2026-10-25, later (the Czech half is certified)

• The sealed Czech pages are read and scored — once (D-127). Eight pages nothing was tuned on, read on the gold re-run's frozen route: F1 0.996, a 95 % page-clustered interval of 0.991 to 1.000 — 502 claims right, 1 false, 3 missed of 505; 104 of 107 entries reproduced exactly; the printed count right on 501 of 502; every token with ten or more instances read at sensitivity 1.000. Eight pages are eight clusters, so the interval is the result. The scorer's metrics were written down before any sealed page was scored, and it refuses to score an edition twice.
• Two defects found on the way, both fixed. The freeze check digested the contrast sheet's file rather than its pixels, so a re-encode of a sheet that was pixel for pixel identical reported the certification run as a different route (D-126); and the scorer keyed an entry on the printed district slug, which the text layer damages — mechanice for Nechanice, novaraka for Nová Paka — so its first run lost 170 of 505 claims that had all been read. The key is the 1910 court now.
• The German number depends on which entries you count. The same reading scores 0.945 over the entries the gold and every reader reached and 0.905 over every entry the sheets print. The gap is one page: on scan 0380 the page image misses the heading Ämter, Anstalten … Reichenau an der Kněžna, so its section runs on as parts and its 28 institutions are dropped rather than misfiled.

2026-10-25 (the route, drawn)

• The reading route opens with a flow chart of itself. The eight steps down the spine, in the runner's own order, and beside them the four ways a mark can be settled — accepted because all four readings agree, the reader of record's alone, settled by Opus, or ruled by the author — each with the share that actually took it. Click a step for what it does, why it is there and the decisions that settled it; the pills switch the edition and the page set. On the Czech gold the author's queue is empty, which the chart shows rather than asserts.
• Its numbers are the run's own. The stages come from run_frozen.py and every count from that run's final_inventory.csv, whose rule column says how each mark was settled; a rule the figure has no band for is counted and named rather than dropped. Switching to the sealed half or the whole volume blanks every number instead of carrying the gold's under another label — which is the page's own argument about what certification measures, made visible.

2026-10-24, later (the court of every section, and a notebook that runs itself)

Decisions D-122 to D-125 and the Roadmap carry the record.

• Every section of the register now carries its 1910 court (D-122). The name a heading prints is matched against the court list of its own land, the register's own names for courts, the courts the page's running head prints, the courts of the okres it names, and last what the readers wrote when they read the heading. Of the Czech volume's 1,248 headings 1,247 now resolve, and three names the old rule matched to the wrong court are now right: on the Broumov pages Teplice is Teplice nad Metují, on the Děčín pages Benešov is Benešov nad Ploučnicí, on the Krumlov pages Planá is Horní Planá. Two courts the register prints were created after the census and are filed under the 1910 court each was cut from. The two editions are paired by court.
• Notebook 15 runs unattended (D-123). It waits for its batches at each collect cell, will not finish a pass while a batch is still processing, sends a failed Opus call once more, and writes its batch queues outside Dropbox — about 150 GB for the four volumes otherwise.
• Gemini can be batched at half price (D-124), through Google's native batch interface with exactly the live request, once a check on the machine of record shows its readings are the live reader's. Identical replies cannot be the test: the model ignores its temperature, so the batch is compared with a fresh live read of the same crops made at the same time.
• A certification or full run says whether it used the gold's frozen route (D-125), comparing the code module by module, the frozen components, the packages and the machine.
• The register's pages are reorganised. Digitising the 1910 register keeps the document itself. The reading route replaces the accuracy-and-plan sheet and the gold-transcription sheet with one page: the route step by step and why, what was tried and dropped, the gold's results, the certification to come, how notebook 15 runs, and the gold — with the dated plan folded at its foot. Links to the two old sheets land on it.
• The certification plan: the sealed pages are read right after the gold re-run, on the frozen route; once the full run exists, a random sample of about 150 Czech municipalities, rare symbols oversampled, measures the analysis variables' own error rates, with its rule written down before the draw.

2026-10-24 (one notebook reads the register, and a third audit)

Decisions D-89 to D-121 and the Roadmap carry the record; the third audit is new on the Audits page. The accuracy-and-plan sheet still opens with its revision of 15 October, so the numbers below are not yet written up there.

• A second reader, and a judge (D-89, D-97 to D-102, D-110). Every page is now read by two model families, Gemini and Claude Haiku. They disagree on about one claim in eight, and Claude Opus settles those disputes: on the gold it settled 99 % of them and was right on 98 %, against 63 % for always taking Gemini (D-100). What it cannot tell goes to the author.
• A picture check of every mark (D-91, D-93 to D-95). A locator puts each claimed symbol on a box of the page image, and an image encoder (DINOv3) names the box from examples cut from other pages. Where the encoder and a reader disagree, the claim goes to Opus. It is never rewritten by rule.
• A contrast sheet for the look-alikes (D-99, D-102, D-103, D-106). The pairs of symbols the readers swap are shown to them side by side. It more than halved Haiku's swaps. For Gemini it did not, so Gemini reads without it.
• No local model joins (D-105, D-113). Six vision models that run on the project's own machine were tried as a third reader. The ones that reached the pair's errors raised alarms on half the claims, and the accurate one shares the pair's blind spots.
• The table and the German edition are read too (D-90, D-92, D-96, D-104). On the gold tables the model reads 99.7 % of the number cells, against 91.4 % for the scan's text layer. The German edition of Bohemia has its own gold and reader, and that gold also serves the Moravian and Silesian volumes, with a spot check in each.
• One notebook runs it all (D-107 to D-117). Notebook 15 freezes every setting, reads each page once with Gemini and three times with Haiku, checks every mark, compares the readers and the two editions, asks Opus each question once, and writes the final inventory. Every mark keeps a history of how each step read it (D-109). The gold, the sealed certification pages and the whole volume run through the same cells.
• The first gold run (D-118). The Czech edition comes out at F1 0.995, for about $0.77 of Opus. The German edition fell to 0.899, because the box-finder learned on Czech pages misplaced German boxes. Three fixes followed, and the German half is read again.
• A third audit (D-119 to D-121). All code now lives under utils/, the register's in utils/lexikon_1910/, and reference/ holds data only. The audit fixed the route's defects (the Czech gold replayed at 0.999: no false claim, two missed), made the printed count the readers' majority (875 of 879 right), runs the Czech–German independence test on the gold, and adds four sealed German certification pages, drawn by rule. Its last two sections say what the output still lacks for place-level variables, and why the certification as planned is not yet enough.

2026-10-15 (the page image decides where the crops go)

Decisions D-83 to D-88, the Roadmap, Validation, Place matching, the 1910 volume note and the accuracy-and-plan sheet follow.

• The notes are cut on the page image (D-83). The reader's images now come from the printed page itself, not from where the scan's text layer thinks the lines are. Each image is still labelled with its section heading; where only the image sees a heading, that one line is read by the model before the section is labelled.
• The postal and railway legend is read from the image too, so the table's post and station numbers can be turned into place names with the reader's spelling.
• Two pages ended the table in the wrong place (D-84), and read a heading of the notes as table rows. Fixed.
• Every reading now checks itself (D-85): how many lines the image shows against how many the reader returned, and whether the reader saw a different heading.
• A municipality split across two pages is checked as one (D-86).
• The district from the page head (D-87): a district name that cannot be matched is now reported as unknown rather than guessed from the page before; Prague and Liberec are read from their heads.
• Loket and Kralupy nad Vltavou (D-88). The register prints both as districts of their own; the district map the project is built on has them inside Falknov and Slaný. Both are right. The register counts people as at the end of 1910 but draws districts as at the end of 1914, and its German edition dates both new districts to November and December 1913 (the Kralupy court to March 1912). So the register's pages now resolve to the 1910 district that held them, and the 1910 district map stays as it is — it is the map of the census, and of every legionnaire's birth. The two districts belong to a later map, for the interwar elections.

2026-10-14 (everything printed is now inside some image)

The accuracy-and-plan sheet opens with a new revision, and Decisions (D-76 to D-82), the Roadmap and Validation follow it.

• 22 pages had never been read (D-76). On those pages the scan's text layer put the end of the table inside the notes, so the notes were never found and no image of them was cut. They are now: 470 entries that no reader had seen.
• Every printed line of the notes is now inside an image. The images of one column now meet edge to edge, so a line the text layer dropped between two sections can no longer fall into a gap. Table rows that used to be sent as notes are not any more.
• The table reads its first and last rows (D-77, D-78). Its top and bottom are taken from the printed rules: 233 rows at the top and 137 at the bottom that had never been read, and none of the gold tables changes.
• The district is read from the page head despite garbled letters (D-79). It changes on 24 pages, and on 19 of them the old answer was wrong.
• The pipeline can describe any volume (D-80). One profile per volume, and the page layout read from the image alone; tried on the German Bohemian, Moravian and Silesian volumes. It is not switched on yet.
• Prague and Liberec (D-82). Their pages are written as running text; the text is now extracted, and a vision read is prepared.
• What is open. Seven section headings the text layer missed altogether wait for the author's ruling, and the stored readings are read again (r5, about $3) once the transcription is done (D-81).

2026-10-13 (a crop that stopped one line short)

The accuracy-and-plan sheet opens with a second revision of the day, and Decisions, the Roadmap and Validation follow it.

• What was wrong (D-75). The notes under each table are cut into small images, one per section, for the reader and for the person transcribing. The last image of a column ended at the last line the scan's text layer had recognised. Where that layer had dropped the final line, the line was on neither image, so the reader and the transcriber both missed it and agreed.
• How often. Counted over the whole Czech volume without asking a model: 13 of 454 pages had print below the old edge — twelve entries and the last line of two more. Every one of the twelve is a number and the same single symbol, the shortest line the register prints. The four pages the reader is scored on lost nothing.
• The fix. The last image of a column now runs to the foot of the column. The two affected transcription images were cut again; no transcription was changed, and three entries on them are still to be read by the author.
• What it costs. Stored readings are matched by the exact image, so six of the 81 images of a stored pass have to be read again before it can be replayed. Not done yet.

2026-10-13 (the order the register prints its symbols in, as a check)

The accuracy-and-plan sheet opens with a new revision, and Decisions, the Roadmap and Validation follow it.

• The symbols of an entry stand in a fixed order, and that is now a check (D-73). The volume's own preface reads one entry out in order — offices, churches, schools, doctor, midwife, fire brigade, savings bank, cooperative, post, railway, and a castle last. On the ten transcribed pages the order has no exception once the gendarme, the canteen and the transport station are set aside, which the register prints in two positions. The check learns the order from those pages and asks no model.
• What it finds. A symbol read as another that belongs elsewhere on the line: above all a castle read as a poorhouse, the reader's commonest confusion, where the castle stands last. It flags about one of the ten or so wrong symbols in a pass and almost never a right one. A wrong symbol standing where the right one would — a railway stop for a station — it cannot see.
• It found slips in the gold. Two entries of the transcription broke the order, and both were slips of the sheet, corrected by the author against the print.
• Where it goes in the reading (D-74), designed and not built. The readers are not told the order, so that it stays evidence about them. It is applied after the read and before the two readers are compared; a symbol it flags is looked at again, on the image, and none is changed by rule.

2026-10-13 (the amateur-theatre database — scraped, tabulated, parked)

A settlement-level source for pre-1914 associational life is on file, and the Roadmap says where it stands.

• What it is. The Databáze českého amatérského divadla lists, place by place, the amateur ensembles of the Czech lands with their years. All 7,368 place entries were fetched on 2026-10-12 and parsed into two tables; 93.6 % of the places tie to a 1910 settlement. Its entry under Data sources (#22) carries the citation and the rights, and it is in the References.
• What it showed. Bohemia and Moravia are covered alike once language and size are held fixed — 30.0 % against 28.2 % of ordinary communes in Czech-speaking okresy have an ensemble starting before 1914. But it records Czech theatre only, coverage rises steeply with the size of the place, and each okres was compiled from its own sources.
• Where it stands. Parked, to be picked up later; nothing is joined to the person files. The chronicle, bibliography and archive citations of every place are kept on disk for that later reading.
• Two new script sheets: the scraper (scraper_amaterskedivadlo) and the coverage tabulation (amateur_theatre).

2026-10-08 (typeface — Libertinus)

• Libertinus throughout: running text, tables and sidebar in Libertinus Sans, headings in Libertinus Serif Bold, matching the paper. The faces are embedded (subset to Latin/Czech/Cyrillic, utils/docs/fonts/, build_fonts.py), so the page no longer depends on which fonts the reader has installed.

2026-10-08 (a second reader that catches nine errors in ten, and the route for a disputed claim)

The accuracy-and-plan sheet opens with a new revision, and Decisions and the Roadmap follow it.

• The second reader has been screened (D-72). Claude Haiku 5.5, read beside Gemini on the same crops, disagrees with it on 87 % of the errors Gemini makes — a second Gemini pass finds 30 %. It also disagrees where Gemini was right, three times as often as the bar allowed, so about one claim in nine ends up disputed. The figures rest on four pages and are to be re-scored on twelve.
• What happens to a disputed claim is laid out, and none of it is built. Both readers read the whole volume; an image encoder names every symbol as a third, different kind of check; what the three do not agree on goes to one call on a costly model and then to the author. Which model is called on what is fixed in code. Re-reading the disputed sections with local models was weighed and set aside, and the sheet says why.
• Two cheap tests come first, and either can stop its half of the design: whether the encoder can tell an outlined symbol from a filled one at all, and whether the costly model settles disputes better than always taking Gemini's word, which is right 57 % of the time.

2026-10-06 (the register's reader at 0.972, and what a second reader would have to add)

The accuracy-and-plan sheet now carries a day's worth of measurements, and its plan says what is done, what is next and in what order.

• The baseline with every fix in: 0.972. Six fresh passes over the gold-sweep pages. Nothing is filed under the wrong part of a municipality any more; what is left is a handful of look-alike symbols the reader gets wrong in every repeat — a castle for a poorhouse, a railway station for a stop — each checked against the print by the author.
• What stands beside a symbol is scored (D-70). A school's count, classes and language and a cooperative's kinds and language are read right on all 116 cases in five of the six passes (the sixth skipped one printed line); the named parts of Části osad are all found with their houses and population exact, and their place-type symbols are the weak column, at 0.92–0.93. Four parsing defects that had passed for reader errors are fixed, and two slips in the hand-read sheets — found by this scoring — are corrected.
• The bar for a second reader (D-71). Reading the same pages again with Gemini finds only 26–30 % of its errors: seven in ten are made by both versions, every time. So a second reader will be chosen on the errors it brings to review, not on its own score. The screen is written and starts with two readers that need no new account.
• The order of work puts transcription first. Every pass already reads eight calibration pages that have no gold yet; with it, each comparison is made on twelve pages instead of four. The stage table's last column is kept current from now on.

2026-10-05 (the register's reader at 0.958, and the plan it had been missing)

The work on the 1910 register now has its own sheets. Until today the page carried the description of the volume and nothing about how well it is being read, though the Decisions register has pointed at that argument since August — and none of those links could be followed.

• 1910 register — accuracy and plan opens with where the reading stands and how the pipeline changes next, stage by stage; the dated record of how it got here follows. The default reader is at settlement F1 0.958 on the development gold, every symbol a regression will use is found each time it is printed, and what is still wrong is rare ([castle], [ruin]).
• 1910 register — gold transcription says where every hand-read sheet lives and what is queued, the German edition included.
• Seven decisions, D-63 to D-69. The scorer now works at the unit the analysis uses; a part of a municipality that opens its own printed line is filed where it belongs; the scored gold is the whole-page sheets; a damaged section heading no longer files a whole section under its neighbour; the German edition gets its own gold; every paid pass runs on one machine; and an image encoder becomes the second method for the rare symbols.
• A price is no longer set as a formula. Two dollar amounts on one line were read as the two ends of an inline formula, and the words between them came out as a run of italic letters. It affected the register's row on the Roadmap, three places on the Decisions sheet and one entry below; it would have affected 26 places on the new accuracy sheet. A closing $ followed by a digit is now left alone, and all 86 real formulas on the Variables sheet still render.

2026-09-23 (the second audit, and a smaller, truer overlap)

The whole cleaning pipeline was read line by line a second time, two months after the first audit, and every finding was measured on the data before it was fixed. The Audits sheet — renamed from Audit — 2026-07-16, because it now holds both — records 8 defects that changed an analysis variable, 25 narrower ones and the documentation drift, each with its before-and-after number. Notebooks 02 and 03 were re-run on the fixed code, so every figure, table and number on this site is from the new build.

• The cross-dataset overlap is 1,340 men, not 2,239. The name test scored surname and given name as one string, so a shared František could carry two different surnames over the line, and 111 fallen soldiers were each matched to two to four legionnaires. The second criterion read a note field that mostly says which interwar regiment keeps the death register, not that the man served in the Legion. Each part of the name must now pass on its own, the match is one-to-one, and the note must name the Legion (D-59). This is the change a reader of the comparison design needs to know about.
• 214 legionnaires born in Galicia had been placed in a Slovak district. Halič is Galicia in Czech and a Novohrad town in Slovak; the town's census covariates went to men born in Lvov and Tarnopol. The seat is no longer resolved by name.
• The exit date — the hazard model's event time — is range-checked now. It ran from 1818 to 2192. Dates known only to the month or year are kept and say so in a precision column instead of being dropped (D-61); 3,827 more records have an exit date.
• US applicants are imputed to France, not Britain (1,659 of 1,685 with a known unit served there), and the Polish Legion no longer lands in the Balkans (D-60).
• Missing stays missing, in four more places. A blank cause of death is no longer "not killed in action" for 8,297 fallen soldiers (D-62).
• The QA gate fails a stale build. data/final had been four days older than the code, and eight checks written for the newer columns skipped themselves silently. It now runs 107 checks, all passing, and the first one asks whether the build is current.
• The Bolshevik coup is on 7 November 1917 on the three timing figures, not 7 September.
• On the reference sheets: Pipeline, Variables, Place matching, Occupations, the Validation register, External validation, Decisions (D-59–D-62) and the Roadmap, which gains a list of what the audit deliberately left open.

2026-09-17 / 09-22 (the reader got good, and it was the view, not the model)

The 1910 register's symbols now have a reader worth keeping. Until September the apparatus below the table — the pictograms that record which municipality had a school, a post office, a gendarmerie post — was read by local models at F1 ≈ 0.70, with 30 % of municipalities exactly right. A cloud reader and one controlled experiment later, the Digitising the 1910 register sheet, the Roadmap, the Decisions (D-54–D-58) and the Validation register say something different:

• Gemini Flash, one apparatus section per call, thinking off: F1 0.962, entry_exact 0.80, $0.33 for the 13-page gold sweep — ≈ $6 for the whole Czech apparatus on batch. Experiment E1 held the crops, the prompt and every flag fixed and changed one thing at a time over 17 passes. With thinking off on both sides, section crops beat the whole page by +0.132 F1 and +0.295 entry_exact; the model version makes no difference; and Gemini's default thinking bought 3.7 nothing and cost 3.8 accuracy.
• The cost ledger had been counting a sixth of the bill. Through the OpenAI-compatible endpoint Gemini reports its answer as completion_tokens and bills its thinking as output without reporting it — ~5,000–5,700 tokens a crop on 3.8. A pass logged at $0.32 really cost ≈ $1.95. The reader now books the hidden share, and thinking is off.
• Claude Opus 5.5 was priced and not pursued: ≈ $160–340 on batch for the full pipeline, against ≈ $6 for the Czech apparatus on Gemini, for a reader no one has measured on these crops. The second reader the German edition needs (D-58) will come from a cheap screen of Qwen3-VL and MiniMax M3 on the same crops instead.
• Two silent defects, found by scoring at the right unit, and fixed (2026-09-23). Scored at the settlement level the analysis will use, the reader had fallen to ~0.85 — not a reading error but a decoder defect: the — 2) marker that opens a municipality's second part was never split, so 12.8 % of correctly read claims landed on part 1 (D-56). Fixed, settlement F1 is 0.915. Checking the fix exposed a second one: the reader joined a tall section's strips in the order they finished rather than the order they were cut, so the same cached answers scored differently live and on replay. Fixed too, and two replays of every E1 pass are now byte-identical.
• What still fails, named. About 0.05 of part attribution is still wrong and not yet diagnosed, and an outline-roofed [castle] is still read as the solid-roofed [poorhouse] by every model version.
• Nothing is certified. Every figure is on four purposive pages read once. The gold is now the binding constraint, and D-57 splits it by purpose so that the symbol validation sample can be read in hours rather than the 48–74 h the whole-page protocol costs.

2026-09-07 (the education ladder finally has a source that could contradict it)

Education was the one major variable with no external check at all. It appeared in the validation register only under "What none of them reach". The ladder was built on two modern histories — Moore (2020) and Lassnigg (2015) — neither of which counts schools, and one of whose figures begin in 1924 in a different country. §5 of the education page had named the missing source by class: "the Austrian statistical administration's own school taxonomy … which classifies schools rather than describing them."

That volume is now held and read: Österreichische Statistik XCIII/1, *Statistik der Unterrichtsanstalten … für das Jahr 1908/1909* (Wien 1912), the k.k. Statistische Zentralkommission's own annual census of every school in Cisleithania — 279 pages, merged from 15 scan parts, with a new External validation — education sheet carrying five tests.

• It confirms the hardest call in the variable. The pokračovací continuation school is kept off the ordinal, and the administration files it exactly so: its own category in both the industrial (p. XX d) and commercial (p. XVI) branches. It was also the mass institution — Bohemia had 504 of them teaching 50,836 pupils against nine state trade schools teaching 3,518, i.e. 84.0 % of all vocational pupils. The pre-rewrite reading, which scored it level with the Gymnasium, would have put the largest and least selective school in the empire at the top of the ladder.
• It breaks education_complete (D-53, deliberately not patched). The column scores a bare elementary schooling complete at eight years, on the law's eight-year obligation. Table 34 counts elementary schools by how many year-groups they actually ran: an eight-class Volksschule existed 10 times in all Cisleithania and twice in the Czech lands, against 1,114 five-class schools in Bohemia alone. So 3,894 records are marked complete against a form that barely existed and 3,137 reading the ordinary full ob. 5 are marked incomplete. The law's years and the school's classes are different units and the column conflates them. Fixing it changes an emitted column and the choice is a judgement, so it is logged with two candidate resolutions rather than quietly patched.
• It shows pok. is badly under-recorded. The Czech lands had 5.3 continuation pupils per full-time vocational pupil; the records name them at 0.86 to one. Men did not write down the evening classes they were obliged to attend, so education_apprentice is a lower bound, not a rate — a caveat the paper needs and nothing internal could have found.
• It cannot touch Hungary, being Cisleithanian by definition, so D-08 stays open for exactly the reason it already was, and the ~10,500 Slovakia-born records sit outside every test.

Two errors in the printed volume turned up on the way, each decided by a second arithmetic path rather than assumed: a roll-up in Table 20 that is one short of its own parts, and a prose figure on p. XXXII (22.341 schools) that contradicts both Table 34 and the volume's own summary table (22.241). The transcription follows the page and declares the errata.

And the QA gate went from 66 to 72. Education had three invariants, all of them behind a guard on a column the current build does not have — so in practice it had none. Six now run unconditionally, including the one that would catch the old defect coming back: a continuation school may act as a floor on the ladder, never as a promotion.

2026-09-07 (two variables that answered questions nobody asked them)

Both variables rewritten in the last fortnight were re-run against the raw field, and both had a defect the rewrite had not touched. Neither was visible in the finished data: the QA gate passed 66/66 with both present, because both are about what the parser did with a string, not about what the frame ended up containing.

• An unrecorded affiliation was False, and now it is <NA> (D-52). Politická příslušnost is blank on 93,377 of 132,993 records — 70.2 % — and every pol_* dummy was writing False there. That is not "not a Sokol"; it is a source that said nothing, and the difference is the same one that fabricated exit_N on 26.4k empty cells in the 2026-07-16 audit and that provenance.py already refuses for its own flags. pol_recorded (39,616 True) marks the base; a record that is filled in but matches no token keeps False, because there the register genuinely failed to recognise something. It matters now rather than later: occupation missingness was found on 2026-08-27 to track the outcome, so a silently-false covariate on 70 % of the sample is not a cosmetic problem.
• A tie between two schools was being broken by word order (D-51). 1,512 education records name two institutions on the same rung — reálka + obch.akademie, gym. + uč.ústav, odb. + hosp. — and the parser took whichever the clerk wrote first. The same pair came out both ways: obch. + odb. gave commercial on 20 records and trade on 164. It is not a labelling question. It sets education_track (academic vs teacher vs military) and education_complete, whose standard course lengths differ across the tied schools — a Realschule is seven years, a Lehrerbildungsanstalt four, so 152 records were being graded against the wrong yardstick. The order is now written down, on two principles: a generic word (odb. = "specialised", střední = "secondary") yields to the school the record actually names, and the terminal professional school outranks the general school that feeds it. The second rests on reading the cell as a chronological itinerary — which the data supports, on 92.4 % of multi-rung records the ordinals never fall along the string.
• Semesters were being counted as years. technika 4 sem. is four semesters at the Technische Hochschule; the bare digit attached backwards to the institution and landed in education_years_secondary on 722 records. Universities counted in semesters and the Mittelschulen in years, and the two were sharing a column. They now have their own.
• The scan is word-anchored, and that found three abbreviations. technik had been matching inside hydrotechnika and ob inside pob. — the unanchored-substring shape that ranks 1.4.0 had to remove after Fidler caught generálmajor scoring as major. Anchoring it exposed spellings the patterns had never covered: univ. (55 records, where only univer… was matched), the lék.fak. / fil.fak. faculty form, and vyš.prům. (37 records, which had been falling to the 3-year Fachschule instead of the 4-year Staatsgewerbeschule). Coverage 99.07 % → 99.10 %.
• One thing that was wrong in the plan, and is recorded as wrong. The review that opened this work claimed the technick… pattern was over-promoting secondary technical schools to Tertiary. It was not: technik and technick diverge at the sixth letter, so the pattern never matched technická at all. The real defect there was the opposite — eight records that parsed to nothing. Checked before it was fixed, and the fix is the one the data called for.
• Three modules now carry a self-test the QA gate runs, alongside the data invariants: provenance_selftest(), education_selftest() and affiliation_selftest(). They hold the rules a finished frame cannot express — that reálka obch.akademie and obch.akademie reálka give the same answer, that the refusal token sl does not match inside slov.líga. A first attempt at an invariant here failed honestly: "semesters never equal secondary years" looked right and would have broken the gate on 13 records that legitimately state the same number twice (ob. 5 gym. 8 univ. 8 sem.). A rule about the parser belongs in a test of the parser.

⚠ data/final is one build behind this entry. The modules, the schema and the gate are at 2.1.0 / 1.12.0 / 1.10.0; the four new invariants are dormant until notebook 02 is re-run, and the affiliation figure in notebook 03 will change base — it must report on the 39,616 recorded records, not on 132,993.

2026-09-03 (the party family that a source dissolved)

The political-party genealogy cited Wikipedia, and it carried an ideological axis in every figure. affiliations.py folded mladočeši, realisté, stát.pokr. and lid.pokr. into nar_dem on the strength of four Wikipedia articles, and kept a separate pokr bin labelled Liberal. A photograph of one atlas plate settled it from the print — and overturned the bin.

• New sheet: Party register. Akademický atlas českých dějin, plate IV.37: 55 parties of the Czech lands 1861–1918 in five blocs, each with its founding year, area of operation, and a notes column recording every merger, renaming and dissolution. Transcribed to reference/parties/atlas_iv37_parties.csv.
• pokr is not a party family. The register has Č13, Č15 and Č17 each as "1918 spoluzakladatelka Č 23", Č23 as "1919 přejmenována na Národní demokracii", and Č7 merged into Č15 in 1908. Masaryk's realists are the same line — Č12, "realisté, 1905 transformace v Č 13". So every pokroková party in this data ends in the National Democracy, and a Liberal group standing apart from it was an artefact of the reading. The canonical is abolished. pol_pokr (869) disappears: 663 already carried nar_dem and only lose a spurious second flag, 78 newly gain it, 42 become volna_myslenka — Free Thought being a freethinker association, not a party, and absent from the register.
• It confirmed two calls, on better evidence than they had. lid.pokr. → nar_dem (Č17 is a co-founder of Č23) and real. → nar_dem (Č12 → Č13 → Č23) were both taken from Wikipedia; the plate says the same thing from the source.
• Three refusals, printed rather than buried. A bare lidová — eight register parties carry it, Catholic and liberal-national and Masarykite alike, so the module having no such token was right all along and is now a documented refusal. A bare demokrat. And SL, 108 records and the largest unmatched token, together with every Slovak formation: plate IV.37 is v českých zemích and contains no Slovak party at all. A source gap, on the roadmap.
• A transcription that defends itself. The plate was read from a hand photograph of a curved page, where the notes column drifts against its rows — the compensating-error shape the validation register already warns about for the 1910 register. But the notes cross-reference by code, so check_atlas_parties.py can require that Č7's "merged with Č10 into Č15" faces Č10's "merged with Č7 into Č15". 12/12 invariants hold; a one-row offset would not have survived them.
• Also dropped: rolník → agr. It mapped an occupation into a party dummy on the theory that a self-described farmer is an Agrarian, while dělník in the same field mapped to nothing — an asymmetry with no source behind it.

2026-09-02 / 09-03 (the school that was never a school)

Vzdělání was being read as a level, and it is an itinerary. ob. 5 mě. 3 pok. 3 names three schools and the years spent at each; the parser took the first token it recognised and threw the rest away. The worst consequence was one token. pok. — pokračovací, the gewerbliche Fortbildungsschule — is the apprentices' part-time continuation school, attended a few hours a week alongside a trade. It scored Upper secondary, which put 8,117 records — the classic apprentice itinerary — in the same bucket as the Gymnasium. The variable was flagged as deferred (D-08, audit B3) on the suspicion; two sources arrived this week and turned the suspicion into a number.

• The ladder now has a source. Moore (2020), Teaching the Empire, gives the general half: compulsory schooling ages 6–14 under the 1869 Reichsvolksschulgesetz, reached as Volksschule 5 + Bürgerschule 3, with Gymnasium (8 years) and Realschule (7) entered at age ten in parallel with the Bürgerschule rather than after it, and the Lehrerbildungsanstalt entered from the Bürgerschule at fifteen. So ob. 5 mě. 3 is the legal minimum, not an achievement; ob. 8 is the same eight years taken in one building; and uč.ústav was sitting a rung too high.
• The vocational branch has one too. Lassnigg (2015, IHS WP 108) reaches the ~17 % of the field Moore does not: the Fortbildungsschule was legally part of compulsory mass schooling, governed communally alongside the Volksschule while the full-time schools ran from the centre, and it selected on nothing — an apprenticeship required no educational credential at all, and the school obligation followed the contract, where obch., prům. and odb. required Bürgerschule grades plus an entrance exam. That is what settles it: pok. is not a rung.
• New sheet: Education ladder. What each school was, who entered it from where, and the two things neither source can give us — the duration of a Fortbildungsschule year, and anything at all about Hungary, whose schools the Ausgleich handed to Budapest. The second is why D-08 stays open: the ordinal is not established for Slovak and Rusyn births.
• education.py 2.0.0 parses instead of classifying. Six rungs — No schooling < Elementary < Burgher school < Vocational secondary < Upper secondary < Tertiary — read off a tokenised itinerary, with year numbers attaching backwards to the school they follow (ob. 5) and forwards when the cell opens with one (8 ob., 3mě), which is how the dot-less variants that used to fall through now parse. Eleven columns replace two. Coverage of non-empty strings 96.7 % → 99.1 %.
• Two refusals worth naming. The ordinal is attendance, not attainment — a bare reálka does not say whether the man finished, so completion lives in its own column that is allowed to be missing, and is missing for the whole vocational branch because no source establishes those course lengths. And the year columns are not summed into a total, for the same reason: continuation-school hours varied by Land and trade under regional laws passed between 1907 and 1931.
• What moved. The old Upper secondary bucket of 18,756 redistributes — 7,579 stay, 3,973 to Vocational secondary, 2,606 to Burgher school, 4,598 to Elementary. 8,815 records (6.6 %) now carry education_apprentice, which never raises the ordinal. Notebook 02 re-run, QA gate 66/66.
• One label changed after the first run. Rung 0 was called None, which in any value_counts is indistinguishable from a missing value — the exact shape of the 2026-07-16 fabricated-exit_N bug. It is No schooling now.
• The chart's colours now carry the order. Notebook 03 was re-run on the new ladder, and at six levels the categorical palette misread: running dark-navy → steel-blue → bright red → dark red → light grey → dark grey, visual weight peaked in the middle of an ordered variable and the top rung whispered. Education now takes CU_SEQUENTIAL, the ramp the style module already documents for one-directional variables — light at the bottom rung, navy at the top, sampled from 0.22 rather than 0.0 so the lowest rung is not white on white. 28 figures and 14 tables rebuilt; the occupation figures, which had predated D-41 since 2026-08-02, are current again as a side effect.
• The ladder sheet shows the ladder. Education ladder now carries the chart itself, not just the argument. It is globbed rather than named, so a second education figure appears there without a code change — the contract the Descriptives sheet already offers.

2026-08-26 (the review flag becomes a question that can be answered)

172 keys of the occupation dictionary carried a review flag, and nothing in the project could clear one. A key looked identical whether nobody had examined it or somebody had examined it and concluded the code was right, so every rebuild of the conflict workbook asked the same 172 questions and a day of adjudication left no trace in the repository. The 2026-08-26 pass (D-41) answered 171 of 181, the answers are now in the code rather than in a spreadsheet, and the flag falls to 10 keys / 163 records. It means unsettled now, not unread.

• The ballot came back as a register. czech_glosses.ADJUDICATED records one verdict per word — keep (130), recoded (41), open (10) — and run_link_bridge.py 1.2.0 applies it last and is the only writer of review, so the German bridge can go on flagging genuinely new ambiguity without re-raising anything answered. The new code always lives in HISCO_OVERRIDE and never in the register, so a reader looking a word up finds one answer; the bridge refuses to run if a recoded key has no override entry, or if an override still claims the flag of an adjudicated key. Both would leave the curation looking settled at the old code. The run rewrote 41 codes, wrote a verdict column on all 181 keys, and took the flag from 172 keys to 10.
• What it moves in the data. 4,731 legionnaire records take a different HISCO code (203 more in the interwar column), and 1,383 move HISCLASS. Nearly all of the class movement is one word: slévač and its spellings, 1,028 records from 7 to 9, recoded 72500 Metal Moulder → 72420 Metal Pourer on OccCANINE at p=0.99. Behind it, 172 estate and forest officials go 3 → 5 (Office Clerk, which is what an estate official actually did), 57 továrník go 3 → 1, 51 ošetřovatel 9 → 4, 28 farmaceut 4 → 2, 16 mlékař 4 → 9. Coverage is untouched: 98.46 % → 98.48 % of non-empty occupations coded.
• What it costs, priced rather than avoided. 219 records trade the class for the code — štukatér (106) → 95530 Stucco Plasterer, cementář (59) → 95240 Cement Finisher, rukavičkář (47) → 79475 Glove Cutter, živnostník (7) → the −1 sentinel. Each is a sharper five-digit code that HSN gives no HISCLASS at all. The sheet says so where the numbers are, because a class variable that quietly shrinks is worse than one that shrinks in writing.
• D-42: a stratification scale stopped ranking men whose occupation was never recorded. The exclusion was believed to be in force and was not. For a HISCO sentinel there is no occupation, so the status modifier alone carried the scales, and HSN reads a modifier as a social position where this project reads it as a refusal to code. 3,098 students — 19 % of that cell — were SOCPO 42, middle class, supervisors of skilled workers, and 160 men holding a school-leaving certificate were HISCLASS 2, higher professionals, and SOCPO 5, Elite. Nothing was fabricated: HSN observed those combinations 378 and 9 times, and they answer a different question. Withdrawing them takes HISCLASS 2 from 4,907 to 4,750 and the SOCPO elite from 3,013 to 2,730; HISCLASS coverage falls 95.8 % → 95.5 %. label_en and the major group stay, because they describe the code and do not rank a man. The defect survived because no invariant covered SOCPO at all; two now do, one per occupation field.
• Ten questions survived, and seven of them are the same question. čsd, zaměstnanec čsd, čsa, pošta and the three spellings of the ČSD workshops at Nymburk are an employer or a workplace written where a job was asked for — and the project has no reason to believe ČSD names a railwayman rather than a clerk, a fitter or a stoker in those workshops. They keep the -2 sentinel and the flag. The other three are real questions: vlakvedoucí (the guard, currently sharing a code with the engine driver), politik, and the bare správce. The workbook now opens on those ten instead of repeating itself — and on rebuilding it, all ten land on Nothing to choose from: no coder ever proposed a rival for any of them, so what is left is not a choice between two codes but a question about what the word means. The Decide and CZ-ISCO says otherwise sheets are empty for the first time, which took a fix of its own (build_conflict_workbook 2.1.1): a sheet with no rows was being written with no header either, and the workbook's own verifier caught it.
• Leaving those ten is itself a decision, and it is on the sheet as one (D-43). Not a backlog item: 156 of the 163 records are in the interwar occupation column, which codes 4,295 records in all, so the ten are 3.6 % of that variable and the ČSD/post-office family alone 3.1 %. Guessing "railwayman" would not round a 133,000-row sample, it would plant a fabricated occupational profile in the smallest occupational measure the project has — the one the interwar-electoral work runs on — and it would read as a finding rather than as missing data. Since D-42 the -2 sentinel is a place for unknown to live: a code, a label and a major group, and no class. The flag stays up until a source answers it — a ČSD staff schematism, the Nymburk workshop rolls — and then the whole family moves together or none of it does.
• A code with nothing behind it, caught on the way in. cementář → 95240 was about to be delivered with no English label and no major group: 32 of the authority's 1,696 codes were only ever observed under a modifier and have no base row, so the lookup missed its own key and the base fallback alike, and the QA gate would have read 59 records as a hole. occupations.py 1.3.1 takes the label and the major group from whichever row the code appears on — they never vary within a code — while every scale stays absent, which is the price D-41 accepted and not a rung more. Same principle as D-42: a description is not a ranking.
• Where to read it on the site. Occupations carries the adjudication and the ten open keys; Decisions gains D-41 and D-42; the Validation register's Open section is now an answered list with its reversals written out — mlékař conceded to CZ-ISCO, skladník kept against it, obchodní příručí kept after asking for the second opinion this workbook was.

2026-08-11 (a page of the register, with the boxes drawn on it)

New sheet: Digitising the 1910 register. Eighteen modules, a 20-page pilot and a hand-keyed transcription worksheet existed for the Podrobný seznam míst and none of it was written down anywhere the project could see — no row in Data sources, no number in Decisions, no entry in the Validation register. All four now exist.

• The figure is a photograph of page 256, in both editions, with the boxes over it. Twenty-two clickable regions, and not one of them placed by eye: every box is derived from the same ALTO geometry the parser uses — the printed column rules, the fitted grid, the block segmenter, the divider, the per-line bounds — so the overlay cannot drift from what the code actually reads. The panel quotes, verbatim, what the text layer made of whatever is under the box. utils/lexikon_1910/build_lexikon_figure.py renders it out of two PDFs (one of them 3.34 GB and gitignored) and writes a 1.5 MB tracked artifact; the docs build only reads that, so it still runs on a machine with nothing from library/.
• Brzice is the worked example, and it earns it. 541 hectares, 88 houses, 420 people over four osady — small enough to check the block sum by eye, and it holds a real defect: a stray inverted comma makes Žďár's male count ‘ 11, so the row comes back unparseable rather than failed, and the grader refuses to guess what the press meant. Under the table, in one printed line, is what the village had: a two-class school, a midwife, a fire brigade and one Raiffeisen credit union. That line is set out three ways — as printed, as the text layer read it (1, 1) g II, $, f, R., four marks and one slot), and as it actually is, drawn with the plate's own glyphs.
• The German volume is on the same figure, and it is not decoration. The identical figures under a different running head, on a different printed page, in a different position in the district — Brzice is municipality 1 in Czech and 3 in German, because each edition sorts an okres alphabetically in its own language. That is the whole argument for matching the editions by name and never by the printed number, and it is now something you can see rather than a sentence in a docstring.
• Both Výklad značek plates, and the 82 designs cut out of them, searchable. Which is also where the difficulty lives: the plate sets its symbols at display size and an entry's whole printed line is 44 px tall, so template matching from the plate peaks at 0.644 and the detections cluster with no separation at all. Across five reader passes, all 109 compared marks are contested and none is unanimous.
• Two decisions recorded that were load-bearing and unwritten — D-38, that a reader is never shown the invariants it will be graded on and a failing block is re-read rather than repaired; and D-39, reversing the assumption that the text layer knows where every marker is, on a measured threefold undercount.

2026-08-05 (the census place register is complete, and a long download is watchable)

All four 1910 Spezialortsrepertorium volumes are now on the shelf. Böhmen, Mähren and Schlesien in German, plus the Czech-language Podrobný seznam míst edition of Böhmen — the census's own register of every Ortschaft and Katastralgemeinde, which is the printed authority behind the settlement crosswalks. Two of them were half-downloaded and finished here: 122 pages from MSVK and 756 from the NDK.

• A second Kramerius generation, and a host that lies about which one it is. ndk.cz presents a modern viewer but answers on the v5 API, so it belongs to scraper_kramerius.py, not the v7 module written for MSVK. Both are now driven from one place, run_lexicon_downloads.py, which runs the two hosts concurrently and stacks a progress bar per volume — 878 pages is long enough that "is it still going?" is a real question. progress.py keeps the bar optional: without tqdm the downloaders fall back to their old log line and nothing else changes.
• --name had been renaming the volume, not the file. It overwrote the title as well as the filename stem, so the OCR text's own header announced itself as NDK_Specialni-lexikon-obci-… instead of Podrobný seznam míst zemí rakouských … IX. Čechy. The stem and the title are separate again.

2026-08-05 (the district map draws the unit you asked for)

The borders were being erased rather than drawn. Every polygon was stroked as if it were a district, and a second layer then over-printed the court-only lines in pale grey to take them back out. On a pale district that reads as a hairline; on a dark one it reads as white scratches, and at political okres — where a court line is not a boundary at all — the map was covered in them.

• Three border layers, one per class, drawn instead of undrawn. Each edge is classified exactly (adjacent MPIDR polygons share their border vertex for vertex, so this is arithmetic rather than tolerance) into court — both sides inside one okres, twin — the city and ring of a pooled pair, and okres — everything else. The polygons now carry fill only. At political okres the map draws only district boundaries; at soudní okres the court lines come in underneath them as pale hairlines.
• Boundaries are chained, not scattered. The okres boundary is ~20,000 individual segments; emitted raw, each gets its own line caps and a stroked border reads as a dotted scribble at any width above a hairline. _stitch walks them into 1,021 polylines that join cleanly — and the court layer got smaller in the process (28.4 kB → 16.8 kB).
• City twins are now a control, not a fixed choice. Ten Czech seats are a Statutarstadt plus the rural okres around it. They are two okresy and the records are split between them, but the court-level census cannot separate nine of the ten, so the analytical unit is the pooled pair (Decisions D-18). Separate draws 158 units; pooled draws 148 — the city border comes off and the arithmetic follows, in the shading, the quantile bins, the district panel and the rank. A merged region showing two shades with no line between them would have been worse than either answer.
• Pooling a mean needed its denominator. rank_reset is stored per okres as a mean, and averaging two means would have given Brno město — a fifth of the records — the same weight as the ring around it (−0.99 and −0.37 average to −0.68; the correct pooled value is −0.50). The mean now travels with its own n, and the rank is recomputed against the pooled totals rather than inherited from a member.
• What the toggle removes is real. In all ten pairs the city is the paler patch — a small light core inside a darker rural ring — so legionnaires were rural-born even relative to their own district town. Desertion runs the other way in the larger pairs (Brno město 2.98 % against the ring's 1.56 %, Liberec 4.21 against 2.44) and vanishes in the smaller ones. Pooling averages every one of those contrasts away, which is the price the census charges.

2026-08-05 (the rolník decision gets an individual-level witness)

Check 6 settled 18 % of the sample and named its own weakness in the same breath. The 1910 Berufsstellung test is ecological: it correlates okres compositions, so it can say that where the census counts more own-account farmers the legionnaires' farm occupations are more owner-coded, and it can never say anything about one man. It also could not tell a holder from a farmer's son. Both limits were written into D-33 rather than glossed over, and both are now partly answered.

• Check 7, from a source that records people one at a time. The Vienna Database household listings pair an occupational title — drawn from a closed list of 521 German terms — with a household position assigned separately, for 163,400 people. BAUER is 0.3 % Gesindeperson; LANDARBEITER/KNECHT is 94.9 %. Among men of Legion age BAUER is 97.5 % household head and 0.3 % son, which is the farmer's-son half: on these listings a farmer's son is written down as SOHN, and the farm title goes to the holder.
• It corroborates, and the sheet says so in the status line. Title and household position are not independent — one clerk assigned both, under a convention that hands the farm title to the head — so this shows what the word meant to the people recording it, not an independent measurement of who farmed what. D-34 records the ecological limit as narrowed, not closed, and D-33 keeps its row and its date.
• The two checks fail in opposite directions, which is why both are kept. The census is the right territory and the right decade at the wrong level of aggregation; the listings are the right level in the wrong territory (44 Austrian localities, no Czech lands) and mostly the wrong period. Neither alone would carry 18 % of the sample.
• The weakest link is named on the sheet, not buried. rolník is glossed Landwirt; the term on the closed list is Bauer, and a bare LANDWIRT does not appear on it at all. The bridge runs through a synonymy the check assumes rather than demonstrates — worth stating plainly, since this project's gloss for sedlák, the Czech word that literally means Bauer, covers six records.
• The source is old enough to be genuinely independent. The Vienna Database on European Family History was built 1974–1984 at the University of Vienna with the MPI for History in Göttingen and the University of Salzburg. Its occupational coding was done by historical demographers reading the listings, decades before this project existed and with no contact with it — which is what "external" is supposed to mean and often does not. Cited in the form MOSAIC prescribes; free on IPUMS registration, with no redistribution grant, so the file stays in gitignored data/ and only derived counts reach reference/.

2026-08-04 (a plain-English doorway on every method sheet)

The register these sheets are written in is correct, and it is also a wall. A method sheet has to be precise, hedged, and free to assume the reader knows what a járás or a HISCLASS is. That is right for the appendix it grounds and wrong for the first thirty seconds of a supervisor, a referee or a historian who has never opened the project.

• "Explain it simply" on 14 sheets. A folded card sits directly under the In short summary on every How it works sheet and on the validation and decision sheets. Open it and the same mechanism is explained the way you would explain it over coffee: an analogy, then what the code actually does, then the honest limit. Posting a letter addressed only to Nová Ves is the place matcher; a job title becoming a rung on a ladder is HISCO; the 26,000 blanks that became 26,000 confident answers is the audit.
• Deliberately a <details>, not a drawer or a modal. It costs no JavaScript, the browser's own Ctrl+F reaches inside it, the sidebar search indexes it, and the print stylesheet already expands every fold — so the plain-language layer travels into the PDF instead of vanishing from it. The one print rule added keeps its title, which the unlabelled source folds do not need.
• It carries no interactive chip, on purpose. The ⇱ badge means this figure responds to input, and the sidebar's interactive dot is derived from those badges rather than declared. A disclosure triangle on a paragraph of prose is not an interactive figure, and labelling it as one would have made the dot claim fourteen sheets had figures they do not.
• The text lives in build_docs.py, not in reference/*.md. Those files are the record; this is the doorway. An analogy about post offices does not belong in the middle of a document the data-construction appendix is built from. A build-time check warns if a sheet in scope loses its entry.

2026-08-02 (the validation evidence is published, and indexed)

The evidence outgrew the page it was appended to. The entry below folded VALIDATION_OCCUPATIONS.md into the bottom of the Occupation → HISCO sheet — the right call while that was the document's only home, but it left the occupation checks with no place in the validation section, where a reader looking for "what has this project been checked against" actually goes. Fidler sat there alone, and the two documents never met. With a third external check added, that no longer holds: this supersedes the "one page, not two" call below.

• Three validation sheets, not one. Validation register (new, reference/VALIDATION.md) lists every check — Fidler, the German link bridge, OccCANINE, CZ-ISCO — with the column that makes it honest: what each one is blind to. External evidence and self-checks are kept in separate tables, because a QA invariant catching a build that disagrees with its own rules is not evidence the coding is right. The two detail pages sit beside it: External validation — units & ranks (the former "External validation") and External validation — occupations.
• A summary on the method page, not a second copy. OCCUPATION_CODING.md §9 was three lines; it is now a table of the three external checks with their headline numbers and blind spots, linking through. The full append at the bottom of the sheet is gone — with the document rendered as its own sheet, keeping the append would have printed every word twice.
• Seven published figures were wrong and are now right. The roadmap had been carrying them as a known issue: hand-assigned keys read "216 of 459" against an actual 429 of 470, unmatched "1,920 (1.6 %)" against 1,840 (1.5 %), the empty-cell invariant quoted 14,044 for both fields when the interwar side is 127,913, and zámečník's 0.12 was called the lowest OccCANINE confidence when brašnář is 0.022. All re-derived from the build rather than edited by hand. Publishing a page is a reason to check its numbers, not to assume someone already did.
• CZ-ISCO as a third coder. The one check that tests the Czech reading itself — the other two both start from a German gloss this project wrote. 99.1 % agreement on the records it can score, and 15 substantive disagreements worth reading.

2026-08-02 (earlier — the occupation method finally has a page)

The most-authored thing in the repo had no sheet. OCCUPATION_CODING.md and VALIDATION_OCCUPATIONS.md were never in the site registry, so every one of the four cross-references to them rendered as a dead local path — the reader was pointed at a file they could not open. Both are now one sheet, Occupation → HISCO, sitting beside Place matching where it belongs: the two documents are siblings, one resolving a birthplace to a district and the other an occupation to a social class.

• One page, not two. The method and the evidence are one argument. Splitting them would have put the OccCANINE agreement table a click away from the coding decision it vouches for, so the validation runs on below the method under its own rule.
• A funnel, and where the code lands. The same grammar as Place matching, because it is the same shape of problem: four rules tried in order, an earlier one wins, and the ribbon down the left is what is still uncoded. The step at exact swallowing five sixths of the lane is the finding, not a caption about it. On the right, the HISCLASS ladder — which is where an occupation sheet has to end, since the class is what the analysis consumes. Click a rule for what it refuses and the strings it actually matched; click a rung for the trades on it and how often those men deserted.
• The robustness check, as something you can do. This sheet has always claimed that the three compound rules are not manufacturing the class gradient. Now you can untick them. The gradient below redraws on whatever is ticked, with 95 % Wilson intervals — the interval that behaves when the sample shrinks and the event is rare — so the bars widening rather than moving is the answer. The readout hands back the exact pandas mask.
• Nothing is transcribed. Every count, ribbon width, example string and rate is recomputed from data/final at build time. The live figure and the Check 4 table in the prose are the same five numbers, which is the only way they can be trusted to stay that way.

2026-08-02 (a second coder for the occupations, and a gate that could not see half the data)

The occupation dictionary was already the most-authored thing in the repo — there is no Czech HISCO source anywhere, so every code is a decision taken here. Two things were missing: an outside opinion, and a gate that watched the whole of what the dictionary produces.

• OccCANINE as an independent second coder. The transformer of Dahl, Johansen & Vedel (2024) was fed the German glosses, not the Czech — Czech is not among its 13 training languages, so feeding it Czech would measure the model going out of distribution rather than measure our coding. Both coders therefore see identical input and every disagreement is about the code, not the translation. 221 glossed-and-coded keys, 96,079 records.
• Read the class row, not the first one. Exact 5-digit agreement is 53.8 % of keys, and that number is close to meaningless here: most disagreements are neighbouring codes inside one trade — kovář 83110 vs 83190 are both blacksmiths, holič 57030 vs 57025 both hairdressers, knihař 92620 vs 92625 bookbinder by hand or by hand-or-machine. What the analysis consumes is HISCLASS and the major group, and there the two coders agree on 88.2 % and 96.7 % of records.
• The model is unsure about the words we had already flagged. Mean confidence is 0.83 where the coders agree and 0.75 where they do not — but the single largest disagreement, zámečník / Schlosser (2,162 records), comes back at 0.12. That is the key this project had marked review for the same reason: Schlosser is genuinely ambiguous between locksmith, machinery fitter and plumber. Two independent coders hesitating over the same word is the check working, not failing.
• A gendarme with no code, hiding behind a half-blind gate. četnictvo was coded 58210, which does not exist in the HISCO authority — so those records resolved to no label, no class, no HISCAM. The reason it survived is the more interesting half: every occupation invariant read occupation_hisco alone, and četnictvo appears 0× at enrolment and 58× in the interwar field. An enrolment-only check was structurally incapable of seeing it. Both halves had to move together — widening the check while the code was still wrong just fails the gate. Now every prefix-scoped invariant iterates both families, and the gate went 37 → 41.
• Four modifier pairings the Dutch data never saw. The authority is keyed on (code, status, relation, product) and built only from combinations HSN observed, so hospodářský adjunkt — 21230 with status 33, Subordinate — matched nothing and returned no scales at all. Those now fall back to the base code, modifier preserved in its own column. The fallback is suppressed wherever relation is set, because HISCO withholds HISCLASS there on purpose: a miner's widow is not a miner.
• A cottager is not a farmer with ten hectares. domkář, chalupník and malorolník move off 61110 General Farmer (HISCLASS 8) to 61115 Small Subsistence Farmer (HISCLASS 10) — the rung HISCO provides between the holder and the farm labourer. rolník stays at 61110, and the split makes that position easier to defend, not harder.
• Coverage 97.3 % → 98.5 %. The curated skeleton doubled to the top 2,000 keys and 459 now carry a code. New entries are explicit, including the typos: an edit-distance rule that turns rolnlík into rolník would just as happily turn one real trade into another.
• And the coding does not manufacture the result. Switching off the compound-matching rules and keeping only whole-string dictionary hits leaves the desertion gradient across social classes unmoved in shape — levels drift up by a fifth of a point, the ordering does not budge. Those rules are the least externally-anchored part of the method, so this was the thing worth testing. See the Occupation and External validation sheets; the calls still open are on the Roadmap.

2026-08-01 (the other half of the map — Slovakia's twins, and a lexicón read too narrowly)

The Czech twins were settled in the morning; the Slovak side had the same structure and none of the fix. Working through it turned up two defects in the lexicón parse that had nothing to do with twins and everything to do with the same habit — writing a rule for the commonest case and never checking what else the page prints.

• Twenty-six járás/town twins, and the town units were empty. A chartered town sat administratively outside the district it gave its name to, so Bratislava, Košice, Nitra, Prešov, Žilina and 21 more name two units apiece. 12,760 records sat in a pair and only 387 on a town. Now: a bare name means the járás, and the municipality promotes to the town — 1,888 records move, 71 with no municipality go to the ring flagged undecided. Košice vonkov resolves by rule instead of through the fuzzy tier, where 589 records had been reaching the right answer by luck. Decisions D-19.
• And here the census can split them — all 52 units have their own row, so unlike the Czech pairs there is nothing to pool and both members keep their covariates. Which is the point: the ring around Žilina was 0.964 Slovak, the town 0.540; Bratislava's ring 0.633 Slovak against the city's 0.149. The asymmetry with D-18 is the source talking, not a second opinion.
• A plural the parser could not read. The lexicón closes each county with its chartered towns under one of four headers; the matcher knew three. The fourth — mestá so zriadeným magistrátom — is the only one seven counties use, so their 27 town rows silently kept the county's last district: obec Nitra, 16,419 people, filed under okres Nové Mesto nad Váhom. Right county, wholly plausible, wrong. Bratislava's four towns escaped only because a singular header sat above them on the same page. Decisions D-20.
• Nitra's 913 Serbs. The numeric columns are right-aligned; the header anchors mark their left edges. Bucketing each number by its centre drifted the narrow ones a column to the left, and dropped two into the same bucket, where they concatenated — Nitra's 9 Croats and 13 Serbs became 913 Serbs. Reading the right edge repairs 27 broken row sums and, more to the point, 135 rows that reconciled perfectly while sitting under the wrong heading. No share-bearing column moved on any row: the damage was confined to Romanian, Croat, Serb and Other.
• Nineteen villages in the wrong valley. Three okres blocks are headed adjectivally (Hradný okres) rather than okres X, which the regex did not match at all — so the Orava villages under Hradný inherited the heading above them and were filed under Trstená instead of Tvrdošín, ~98 records. Two more blocks lost their district outright. The three district names are curated onto the backbone; unmapped obce fall from 93 to 57, and the remaining 57 are the four districts that genuinely sit outside the Felvidék frame. Decisions D-21.
• A coin flip, retired. Ilava nebo Zlaté Moravce scores 100 against both, and the winner was whichever key the lookup happened to hold first. 38 such records across 32 strings, every one a cell naming two districts or truncated past recognition. A tie is now a refusal — the same rule the settlement pass has always applied. Decisions D-22.
• The same coin flip on the Czech side — where the record can be asked instead. Checking whether the defect had a Cisleithanian twin found 5,100 records resolved by dict order. But refusing them, as Slovakia does, would have been the wrong answer: the record's own municipality knows which okres it sits in, and it disagrees with the arbitrary winner on 59 of the 224 births it can decide. 22 men recorded as born in Ostrava were born in okres Frýdek — the Silesian side, not the Moravian city; 14 Nové Město are na Moravě, not nad Metují. Curating the frequent cases by population, which was the obvious alternative, would have got every one of those backwards. What the municipality cannot decide now refuses. Decisions D-23.
• Praha's suburbs were not Praha. Smíchov, Žižkov, Karlín and Královské Vinohrady were political okresy in their own right in 1910, so Praha - Smíchov names Smíchov — and the matcher had been reading it two ways at once, sending Smíchov to the capital while sending Žižkov and Karlín to their own units. 68 births move out of Praha, and the court districts come with them (Praha - Říčany → okres Žižkov). It never overrides the municipality: where both can speak, all 217 agree.

2026-08-01 (the twins, settled — and a census that cannot split nine of them)

The morning's fix asked a question the afternoon had to answer: a Statutarstadt was its own district, so what does a record naming one actually mean? Two answers, and the second was not the one anyone expected.

• One naming convention, so a bare name means one thing. Nine pairs were X / X (město), Brno alone Brno (city) / Brno-venkov (ring) — so a bare district resolved to opposite members, and nine city units held 0 records while Brno city held 2,768. Brno's two rows are renamed to match: city X (město), rural bare X, everywhere. That is also what the ČSÚ register means by a bare name — all 118 ring villages around Brno record okr. Brno — and under the old naming 117 of them sat in the city okres, which is why okres Brno-venkov had not one birth coordinate. Decisions D-17.
• The municipality picks the member, by definition rather than by inference. A statutory city was a single municipality — one locality apiece in the census size bands — so naming the city puts a record in the city unit and naming anywhere else puts it in the ring. 3,827 records move. The 66 with no municipality at all go to the ring, flagged undecided. A trailing venkov now resolves by rule instead of through the fuzzy tier, where it had been reaching the right answer by luck and was afterwards indistinguishable from a bare name.
• The census cannot separate nine of the ten pairs at all. A statutory city had no Gerichtsbezirk of its own, so the court-level workbooks carry 330 of the 339 Czech-lands courts and the nine missing ones are exactly the shared seat courts. Opava (město)'s 0.463 Czech share is not the city — it is Gerichtsbezirk Troppau, city and ring; Opava's 0.069 is only Odry and Vítkov. Not coarse: mislabelled. Jihlava, Olomouc and Frýdek had the whole pair on one row and nothing on the other. Only Brno genuinely splits.
• So the baseline analytical unit is the pooled pair — 148 units, counts summed and every share re-derived from them, never averaged across the two rows. The 158-row member panel keeps the okres-level covariates (literacy, illegitimacy, occupation, population) split for all twenty and blanks the court-derived ones on the eighteen the source merges. Decisions D-18.
• Coordinates, finally. Brno-born legionnaires go from 0 to 773 of 775, and Praha from 0 to 2,691 of 2,693 — the register's okres field had been swallowing the "se člení na osady" narrative, so the key that matches Wikidata was a whole sentence. Praha, Frýdek, Těšín and Bílsko are absent from Wikidata's reach entirely and now carry curated points.

2026-08-01 (the map says what it means; a city-shaped hole closed)

The evidence panel added this morning earned its place by lunchtime: it surfaced Brno → okres Jilemnice, and pulling that thread found a defect with a single cause and a wider reach than the misplacements suggested.

• The ČSÚ register leaves a statutory city's okres empty — the city was its own political district — and the settlement lookup is built only from settlements that have one. So the city vanished and a like-named hamlet won by default: 456 records into okres Jilemnice, plus Těšín → Jičín and Lvov → Falknov. Twelve headwords are now curated in, and a bare municipality may no longer resolve onto an osada — a named part of another village — which closes the class rather than the instance. Decisions D-16.
• The coordinates did not come back with them. A city-born record still has none: 775 for Brno, 154 Kroměříž, 132 Olomouc, 119 Jihlava, 94 Liberec, 86 Opava — and 2,693 for Praha. The curated row carries an okres_id for the city district, while a bare recorded Olomouc resolves to the rural one, so the (municipality, okres_id) join that attaches coordinates misses by exactly one unit. That is the same inconsistency the twins entry below describes, seen from the other end, and it is fixed with them.
• The map draws court districts but measures political okresy. Every internal court line used to read as a district boundary. They are now told apart exactly: adjacent polygons in the MPIDR file share their border vertex for vertex, so an edge whose two owners sit in the same okres is a court line and everything else is a district boundary. Court lines are pale hairlines, district boundaries solid. A unit toggle shades the courts themselves from birth_court_id — a coordinate-derived layer covering 53,746 records, so a cross-check rather than a finer version of the same count.
• Click a district and the map runs backwards, from the place to the men: how many were born there and where it ranks, the desertion rate with a 95% interval against the national one, the theatre mix, and the mean rank reset on combat ranks. Prague, for instance: 7,250 legionnaires, 2.87% ± 0.38 against 1.98% overall. Rates are suppressed under 30 births.
• The town twins are not an error — but the matcher reads them two different ways. Ten districts come in city/rural pairs — Brno / Brno-venkov, Jihlava, Olomouc, Opava, Znojmo, Liberec, Kroměříž, Uherské Hradiště, Frýdek, Bílsko — because a Statutarstadt was its own district. Nine of them are named X (rural) and X (město) (city); Brno alone is named the other way round, Brno for the city and Brno-venkov for the ring. So a bare recorded district name lands rural for nine cities and city for Brno: 9,355 legionnaires and 8,983 VÚA records sit in a twin pair, nine of the ten city units hold 0, and Brno city holds 2,768. Measured, not asserted — and being corrected next, along with the census question underneath it. Decisions D-17.
• How good are the boundaries? Asked directly, answered on the sheet: about 95%, decomposed into unit identity (~99–100%), topology (every interior edge shared by exactly two polygons, none by more than two), a point well inside a district (~98–99%) and a point near a line (lower — the file is generalised to ~41 vertices per polygon).

2026-08-01 (place matching: a funnel, a filter and a map)

The place-matching figure was a picture of the pipeline. It is now an instrument for interrogating it — three additions, all recomputed from data/final at build time.

• The cascade is a funnel. A ribbon runs down each lane showing what is still unplaced, narrowing by exactly what each tier placed, and each box carries a bar for its tier against the largest in its lane. The dominance of the exact-okres pass — two thirds of the ribbon gone in one step — is now visible before a number is read. The residual box at the foot of the Czech lane and the "in" caption at the head of the Slovak lane are the same count, which is what "the second frame runs on the first frame's residual" actually means.
• The tiers can be ticked off. This sheet has always said an analysis can drop the weak tiers and see what moves; now it can be done here. Untick a tier and the readout gives the surviving N, its share, and the three-line pandas mask that reproduces it — copyable. The headline: 113,571 of 132,993 legionnaires (85.4 %) carry a birth district on all tiers; dropping the exact-okres pass alone leaves 20,336.
• A map, with both frames on one plane. Czech okresy (MPIDR, EPSG:3035) and Upper-Hungary járás (GISta Hungarorum, EPSG:3857) drawn together — the Hungarian coordinates inverted to lon/lat and re-projected through the same LAEA forward as the Czech side. Validated against okres_centroids_1910.csv (median 3.1 km) and the known positions of Bratislava, Košice, Prešov and Žilina (1–8 km). Shading is quantile bins, not a linear ramp: birth counts are skewed enough (Prague at 7,250 against a median in the hundreds) that scaling to the maximum washes every other district to the same near-white.
• Each stage now shows the strings it actually matched. The panel lists the commonest recorded string → resolved district pairs for the selected tier — aggregates over place names, no individual records. This is the part that earns or loses trust: fuzzy turns out to be doing honest work (Ostrava → Moravská Ostrava, Dvůr Králové → Dvůr Králové nad Labem), while muni immediately shows Brno → okres Jilemnice (3 records) — the Praha problem again, a big city missing from the ČSÚ register so a like-named hamlet captures the births. Surfaced, not silently fixed; the whole tier is 241 records and is ticked off in one click.

2026-08-01 (affordance chips)

• The "this responds to input" label is now one rule, applied by the builder. It had drifted: Place matching and Rank ladders showed an ⇱ chip but carried no red dot in the sidebar, one chip sat inside a toolbar where nobody scanning headings would see it, and two sheets announced themselves with the word "Interactive:" in the blurb while the rest did not.
• A chip now sits at the end of the heading that introduces the block, and nowhere else — emitted only by ix_head(), one chip per affordance, so a figure that both toggles and carries tooltips shows both. A build-time check warns if a chip ever escapes a heading again.
• The sidebar dot is derived from the page, not declared. A sheet is marked interactive iff it actually rendered an ⇱ chip, so the registry can no longer disagree with what a reader sees. Six sheets carry the dot; Place matching and Rank ladders are the two it gained.

2026-07-30 (rank bridge, second pass)

• Both ladders are now clickable, in both directions. A Habsburg rank still answers "where did these men end up"; a Legion rank now answers "where did they come from", which is the direction the endogeneity question is actually asked in. Same joint distribution, conditioned the other way.
• At OF-3 and above the panel lists the individual records — 94 of them, 74 Legion and 20 Habsburg — because at those counts an aggregate is not checkable: a bar reading "31 majors" hides whether those are 31 real majors or a parser artefact. Each row gives the legion_id as a link to its csol.cz entry, the birth year, the rank on the other ladder and the theatre. Ids only, no names, so the published page still carries no named individuals.
• The fine common grade is back, as a bare muted numeral on each ladder's outer edge — the rung rank_reset is actually measured on, next to the coarse NATO band it collapses into.
• An open question logged, not silently fixed. svobodník is coded OR-3, but §G2 reports the APersP-01 CZE note as vojín, svobodník = OR-1, and Bednařík's 1916 POW-camp pay scale pays "pěšák a svobodník 15" — private and svobodník identically, the first real step at desátník. So the enlisted band rests on the weakest evidence in the figure. It is now Decisions D-15 (Open) with the full argument, flagged under the figure, and nothing in the data was changed: the band is only the robustness axis, and repairing one rung would leave the rungs above it inconsistent.

2026-07-30 (later)

• References is now the master bibliography, and complete. Every work cited anywhere on the site is listed — which it was not: the three Legion chronicles (Vaněk, Bednařík, Boháč), the k.u.k. rank tables (Dolejší et al., Rest et al.), the NATO APersP-01 / STANAG 2116 standard the whole common scale is expressed in, the Vašátko insignia plates, the four interwar statutes and the superseded 1910 Ortsverzeichnis were all cited but absent. 41 entries, in three sections (literature · statutes and standards · data sources).
• reference/references.bib is generated from that list on every build, with the Zotero citekeys pinned so a re-import updates records instead of forking them. A build-time assertion refuses to emit one without the other, so the sheet and the .bib cannot disagree. The hand-maintained rank_sources.bib is retired, its entries folded in.
• Citation linking fixed. A work cited in two textual forms on one sheet — Vaněk et al. (1925–1929) in one paragraph, Vaněk et al., 1925–1929 in the next — had only the first form linked, because a match on one pattern stopped the others being tried. Inline citations went from 200 to 263.
• Rank bridge corrected. The right-hand connectors were mirrored: they left the grade rung at the rank's height and arrived at the rank at the rung's height. Each rung is now labelled with its NATO band alone; where consecutive rungs repeat a band, that repetition is the point — the three subaltern rungs all reading OF-1 is why the fine scale exists.
• Figures regenerated by re-running 03_visualisation against the current data/final — all twelve, including the exit-code table that the audit flagged as stale.

2026-07-30

• Structure given rules. Every sheet now has the same bones — one title, a group chip, a TL;DR card, a contents strip, a "see also" footer — instead of per-sheet improvisation. Links follow one rule too: prose links underline and firm up on hover, navigation never underlines, and outbound links carry a ↗.
• The record of change split three ways, each answering one question and nothing twice: Roadmap (what is open) · What's new (this sheet — what changed and why) · Commits & versions (the git record). A strip at the top of each says which is which.
• Decisions — a register of every judgement call, with a one-line reason and a link to the argument. The audit sheet is now titled by its date, because that is what it is: one dated review, not the decision log.
• Place matching gained a flow diagram — the Czech and Slovak passes side by side, each stage carrying its live record count and its refusal rules.
• Rank ladders gained a cross-army bridge: Habsburg ladder, the common grade scale, and the Legion ladder in one picture, which is what rank_reset actually subtracts.
• Overview gained a reading-path map for first-time readers.
• Cross-sheet section links now work (a link to another sheet's heading used to change the URL and nothing else), and the page has a favicon.

2026-07-23

• 1914 Habsburg order of battle — the complete peacetime army digitised from Ehnl (1934) Tabelle I (443 units: common-army, Landwehr and Honvéd infantry, Feldjäger, cavalry, artillery and technical troops). Each regiment carries its Ergänzung home district (recruiting Bezirk, standardised to modern place names — Trutnov, not Trautenau) and its national composition (mother-tongue shares, dominant language, Slav share, Herfindahl).
• Habsburg army & recruitment sheet — how the territorial Ergänzung system tied a regiment to a home district and, through it, to a nationality; the natural-experiment angles this opens; qualitative context from the reference works.
• Regiment explorer — a filterable, sortable table of all 443 units (filter chips by branch and by language, live search, click-to-sort).
• Descriptives waterfalls now separate "Habsburg unit recorded (present)" from "Habsburg unit matched to Tabelle I" — a recorded unit string is not the same as one that resolves to a real 1914 regiment (and hence to a home district and nationality). Legionnaires: 100,531 present (75.6%) → 81,911 matched (61.6%); fallen: 116,138 present (62.4%) → 89,378 matched (48.0%). A short decomposition explains what the unmatched remainder is (mostly no-unit records, plus wartime Landsturm absent from a peacetime order of battle).
• Source provenance gained the new Habsburg-army sources (Ehnl, the 1914 Schematismus, the secondary literature) and mechanism nodes for the recruiting district, unit nationality and deployment; the side panel now shows each source's full reference and a one-line note on how it enters the pipeline.
• Terminology unified: the "A-H" abbreviation is now spelled out as "Habsburg" throughout.

2026-07-22

• Habsburg comparison-population waterfall added next to the legionnaire one on the Descriptives sheet — same interactive machinery (toggle covariates, live theatre composition). The control group is the fallen soldiers who did not cross into the Legions; a legion link disqualifies a control, so the first toggle excludes the legion-linked and the highlighted count tracks the valid controls.
• Overview made less committal about the headline regressor and estimation method, which are still in development.
• This What's new sheet added, with the live-page link.

2026-07-21

• English value translations surfaced (units, religion, rank tiers) in the data dictionary and variable notes.
• References sheet (APA): literature and data sources (census, lexicons, GIS, the two scraped databases). Inline author-year citations across every sheet are now clickable and preview the full reference on hover.
• Roadmap / to-do sheet (done vs. remaining).
• Interactive vs static affordances: an "⇱ Interactive" chip on figures that respond to input, a red dot on those sheets in the sidebar, and a "↕ Hover for values" chip on the static-with-tooltip histogram.
• Source texts consolidated under library/ (lexicons, reference works, books).
• --noindex build option for link-only publishing.

2026-07-20

• Interactive sample-size waterfall (legionnaires): cumulative coverage bars, a live filter panel with per-toggle marginal cost, desertion rate by theatre, a shareable link, and a runnable pandas snippet.
• rank_reset: diverging histogram, by-theatre overlapping areas, and a combat-ranks-only / inclusive toggle (the legion_noncombat decision).
• Units & armies sheet (regiment vs unit_key, Fidler cross-check).
• Source provenance — interactive map from each source to the mechanisms it grounds.
• Architecture map — click a module for an info panel (role, version, imports).
• External validation sheet (vs Fidler 2025) and the Data QA gate panel.

Earlier

• First generated site: pipeline, variables, place-matching and rank-ladder references; live descriptives recomputed from data/final; per-script sheets; symbol-usage index; data dictionary and census codebook.

PipelineHow it works

What every script does. · source: reference/PIPELINE.md

Script index

One sheet per module — its docstring, public API, version, callers and git history, all read from the source. Pick one in the sidebar under Pipeline scripts. The narrative below explains how they fit together.

155 rows

Audit status: audited and fixed three times — 2026-07-16, 2026-09-22 and 2026-10-24, each with its findings and measured before/after numbers. Brought back in line with the code on 2026-11-02, when this guide named about a third of the modules in utils/: every module is now named here, at least in one line. The docs site gives each module its own sheet, built from its docstring and source, so that is where the detail lives; this guide says what each one is for and where it sits in the run.

The pipeline at a glance

Five notebooks orchestrate everything; all logic lives in utils/, the notebooks only call it in order. Three are the person pipeline (collect → clean and link → draw); two drive the slow, paid reading passes, kept apart from the fast code as the code standard asks: the 1910 place register (15) and the amateur-theatre model pass (16). A few measures are still pilots run from the command line (utils/reference_tools/, §7) and write only to data/interim/. data/ holds machine inputs/outputs (gitignored); reference/ holds small curated inputs (tracked).

01_data_collection      scrapers → data/raw/{legionnaires,habsburg}/*.csv  + post-scrape QA
02_data_cleaning        §1 legion parse      raw csv → interim parquet   (clean → parse → to_english)
                        §2 habsburg parse    raw csv → interim parquet
                        §3 places+census     §3a districts_1910  §3b census_okres_1910
                                             §3c settlements_1910 (+coords, courts)
                                             §3d districts_hu_1910 + census_hu_jaras_1910
                                                  + settlements_hu_1910
                                             §3e today's geography (RÚIAN, Geonames, PRNG,
                                                  vanished places) + settlements_1910_geo
                        §4 linkage + place standardisation → data/final/persons/*
                        §4b wide analysis files (persons × census) → data/final/analysis/*
                        §5 data_dictionary.csv
03_visualisation        reads data/final/ ONLY → output/ figures + LaTeX tables
15_frozen_run           the 1910 register: freeze → read (batched) → check (E3) → Opus →
                        ledger → output/frozen/<edition>_<pages>/ → §18 data/final/register/
16_theatre_model_pass   amateur-theatre chronicle clauses the rules leave open, read by a
                        model → data/interim/theatre/  (gold → full; SPEND off, D-157)