Poengkart: Open Admission Thresholds and a Calibrated Forecast for the Norwegian Upper-Secondary Intake

Abshalom Dayan
Technical report · October 2026 · v1.32 (version history in Appendix D)
Application: poengkart.no · Code and data: github.com/avshalomd/poengkart

Abstract

1. Introduction

In most Norwegian counties, admission to upper-secondary school (videregående opplæring) is a points-based queue. Applicants rank programme–school combinations in the counties' joint application portal, vigo.no, and each programme admits by descending lower-secondary grade points until its capacity is reached. The points of the last admitted applicant — the poenggrense, or admission threshold — is a market-clearing cutoff, re-set every year by that year's cohort. For a family choosing where to apply, the distribution of plausible thresholds at each candidate programme is the central planning quantity.

That quantity is public only in fragments. Eight of Norway's fifteen counties (fylker) publish thresholds today — one of them only inside a Power BI dashboard, obtained here as an extract on request. A ninth released figures on request that cannot be compared with the rest, and two more printed tables in the 2010s and no longer do; all three are left out of this report (Section 4.4). Those that do publish tables that differ in structure, in retention (Oslo's own archive reaches back to 2017, and its older years survive only as copies; Akershus publishes only the latest year), in which of up to three intakes the figures describe — often without stating which — and in the notation used for special cases, such as programmes that filled with applicants who had no points. No national, machine-readable, or historical compilation exists. Forecasting from such data is hard for structural reasons: a threshold exists only where demand exceeded capacity, so the panel is censored by construction; a quarter of the series have a single observed year; and the published figure is intake-specific, so identical demand can print different numbers in different counties.

To the best of our knowledge, no prior work forecasts admission thresholds in the Norwegian intake, and published work on cutoff forecasting in any centralised score-based admission system is thin (Section 3). This report describes Poengkart ("points map"), an open dataset, forecasting model, and deployed application. Our contributions are as follows:

• An open, register-normalised dataset of every retrievable published threshold — 227 schools, 2,936 programme rows, 19,632 cell-level observations across 8 counties and the years 2012–2026 — in which every row is resolved against the national Grep register and every cell's state (numeric threshold, filled at zero, no waiting list, quota-based admission) is preserved as data (Section 4).
• A calibrated probabilistic forecast of the next intake, P(admission | applicant's points), from a hurdle model with hierarchical partial pooling and a random-walk county–year component, whose predictive spread is estimated from walk-forward forecast errors rather than model internals (Sections 5–6).
• A held-out evaluation on the 2025–2026 intakes, with cluster-bootstrap uncertainty, showing the model beats persistence and group-mean baselines on RMSE in every history stratum (5.14 vs 6.18 and 6.01 points in the deepest stratum), covers 80.4% of outcomes with nominal 80% intervals, and produces calibrated probabilities that beat both a deterministic and a probabilistic persistence rule (Brier 0.091 vs 0.157 and 0.096), plus ablations of the structural choices (Section 7).
• Measurements of the publication practice itself: a paired within-year estimate of the intake-round effect (−3.0 to −3.1 points, 14–37% of queues cleared between intakes), a decomposition of raw school means into demand, programme mix and market timing, and a residual-based audit that distinguishes genuine extremes from parsing damage (Section 8).

Section 9 describes the deployed application; Sections 10–12 give limitations, ethical considerations, and recommendations to the publishing counties.

2. Institutional background

Points and applications. Norwegian pupils finish compulsory school with a points total (grunnskolepoeng): ten times the mean of their final grades on the 1–6 scale, a maximum of 60; pupils without grades compete "without points". Applications go through vigo.no, the counties' joint application portal, by 1 March. An applicant to the first year (Vg1) must name three programmes of study (utdanningsprogram) in order of preference and has a statutory right to a place on one of them (opplæringslova § 5-1); the form takes up to ten ranked wishes, each a programme at a school. For the 2025–26 intake, 80,144 pupils applied to Vg1 and 210,866 to all levels (Utdanningsdirektoratet, 2025). The programme structure is national, maintained by the Directorate for Education and Training (Udir) in the Grep register: 15 utdanningsprogram subdivided into programområder (programme areas) across year levels Vg1–Vg3, plus the one-year påbygging to general university admission, which the counties publish beside them and which we carry as a sixteenth programme.

Intakes. Counties run the intake in up to three publications (inntak) in July–August, numbered "1. inntak", "2. inntak", "3. inntak" as on the applicant's letter; we use inntak and intake interchangeably below and say "round" only for the code's round_bridge. In each intake every programme admits by descending points among remaining applicants; seats declined in one intake are re-offered in the next, so thresholds fall between them (Section 8.1 measures how much). The published threshold is the points of the last admitted applicant in the published intake — and counties differ in which one they publish, sometimes without saying. Statutory priority admission (fortrinnsrett) and admission on documented grounds (inntak etter dokumentasjon: International Baccalaureate, elite sport) bypass the points queue entirely.

Catchment. Whether an applicant competes county-wide or within an intake area (inntaksområde) is set by each county's own admission regulation. Oslo and Rogaland run a county-wide points queue (Oslo: FOR-2025-01-29-147); Møre og Romsdal likewise. Akershus admits within three intake regions (FOR-2025-02-13-354, in force from 1 January 2025), Innlandet requires the nearest school offering the programme as the first wish (FOR-2024-12-17-3529 § 2-4), and Buskerud, Trøndelag and Vestland give residents of an intake area priority. Where catchment applies, a published threshold binds only applicants resident in the area, and the application says so.

Two shocks in the panel's window. Written examinations were cancelled in 2020, 2021 and 2022, so those cohorts' grunnskolepoeng rest on classroom grades alone; the national mean rose from 41.9 in 2019 to 43.2 in 2020 and fell back to 42.4 in 2023 (Statistics Norway, table 07495), a shift of the whole points scale that a threshold model sees as a county-year level move. And the county map changed twice: Innlandet (2020, from Hedmark and Oppland) and Vestland (2020, from Hordaland and Sogn og Fjordane) were formed at the start of the panel's dense years, and on 1 January 2024 Viken was dissolved into Akershus, Buskerud and Østfold, so Akershus's and Buskerud's recent series start in 2024; their older years (2012–2015) are the pre-Viken counties' own, with no Viken year between them.

Two structural facts shape everything downstream. First, a threshold exists only where demand exceeded capacity: undersubscribed programmes have no waiting list (ingen venteliste) and no threshold, so the dataset is censored by construction, and the censoring is informative — it is exactly the "no queue" event our fill model predicts. Second, the published figure is intake-specific: comparing thresholds across counties without knowing the intake conflates market differences with administrative timing.

3. Related work

School choice and matching markets. Centralised admission mechanisms descend from deferred acceptance (Gale & Shapley, 1962); their analysis and design as school-choice mechanisms is a mature literature (Abdulkadiroğlu & Sönmez, 2003). In large score-priority markets, stable outcomes are characterised by market-clearing cutoffs (Azevedo & Leshno, 2016) — the very object we forecast — and cutoffs figure as sufficient statistics in empirical work on centralised admission (Fack, Grenet & He, 2019). In Norway, Kirkebøen, Leuven & Mogstad (2016) use the university-admission cutoffs as a regression discontinuity to estimate returns to fields of study; the upper-secondary cutoffs studied here have not, to our knowledge, been compiled or modelled. This literature designs and analyses mechanisms; it does not forecast next year's cutoffs for applicants.

Information and beliefs in centralised admissions. Applicants misjudge their admission chances, and the misjudgement has consequences: Kapor, Neilson & Zimmerman (2020) show that families in New Haven hold inaccurate beliefs about admission probabilities and that the resulting application mistakes are costly, and Arteaga et al. (2022) show in Chile and New Haven that many applicants submit "smart" lists that nonetheless leave them unassigned, and that a warning about the risk changes behaviour. A calibrated public forecast of cutoffs is the informational instrument those papers point to; our contribution is the forecast and its evaluation, not the effect of providing it, which we cannot measure (Section 10).

Norwegian school choice. The Norwegian debate has been about the admission rule rather than its outcome data: Sandsør (2020) reviews the Norwegian and international evidence on free school choice versus catchment for the Ministry's consultation, and Serediak & Helland (2020) simulate alternative intake models on Oslo's applicant data and their effect on segregation between schools. Our data are the thresholds those rules produce, which is why we present them without ranking schools (Section 11).

Cutoff prediction. Direct prior art is thin. The closest published work we know of predicts admission score lines in China's Gaokao system (Chen, Peng, Gao & Cai, 2022). Our setting differs in scale and structure — many small county-level queues with short histories rather than one national exam — which motivates partial pooling and makes calibrated probabilities, not point predictions, the deliverable.

Hierarchical models and shrinkage. Borrowing strength across small groups by partial pooling is classical (James & Stein, 1961; Efron & Morris, 1975) and is standard multilevel practice (Gelman & Hill, 2007). We apply it to a censored panel in which 396 of the 2,597 series with any numeric threshold have a single observation, with variance components estimated by an EM-type procedure (Dempster, Laird & Rubin, 1977).

Censored and two-part models. Treating "no queue formed" as censoring of a latent threshold descends from Tobin (1958); we instead use a two-part hurdle specification (Cragg, 1971), because the fill event is observed and economically distinct from the threshold's level, and because the application needs the fill probability as its own output.

Calibration and forecast evaluation. Post-hoc recalibration by a fitted logistic map is Platt scaling (Platt, 1999; Niculescu-Mizil & Caruana, 2005; Guo, Pleiss, Sun & Weinberger, 2017). Probabilities are scored with the Brier score (Brier, 1950), a strictly proper scoring rule (Gneiting & Raftery, 2007), and reliability diagrams (DeGroot & Fienberg, 1983; Murphy, 1973); interval forecasts are judged by the calibration-and-sharpness criterion (Gneiting, Balabdaoui & Raftery, 2007). Taking the predictive distribution from the forecaster's own past errors rather than from the model is an old idea in forecasting (Williams & Goodman, 1971) and is the principle behind conformal prediction (Vovk, Gammerman & Shafer, 2005; Lei et al., 2018); Section 6 is a stratified, non-exchangeable version of it, and Section 7.3 reports where its coverage guarantee holds and where it does not. Our temporal evaluation follows rolling-origin practice (Tashman, 2000; Bergmeir & Benítez, 2012; Hyndman & Athanasopoulos, 2021).

Open government data. The fragmentation documented in Section 4 is a known failure mode of public-sector open data — publication without machine-readability, stable identifiers, or retention (Janssen, Charalabidis & Zuiderwijk, 2012); Section 12 turns our error classes into concrete publishing recommendations.

4. Data

4.1 Sources and extraction

The sources are the counties' own publications — typically one PDF per county per year, occasionally per intake — together with tables released to us on request (Section 4.4). Each county has a dedicated extractor feeding a shared normaliser. PDFs are read by coordinate rather than by text flow: columns are sliced by x-position, and rotated column headers (Oslo, Vestland) are reconstructed glyph by glyph. Editions that counties overwrote or removed were recovered from the Internet Archive where it holds them, and where it does not, from a copy printed elsewhere: a newspaper that reprinted the county's list, a master's thesis, a document archive's page text of the county's own file. Those copies, and the tables that survive only as images, are typed by hand, and each typed file names in its header the document, where it was reprinted, the intake as the copy states it, and who typed and checked it (Section 4.4). "Every edition we could retrieve" is the honest description of the panel, not "all editions that ever existed": the panel starts in 2012, and the older editions the archive holds (Oslo's 2009 table, Hedmark's 2008–2011 columns) are kept under sources/ but not read. School names are matched to the national school register (NSR) within county and geocoded via NSR, the national address API, and the national place-name register, in that order. Where two publications disagree on a cell, the newer wins and the disagreement is retained in data/source-drift.json; a disagreement near a full grade point flags the school-year as uncertain, and the application says so in words. A bare integer below 8 in a cell is read as a fragment of a course code rather than a threshold. A printed figure between 0 and 10 is kept as printed. Karakterpoeng start at 10,0, a 1 in every subject, so such a figure means the last applicant admitted had grades missing; Oslo's intake office explained that an applicant with fewer than seven counting numeric grades is scored as the sum of their grades divided by 16, times ten, instead of 0. A printed 0,0 stays where the county documents what it means. A suite of 141 regression checks locks known failure modes: shifted year columns, implausible values, unmatched schools, county-specific quirks, and the decimal rule itself.

4.2 Register normalisation

Programme labels vary in spelling across counties and across years within a county. Every row is resolved against Grep: 2,930 of 2,936 rows carry a register code (the remaining six are International Baccalaureate, outside the register). One code is vigo's rather than Grep's: the Vg4 year of påbygging after a vocational qualification is offered as PBPBY4YK--, which Grep does not list. The classification into utdanningsprogram is therefore the state's own, not ours — an earlier keyword classifier misfiled gartnernæring (horticulture) under restaurant and food processing because the substring "ernæring" (nutrition) matched. County intake groups are finer than Grep codes (one school can run several queues under one code), so each row keeps the county's own label as its identity and carries the register code and official name as metadata. Levels follow the counties' tables: mostly Vg1, with Vg2, Vg3, a fourth-year Vg4 (in Rogaland and Vestland) and a combined Vg2/Vg3 group where the county prints one.

4.3 Cell semantics

A cell in a county table is not always a number, and the non-numbers are distinct events, not missing data. Table 1 gives the taxonomy, in the counties' own terms, and how each state enters the analysis.

Table 1: Cell states in county publications and their treatment. "Level" and "fill" refer to the two model components of Section 5.

One county's "ingen venteliste" is a published rule rather than an observed state. Møre og Romsdal's dashboard masks every figure of 25 points or less with * and legends it «alle kom inn, eller at laveste karakter var under 25» (everyone admitted, or the lowest score under 25); the extract behind it prints the number, and we apply the county's rule so that the panel shows what the county publishes (Section 4.4).

Two consequences are enforced throughout. The dataset is censored by construction, so averages of published numbers are averages over the queued subset, and the fill indicator carries the rest of the signal. And 0.0 is a trap: it prints like a number but is the bottom of the scale, not a height on it — including it in a mean drags a school's level toward zero for the opposite reason a genuinely low threshold does.

4.4 Coverage

Table 1b lists the panel county by county: the years held, the intake the county publishes, and how the cells divide among the states of Table 1.

Table 1b: The panel by county. "Intake" is the publication the figures describe, as stated by the county or its copy, or confirmed to us; "—" where the source does not say, with the years that differ in parentheses. Cell columns use the tokens of Table 1.

Innlandet's 2020–2022 tables and Akershus's 2024/25 and 2026/27 workbooks were released to us under freedom-of-information requests in August 2026 rather than published (Innlandet confirmed nothing older survives; the Akershus workbooks are 2. inntak only and were never published). Møre og Romsdal's extract — fifteen years behind the county's dashboard — was released in September 2026, and the county confirmed the figures describe the second, final intake. Vestland's 2014–2019 editions are its predecessors': Hordaland's Bergen-area studiespesialisering lists (the county's press releases, and for 2014 and 2015 two newspapers' copies of them) and Sogn og Fjordane's Vg1 table for 2018–2019.

The years before 2017 are the least complete part of the panel, and most of them are not the county's own files. Where the county map has changed since, they are the predecessor county's tables: Innlandet's 2012–2018 are Hedmark's (Oppland's were not found), and Akershus's and Buskerud's are the pre-Viken counties'. Where neither the county nor the Internet Archive still holds an edition, the figures come from a copy printed elsewhere — a regional newspaper that reprinted the county's list (Akershus 2015, Oslo 2013 and 2016, Rogaland 2012, Hordaland 2014–2015), a master's thesis (Oslo 2012), or a document archive's page text of the county's own file (Hedmark 2012–2018) — and Buskerud's 2012–2014 booklets print their tables as images. All of these are typed by hand; where two editions overlap they are compared cell by cell like any others (Section 4.1), the dataset marks the years that rest on a copy, and the application says so beside them. Most newspaper copies are excerpts — one region, or one or two programmes — and such a county-year is pooled out of the county's random walk (Section 5.2). Hedmark printed a grade average to one decimal, so its figures are whole points; its 2017 table marks only 14 programmes «ledige plasser» (free places), against 60 to 83 in its other years, and the two editions that print 2017 agree, so it is kept as printed. Rogaland's 2015 figures, from the county's school portal of the time, state no intake, and neither does the application for that year (Section 5.4). Rogaland's 2017 figures are the county's own: a news article of July 2017 that the county's site no longer serves and the Wayback Machine never captured, recovered from Common Crawl's capture of 20 July 2017. It names the round, 1. fellesinntak, and its table is «examples of programme areas with many applicants and long waiting lists» (57 figures, Vg1–Vg3), so the year is an excerpt like the newspapers' and is pooled out of the walk in the same way.

Of the 2,597 school×programme series that ever carry a numeric threshold, 396 have exactly one observed year. Among the 19,632 cells, 18,796 competed on points and inform the fill model, and 13,049 carry a numeric threshold and thus inform the level model. Møre og Romsdal's figures also carry the admitted mean (Gjennomkar), which no other county publishes; where it equals the threshold, one applicant set the figure. 65 of the county's 4,252 numeric cells are of that kind, and Section 7.5 reports what down-weighting them does. Møre og Romsdal's 1840 "ingen venteliste" cells are of a different kind from the other counties'. The extract behind the dashboard has no fill state of its own — every offered programme carries a number, down to 5.7 — but the dashboard the county publishes never prints a figure of 25 or less: it masks the cell with and legends it «Ruter markert med betyr at alle kom inn, eller at laveste karakter var under 25», and the extract reproduces the mask exactly (every starred 2026/27 cell has a figure of 25 or less, every unstarred cell more than 25; the legend says under 25, but an exact 25,0 is starred too). We apply the mask as the dashboard draws it: a figure of 25 or less enters the panel as "ingen venteliste", the state the county shows rather than the number it hides, which the county's intake adviser confirmed as the intended reading for Vg1 given the county's high fill rates. It is a proxy label. A programme with a queue whose cutoff was 24.6 is labelled open, and one that admitted everyone with a weakest applicant at 27 keeps its number; the share it produces, 30% of the county's cells, sits between Vestland's 6% and Innlandet's 43%. Section 7.5 measures what the proxy labels are worth — to the county's own fill forecast and to the seven counties whose labels are observed — and the county may link capacity data during 2027, which would replace the rule with the state. From 2 to 5 September 2026 the county was instead held out of the fill model with its fill probability fixed at 1; Appendix D records the change.

Telemark and Vestfold are outside this report and outside the model. Neither county publishes thresholds. Telemark released a workbook on request in September 2026 (Vg1, 2024–2026, eleven schools, 166 cells) that gives, for every offered programme, the lowest grade points among those admitted, down to 10.0. There is no fill state and no rule to read one from: where everyone who applied got a place the file still prints a number, so a low figure may be a real cutoff or no queue at all, and the county has not yet said whether applicants admitted with fortrinnsrett or after individual assessment are among "those admitted". Vestfold answered the same kind of request in October 2026 with the same kind of figure (Vg1–Vg3, 2024–2026, ten schools, 731 cells): the lowest grade points among those admitted for every offered programme area, no fill state, no intake round stated, and the county's request that the figures be read as a guide only, since places and search patterns move them from year to year. Those numbers are not poenggrenser in the sense of every other county in Table 1b, and fitting them beside the others would put figures of a different kind into the level model and the held-out scores. The two counties' cells therefore enter none of the fits this report evaluates, no walk-forward fold and no number in this report. The application still publishes their figures, with a note saying what they are, and still forecasts their schools: once the model of Section 5 is fitted, its intercept and its utdanningsprogram and programme-area effects are held fixed as an offset, and only each county's own school, series and county-year effects are estimated, on its own cells, with the fitted model's variance components (a satellite fit, one per county; Satellite in tools/model.py). The fill probability is pinned at 1, as for any source that carries no fill state, so a reader's chance there rests on the threshold alone. Nothing of that fit reaches the panel, the backtest or the evaluation. What it is worth is measured on each county's own years instead, by the same walk: the panel fitted on the other counties' years before T, the satellite fitted on the county's years before T, its year T predicted. For Telemark, over 110 cells in 2025 and 2026 that gives an RMSE of 7.78 points against 8.98 for persistence, and it is that number, not the panel's, that its forecasts carry as their spread — the panel's spread covered 63% of those outcomes where it claimed 80%, being measured on cells of a different kind. For Vestfold the same walk over 498 cells gives an RMSE of 5.89 points against 6.96 for persistence, and a spread of 5.9; there the panel's spread covered 78%. The application quotes each county's own measurement on every one of its schools. Each county joins the panel here when it states, per programme and year, whether everyone was admitted.

Agder and Nordland are in the dataset and outside this report for another reason. Neither publishes thresholds today, but both once printed them: Agder's intake office showed school counsellors its 2020 and 2021 Vg1 thresholds, and Aust-Agder a selection of 2016–2017 figures (19 schools, 204 cells in all), and Nordland printed every programme area in its yearly statistics booklet in 2013–2015 and 2019–2021 (21 schools, 2,144 cells). They are the only public record of those markets, and the application shows them with notes on what each figure is. They are not forecast: the newest is five intakes old, and the tables are of other kinds than the panel's. Agder's 2020–2021 figures carry a priority tier in the hundreds, which we remove by an arithmetic the county never states (the hundreds are always 0, 200, 300, 700, 800 or 900 and the rest always lies on the 0–60 scale); Nordland's 2013–2015 figures include an 800-point county supplement that its legend states and we subtract; and its 2019–2021 table gives the lowest points among those admitted with no fill state, as Telemark's does. Neither county is fitted, walked forward, scored or forecast, and no number in this report includes them.

The counties that publish nothing today either state that they choose not to (Agder, Nordland, Østfold) or publish aggregate statistics without thresholds. The unevenness is a property of the publication regime, not of the collection: for several counties and periods, county-wide thresholds were never computed, because admission ran on catchment rather than points.

5. Forecasting model

5.1 Notation and target

A series i is one school×programme×level queue; sc[i], p[i], a[i], and c[i] denote its school, utdanningsprogram, programme area, and county, and r(i,t) the intake its county published in year t. In year t, Q_it ∈ {0, 1} indicates that the cell filled (a queue formed), and y_it is the published threshold, observed only when Q_it = 1 and the last admitted applicant had points (y > 0). For an applicant with points x, the deployed quantity for the county's next publication year is

where π is the forecast probability that a queue forms, m the forecast threshold conditional on a queue, s the forecast spread (Section 6.1), and Φ_F the CDF of the standardised forecast error (Section 6.2). Equation (1) is the hurdle decomposition: either no one is turned away, or admission requires the cutoff to land at or below x. One state sits outside the decomposition: a cell that fills with the last admitted applicant holding no points (the 0.0 state of Table 1) counts as Q = 1 but contributes no y — admission there is certain at any score, and the calibration in Section 7.4 scores it as such; the empirical recalibration of π absorbs the approximation.

5.2 Two components, one hierarchy

The level component models the observed thresholds:

and the fill component the queue-formation event:

μ and ν are intercepts. Four terms are random effects with mean-zero Gaussian priors and estimated variances: school (α), programme area within level (γ), the series interaction u_i (a school can be strong in music and ordinary in electrical), and the county–year effect w. The programme-area effect γ is keyed by the county's own label and level, not by Grep code, because one code can cover several county queues (Section 4.2); the register enters as metadata, not as a pooling key. The utdanningsprogram effect β is an (essentially unpenalised) fixed effect — with only 16 levels it needs no pooling — and the round terms ρ are fixed too: in the fill model an unpenalised 3. inntak indicator, and in the level model a plug-in constant taken from the round bridge rather than an estimated coefficient (Section 5.4). The county–year effect follows a random walk,

so a county's market level moves smoothly and its latest state is the forecast for next year; an innovation spans two consecutive published years, so a publication gap is one step, not one step per calendar year. Sixteen partial county-years are pooled out of the walk. Seven are Vestland's 2014–2020: the Hordaland studiespesialisering lists of 2014–2019 (fifteen or sixteen Bergen-area cells a year), Sogn og Fjordane's Vg1 table for 2018–2019 (ten schools) and the Vg1-only 2020 publication with no "ingen venteliste" state. The others are the excerpts of Section 4.4 — Akershus 2012–2015, Oslo 2013 and 2016, Rogaland 2012 and 2017 — and Trøndelag 2024, the Trondheim region's table alone. (Akershus 2012–2014 is Røyken alone, from Buskerud's table while the school was in Buskerud.) Their cells still train the school, programme and series effects, but each county's partial years share one pooled level of their own outside the walk, so the county's level starts where its first full publication does, and one region or one newspaper's selection does not set it. The pooled partial level serves as a fallback for the level walk only where a county has no non-partial fitted year at all — Vestland in the 2020 and 2021 walk-forward folds, Trøndelag in the 2025 fold — while the fill walk keeps the neutral level there, because the partial years' fill labels are uninformative by construction. The level component is fitted on the 13,049 cells with a numeric threshold; the fill component on the 18,796 cells that competed on points (Section 4.4). The hierarchy exists to borrow strength: a series with one observed year inherits its level from the hundreds of comparable series around it — partially pooled toward its school, programme, and county means — instead of being trusted alone.

5.3 Estimation

Effects are estimated by penalised maximum likelihood — a ridge whose penalties are the inverse prior variances — with the variance components updated by a few rounds of the standard normal–normal EM approximation, in which each posterior is taken as diagonal (the posterior variance of an effect is computed from its own precision, ignoring correlation with the other effects). This is the usual first-order approximation; its known bias is to understate posterior variance and hence to shrink the estimated τ slightly toward zero, which we accept because the spread that the application shows comes from Section 6, not from these components. The residual σ is the RMS of the fitted residuals without a degrees-of-freedom correction, and each τ is floored at 0.3 points. The school, programme-area and series effects are separately identified only through their priors — a series effect and its school effect are exchangeable on a series-only school — so the random-walk levels carry a small ridge and the intercepts absorb the common level; the reported decomposition of Section 8.2 is therefore the penalised one, not a unique one. The pipeline is deterministic. Observations are exponentially down-weighted with age; the half-life of 4 years was selected by the backtest and mattered little (Section 7.5). A threshold that equals the admitted mean — one applicant's score, flagged in 65 Møre og Romsdal cells — enters the level fit at a weight the backtest chooses among {1, ½, ¼, 0}; it chose 1 (Section 7.5). Table 2 reports the fitted standard deviations.

Table 2: Fitted variance components (standard deviations) of the four random-effect families. Level components in points; fill components on the logit scale. The fixed effects (μ, ν, β, ρ) have no variance component.

Coupling the components — the level model's school effect entering the fill model as an offset, so "in demand" is one trait read two ways — is adjudicated by the backtest on every refit. Earlier builds rejected it; with the Innlandet backfill it wins on the calibration years, by a margin whose interval includes zero (Section 7.5), and the deployed hurdle is coupled.

5.4 Intakes

Within a county the published intake is constant, so it is absorbed by the county level and need not be known — which is why counties that do not state their intake remain usable. A year published from another intake inside a county's series is handled with fixed per-programme offsets taken from the round bridge (Section 8.1), so the random walk does not read an administrative dip as a market event. These offsets enter the level model as known constants — ρ in (2) is not estimated — and the affected cells are also excluded from evaluation (Section 7.1). Vestland 2023 was that case in every build up to v1.7, because the county's page offered only its

1. inntak file; the 1. inntak file was recovered on 5 September 2026, and no

county-year currently triggers the mechanism. The older years bring three intake differences it does not cover, because no bridge measures them. Akershus's 2012–2015 and Rogaland's 2012 and 2017 figures are 1. inntak in counties that publish 2. inntak today; all are excerpts, so the county's own pooled partial level (Section 5.2) absorbs the difference instead of the walk. Buskerud's 2012–2014 tables state 1. inntak where its current ones state nothing, and Rogaland's 2015 figures state no intake; these are whole county-years inside the walk, so whatever difference they carry is one innovation of it, taken across the gap to the county's next published year, and at the half-life of Section 5.3 they carry about a tenth of the weight of the county's newest year.

6. Uncertainty

6.1 Spread from walk-forward errors, not from the fit

A hierarchical model is sure of itself: the residual standard deviation is 4.6 points, but a next-year forecast also carries the uncertainty of every borrowed effect and of the market's next move — and for a one-year series the effects are mostly borrowed. The spread s in (1) is therefore not read from the final fit. It is the RMSE of the model's own walk-forward forecasts on the calibration years (Section 7.1), stratified by the history the series had at forecast time, and floored at the residual sd of the newest fit that saw no evaluation year (4.5 points, from the fit trained on data through 2024), so the history component can never claim to beat the model's own in-sample noise (Table 3); that component is then scaled by the band the forecast level falls in (Table 3b), by a factor for the series' level group, Vg1 or Vg2 and up, and by a factor for a series whose newest step was unusually large. An earlier version floored at the final fit's residual instead — a small leak of the held-out years into their own intervals, worth 0.4 points of coverage; both review passes flagged it, and the floor now uses only pre-evaluation data.

Table 3: Forecast spread s, estimated from walk-forward errors on the calibration years 2020–2024.

The gradient is the empirical price of borrowing: a never-observed series is forecast 6.8 points loose; four observed years buy the spread down to 5.1. History is not the only conditioning variable. A queue cannot outgrow its applicants' scores, so a forecast in the forties has less room to miss than one in the twenties, and v1.6 found its intervals too wide above 45 points and too narrow below 25 (Section 7.3). The spread is therefore the product of the history bucket's RMSE and a multiplier for the band the forecast falls in, the two fitted by backfitting on the calibration years — each is the RMSE of the errors standardised by the other — with the multiplier constrained to fall with the level (pooled adjacent violators, weighted by cells), so that six free numbers cannot chase the calibration years' noise, and clipped to [0.5, 1.5].

Table 3b: Level multiplier of the spread, by forecast band, from the calibration years 2020–2024.

The middle barely moves: a forecast above 45 points gets a band 25% narrower than its history alone would give, one below 25 a band 6% wider. Section 7.3 reports what that bought on the held-out years.

The last factor is the level group's. The two factors above are fitted on Vg1 and Vg2 and up together, and v1.17 found the result too wide for Vg1 and too narrow for the rest (Section 7.3). Each group's factor is the one that makes its own 80% band cover 80% of its calibration-year errors: a quantile rather than an RMSE, because the errors are heavier-tailed than a bell curve and the band is what the application draws. It is ×0.892 for Vg1 and ×1.001 for Vg2 and up.

A series whose two newest figures are more than 8 points apart misses by more. The first-visitor walkthroughs of 26 September 2026 asked for such steps to be flagged (Randaberg's Vg1 Teknologi- og industrifag went from 30,0 to 11,3, as Rogaland printed it), and the walk-forward says the band is too narrow there: without a factor of its own it covered 72.2% of held-out outcomes after such a step, and 80.9% of the rest. The factor is fitted by the same quantile rule, once on the calibration-year cells after a jump and once on the rest, so the pooled band still covers 80%: ×1.226 after a jump and ×0.978 otherwise (meta.sigma_jump_multiplier). Held out, the two cover 80.3% and 80.2%. About one forecast in six carries it, and the application flags the step itself as an unusual change.

Every county is forecast for one intake, the next that has not happened yet: the target moves on each 1 July, when the offers go out and before any county publishes that year's thresholds (forecast_intake, in tools/forecast_year.py). A county that has not yet published the intake just held is then forecast two years past its newest figure. The level is the same random-walk forecast; the spread gets a factor fitted by the same quantile rule on a second walk-forward that withholds the year before each test year: ×1.033 (meta.sigma_ahead_multiplier). Held out, the two-year forecasts miss by an RMSE of 5.92 points against 5.47 a year ahead, and their 80% band covers 79.9% with the factor and 77.8% without it.

6.2 An empirical error distribution

Φ_F in (1) is the empirical CDF of the standardised walk-forward errors (41 quantiles), not a Gaussian. It is mildly left-skewed — thresholds collapse more often than they jump, because a queue can vanish but cannot exceed its applicant pool — though on this data a Gaussian would have scored almost identically. The empirical CDF is clamped to [0.005, 0.995], so no single cell's probability is ever exactly 1 − π or 1.

6.3 Recalibrating the fill probability

The raw fill model is overconfident: in the calibration-year walk-forwards on all eight counties (Møre og Romsdal's proxy labels included), cells the raw model gave a mean π of 0.97 (the ≥ 0.9 bin) filled 94% of the time, and in the held-out years cells given a mean 0.96 filled 88% — the series effects fit the training panel too well. We therefore recalibrate by Platt scaling, fitting

on the calibration years only, which are disjoint from the held-out evaluation years. The slope well below 1 is a uniform confidence haircut. The map is one pair of coefficients for all eight counties; it transfers a calibration-year haircut to the held-out years on the assumption that the overconfidence is stable in time, which the held-out reliability table (Appendix C) bears out in the upper bins and not in the sparse 50–70% bins. Held-out Brier for the fill event: 0.157 against 0.208 for the base-rate forecaster (base rate 0.704; difference −0.052, cluster-bootstrap 95% interval [−0.059, −0.043], Section 7.2). The recalibrated π′ is still uneven in the mid-range (54% observed in the 50–60% bin, 67% in the 60–70% bin, but 78% in the 70–80% bin), which the deployed quantity (1) — the only probability shown to users — absorbs, as Section 7.4 shows.

7. Evaluation

7.1 Protocol

Evaluation is walk-forward with an expanding window (Tashman, 2000; Bergmeir & Benítez, 2012): for each target year t ∈ {2020, …, 2026}, the model is fitted on everything published before t and forecasts year t. Years 2020–2024 are calibration years: they set s (Section 6.1), Φ_F (Section 6.2), and the Platt map (5). Years 2025–2026 are held out entirely and are the only years reported in this section. For the held-out years, no information from 2025 onward touches the forecasts: model fits, spread, error distribution, Platt map, and the half-life were all fixed on earlier data. (Within the calibration years the selection is pooled across 2020–2024, as calibration always is.) One caveat is moot in this build: the round bridge of Section 8.1 is computed once on all paired publications, including 2025–2026, but its offsets enter a training fit only for a year published from another intake (Section 5.4), and with Vestland 2023 now read from its 1. inntak file there is no such year. Should one return, its cells are excluded from all scoring: no earlier year can teach any forecaster an administrative event, and grading the model on an offset it is told about would mislead in both directions.

Two things the protocol does not do. It scores the level model only on cells that published a number in the target year, so a series that collapsed to "ingen venteliste" is scored by the fill component and not by the level one; the level results are conditional on a queue surviving, and Section 7.4 is the unconditional view. And the base-rate baseline of the fill event uses the held-out base rate itself, an oracle that a real competitor would not have; the calibration-year base rate is within a point of it.

Baselines. For the threshold: persistence — the series' most recent published figure, at whatever lag — where the series has history, and the programme–county mean of prior years, where the programme has been published in the county before. For the admission probability: the deterministic step rule ("the last published figure is this year's cutoff", probability 0 or 1) and a probabilistic persistence forecast that centres the same spread s, error distribution Φ_F and recalibrated π′ on the last published figure instead of on m — so that the comparison isolates the model's point forecast from its uncertainty treatment.

Uncertainty on every comparison. Held-out cells are not independent: a county's year moves all its series together, and the two held-out years share the fitted structure. Every difference reported below therefore carries a 95% interval from a cluster bootstrap over school×year (415 clusters, 1,000 resamples), which respects the within-school co-movement that a plain binomial standard error ignores but still treats county-years as exchangeable within a school; with only fifteen held-out county-years it cannot do better, and the by-county coverage of Section 7.3 shows the size of what it leaves out.

7.2 Threshold accuracy

Table 4: Held-out threshold accuracy, 2025–2026 (2,730 cells with a published number), by history stratum. RMSE and MAE in points (↓ better); bias is mean forecast minus outcome; ±3 is the share of forecasts within three points (higher better); best per row in bold. Persistence is undefined for 0-year series, and the programme–county mean is defined for only 109 of the 146 0-year cells (the model's RMSE on those 109 is 6.72). EWMA is exponential smoothing of the series' own past figures with α = 0.4 (Muth, 1960); with one observed year it is persistence.

The model beats all three baselines on RMSE and MAE in every stratum where they are defined (Figure 1), and the margins over persistence are not sampling noise: the model-minus-persistence RMSE difference is −1.19 points [−1.70, −0.73] with one year of history, −0.97 [−1.39, −0.56] with two or three, and −1.04 [−1.26, −0.80] with four or more; against the programme–county mean the intervals are [−1.44, −0.53], [−1.02, −0.41] and [−1.13, −0.66]. The 0-year stratum, where the mean baseline exists for three cells in four, is the least certain: the difference, −1.25 points on those 109, has an interval [−2.47, −0.13] that stays below zero, if only just (it touched zero until v1.31). The ±3 hit rate tells a subtler story: persistence edges it in the deepest stratum (50% vs 49%), a difference whose interval [−0.04, +0.02] is centred on zero — repeating the last figure lands inside a narrow band about as often, and misses by more when it misses, which is the trade a squared-error forecast makes. The bias column shows where the borrowing costs: series with no history are forecast 3.0 points too low on average and one-year series 1.4 too low, because a new queue tends to open above its programme's county mean; from two years on the forecast is unbiased. Overall: RMSE 5.47 [5.26, 5.68], MAE 4.14, bias −0.22, 47% of forecasts within ±3 points. For scale, the same series moves with a standard deviation of 6.2 points between consecutive published years (8,570 pairs; 49% of moves within ±3), so persistence is a strong baseline and the residual year effect is large — which is why the deliverable is the distribution, not the point forecast. Per-year results are in Appendix B.

Persistence and the programme–county mean are the two ends of one family: the last figure with all the weight, or every past figure with equal weight. Exponential smoothing sits between them, and Muth (1960) showed it is the optimal forecast when a series is a random walk observed with noise — close to what a threshold is. With α = 0.4 (the weight on the newest figure; the rest decays geometrically) it is a far stronger baseline than either end on deep series: RMSE 6.03 against persistence's 6.73 with two or three years of history, and 5.29 against 6.18 with four or more. The model still wins, but by less — −0.27 points [−0.51, −0.04] and −0.14 [−0.26, −0.02] — and the two-to-three-year interval stops 0.05 short of zero. That is the honest size of what pooling across schools and programmes adds once a series can be smoothed on its own: most of the model's margin over persistence on long series is the smoothing, not the hierarchy. On short series the hierarchy is everything — with one observed year exponential smoothing is persistence, and the model beats it by 1.19 points.

7.3 Interval coverage

The nominal 80% interval m ± 1.2816 s contained the published figure 80.2% of the time on held-out cells (n = 2,730; cluster-bootstrap interval [78.5, 81.8]), at a mean width of 13.1 points; the 50, 90 and 95% Gaussian intervals covered 52.1, 89.6 and 93.7%. The deployed distribution Φ_F is the empirical one, and its central 80% band covered 80.3% (50/90/95: 50.1, 90.3, 94.6%): the Gaussian band lands on its nominal level, and the empirical quantiles, learned on the calibration years, transfer to the held-out years a little too tight. Coverage is the target and width the price (Gneiting, Balabdaoui & Raftery, 2007): roughly ±6.6 points is what an honest 80% claim costs on this data. The graded interval is the symmetric Gaussian one because that is what the application displays, and coverage is conditional on a numeric threshold materialising at all.

The marginal figure hides a conditional pattern, and v1.7 acts on it. Table 4b stratifies the 80% Gaussian coverage by the forecast level and by county. Errors are smaller where the forecast is high — the cutoff cannot exceed the applicant pool, so a programme forecast above 45 points has little room to surprise — and with the spread conditioned on history alone (v1.6) the intervals covered 95.6% above 45 points and 61.9% below 25. The level multiplier of Table 3b, fitted on the calibration years only, moves those two bands to 80.9% and 69.0% on the held-out years and leaves the three middle bands where they were (78.0–81.9% against 77.4–81.7% before); the overall coverage, the mean width and the admission-probability Brier score are unchanged (80.2%, 13.1 points, 0.0919 against 0.0920), which is what moving width from one end to the other should do. The 40–45 band, at 84.8%, is still too wide: the calibration years give it a multiplier barely below the middle bands', and the held-out years say it belongs with the top. Across counties, coverage runs from 72% in Trøndelag to 89% in Oslo; Møre og Romsdal, forecast from proxy-labelled cells, sits at 82%.

The pooled figure is two different ones averaged. The application shows Vg1 by default, and Vg2 and up are nearly half of the held-out numeric cells; Table 4c scores the two apart. Vg1 forecasts are more accurate (RMSE 5.36 [5.08, 5.65] against 5.60 [5.31, 5.89] for Vg2 and up). With a spread pooled over levels, Vg1 intervals were too wide and the others too narrow (held-out coverage 83.5% and 77.4%, the same split on the calibration years), so since v1.18 each level group carries its own factor of the spread (Section 6.1). The nominal 80% interval now covered 81.4% [79.2, 83.6] of Vg1 outcomes and 79.5% [77.1, 81.8] of the rest, and the Vg1 band is 12.4 points wide instead of 13.2. The factor is fitted to cover 80% on the calibration years (80.3% and 79.7% there by construction), so the held-out figures are the test. The application quotes the Vg1 coverage beside the Vg1 view and the pooled one when Vg2 and up are shown.

Table 4c: Held-out scores by level. Intervals are the cluster bootstrap of Section 7.1; fill Brier against the level's own base rate.

Table 4b: Held-out coverage of the nominal 80% interval, by forecast level and by county. RMSE and mean s in points.

7.4 Admission-probability calibration

The deployed quantity is (1). For every held-out cell and every score x ∈ {20, 25, …, 55} we ask "would an applicant with x points have been admitted?" — the outcome is determined by the published threshold and fill state — and score the predicted probability over all 31,544 score–cell pairs (3,943 cells), with π as deployed: Brier score 0.092 [0.089, 0.096]. The step rule is defined only where the series has a prior figure, 27,584 of those pairs; on that common subset the model scores 0.091 against the step rule's 0.156 (difference [−0.070, −0.060]) and the probabilistic persistence forecast's 0.096 (difference −0.005 [−0.007, −0.003]). The second comparison is the fair one: most of the model's advantage over the step rule is the uncertainty treatment of Section 6, which any centre could carry, and the model's own point forecast is worth a further 0.005 of Brier on top of it — small in absolute terms, but its interval excludes zero.

Table 5: Reliability of the held-out admission probability (Figure 2 shows the same data as a diagram).

The largest gap between prediction and outcome in any decile is 8.2 points, in the 60–70% bin, where the forecast is cautious: a stated 65% was realised at 73%, so the upper middle of the scale understates the chance rather than overstating it (the 70–80% bin, next, is cautious by 7.6 points). Below 60% the forecast is within 3.8 points of the outcome in every bin, with predicted minus observed at most -0.3 points in the three lowest bins — a stated 15% was realised at 17% — a region the application's coarse bands (likely ≥ 70%, possible ≥ 35%, otherwise unlikely) absorb in any case.

7.5 Ablations

Four structural choices that could have gone the other way were adjudicated on the calibration years, never the held-out years; the intervals are the cluster bootstrap of Section 7.1 applied to the calibration-year folds:

• Recency half-life. Over half-lives {1.5, 2.5, 4, ∞} years the calibration-year walk-forward RMSE was {5.758, 5.741, 5.739, 5.766}: 4 years wins, by 0.027 points over no decay [+0.012, +0.042] — kept because it wins, reported because the interval says the choice does not matter.
• Level→fill coupling. Plugging the level model's school effect into the fill model lowered calibration-year fill log-loss from 0.449 to 0.448 (difference [−0.003, −0.000]) on the current panel, so the deployed hurdle is coupled. On the held-out years the two variants score the same fill Brier (0.156 against 0.157). Every earlier build rejected the same coupling (0.403 vs 0.406 before the Innlandet backfill) — the verdict belongs to the backtest, this report documents the flip rather than smoothing it over, and the interval says the two hurdles are indistinguishable on this data.
• A county's proxy fill labels. Møre og Romsdal's "ingen venteliste" is the dashboard's rule of Section 4.4, not an observed state. Holding the county out of the fill fit instead — its π forced to 1, as deployed before the rule — moves the Platt slope from 0.661 to 0.586 and the held-out fill Brier on the seven counties whose labels are observed from 0.1584 to 0.1577: the proxy labels do not distort the other counties' calibration; they cost it a trifle. On the county's own 732 held-out cells the proxy-labelled hurdle scores 0.150 against 0.180 for its base rate, which says the rule is predictable — a low cutoff one year foretells one the next — not that it separates queues from empty places.
• Level-conditioned spread. Section 6.1's multiplier is fitted on the calibration years and judged on the held-out ones (Section 7.3): the top band's coverage falls from 95.6% to 80.9% and the bottom band's rises from 61.9% to 69.0%, at no cost to the overall coverage, width or Brier score. Kept.
• Single-applicant cells. A threshold equal to the admitted mean is one applicant's score. Over level-fit weights {1, ½, ¼, 0} for the 65 such cells the calibration-year RMSE was {5.739, 5.741, 5.743, 5.746}: full weight wins by less than a thousandth of a point, and the paired interval is [0.00, 0.00]. Sixty-five cells cannot move a backtest of this size, so the choice is a prior, not an estimate; the pipeline keeps the flag, the search and the full weight, and reports the verdict rather than the intuition.
• Pooled fit versus a Vg1-only fit. The school effect and the county walk are shared across levels, and Vg2 and up are about half the numeric evidence in the three counties that publish them. Refitting the whole walk-forward on Vg1 cells alone, with its own spread and fill recalibration, and scoring both on the same Vg1 cells: the calibration-year RMSE is 5.52 pooled against 5.54 Vg1-only (Vg1-only minus pooled [+0.000, +0.052]), and on the held-out years 5.36 against 5.31 ([−0.070, -0.011]). The other levels help the Vg1 forecast a little and never hurt it, so the pooled fit is kept and no school-by-level effect is called for. With its own spread, the Vg1-only fit's intervals cover 81.6% held out against the pooled fit's 81.4%: what was wrong with the Vg1 interval was the spread, not the pooling.
• Per-level spread factor. The history and band factors of Section 6.1 are pooled over levels. Scaling each level group to its own calibration-year 80% band moves held-out coverage from 83.5% to 81.4% on Vg1 and from 77.4% to 79.5% on Vg2 and up, and the Vg1 band from 13.2 to 12.4 points. The admission probability hardly notices: its Brier score goes from 0.0920 to 0.0919, because the empirical error distribution of Section 6.2 already absorbs most of a scale error. Kept for the interval's sake.
• Jump factor. Scaling the spread after a newest step of more than 8 points (Section 6.1) moves held-out coverage after such a step from 72.2% to 80.3%, and on the rest from 80.9% to 80.2%; the admission probability's Brier score goes from 0.0920 to 0.0919. Kept for the interval's sake, like the level group's.
• Trend on steadily rising series. The model has no trend term: a series that climbed four years running is forecast back towards its school and programme, sometimes 8 points below its last figure (Oslo's Persbråten, Studiespesialisering, 48,1 in 2026). The walk-forward has made 81 such Vg1 forecasts (at least three earlier figures, none falling over the newest four, forecast 4 or more points below the newest). The published figure came in below the last one 82.7% of the time; the model's RMSE there was 5.88 against 7.57 for persistence, and it under-forecast by 1.23 points [−0.04, 2.52]. The drop is real; that it is a little smaller than forecast is likely, though the interval now reaches zero. A trend term, the error regressed on the forecast's gap below the last figure with the slope (0.123) fitted on the calibration years, takes the held-out Vg1 RMSE from 4.891 to 4.827, and would lift the forecasts the walkthroughs questioned by half a point to a point; not added (meta.halflife_search.rising_series_check).

8. Findings about the publication practice

8.1 The intake-round effect

Two counties publish the same programmes in the same year from two intakes — a paired, within-year measurement of what a later intake does (Table 6).

Table 6: The round bridge: paired within-year threshold changes between published intakes. Standard errors of the mean in parentheses; "queues cleared" counts earlier-intake queues that no longer existed in the later one.

The drop conditional on the queue surviving is half the story; the cleared queues are the other half. The two rows are not one effect: Akershus's rests on two years and a step of one intake, Vestland's on nine years and two intakes (2017–2019 pair Hordaland's press releases with its full 3. inntak table), and what makes them similar in size — a first intake that over-offers by about the same margin in both — is a coincidence of two counties' practice, not a national constant. The effect is heterogeneous by programme — in Vestland, studiespesialisering −5.3 and påbygging −5.9 against electrical −1.7 and building −2.6 (full table in Appendix A). Cross-county comparison of raw thresholds that ignores the intake is therefore systematically misleading by 3–6 points — and several counties do not state the intake in their publication. We deliberately do not project a common "1. inntak equivalent" scale across counties: applying Akershus's offsets to Rogaland would be an assumption dressed as a measurement.

8.2 What a raw school mean measures

The school effect α_s in (2) is a school's threshold level relative to the same programmes elsewhere in its county — programme mix held constant. Writing a school's raw mean threshold as the mean of its fitted cells and attributing the between-school variance of that mean to the components of (2) gives a decomposition of what the map's colour actually encodes. Over the 194 schools whose α_s rests on five or more fitted cells (between-school sd of the raw mean 5.2 points), the school's own effect accounts for 44% of the variance, the programme mix (which utdanningsprogram and programme areas it offers) for 25%, the county-year level — which intake the county publishes, and its market that year — for 12%, the series interactions for 10%, and residuals and covariances for the rest. Ranked within their own county by α_s instead of by raw mean, schools move 3.2 places on average and at most 26; in Møre og Romsdal, where mix explains 55% and the school effect 34%, the average move is 6.3 places, in Oslo 2.5. A substantial part of a raw school mean is what the school offers and when its county publishes, not how contested it is — which is why the application prints the mix-adjusted effect with a standard error, states that it measures demand, not quality, and does not rank schools across counties at all (Section 11).

8.3 The model as a data audit

The cells the fitted model finds least plausible (|z| > 3: 106 of 13,049, 38 of them in Vestland, the county with the most cells; the 25 most extreme are published in the model metadata) were checked against sources. The largest deviation and two further Vestland-2022 extremes (Slåtthaug 18.0, Dale 12.5, Fitjar 48.8) appear verbatim in the county's own PDF — genuine extremes, not parse damage. One case remains open: Kongsberg's 2025 figure of 4.0 for musikk, dans og drama, following a "by documentation" cell in 2024, is printed so by the county and is kept as data. Residual screening of this kind is a continuous quality control of both pipeline and source, and would catch the header-shift class of parser bug if it recurred.

9. The deployed application

The map colours each school by the mean of its current-year thresholds and sizes each marker by the share of its programmes that filled — the censoring is displayed, not hidden inside the mean. One global filter follows the register's hierarchy (all → utdanningsprogram → programområde), with county row labels preserved as queue identities and official register names attached where spellings diverge. With the applicant's points entered, every row shows the probability (1); a choice list enforces vigo's real constraints — ten wishes, at most three distinct Vg1 utdanningsprogram — and reports the probability of at least one place as 1 − ∏_k (1 − p_k) under an explicit independence caveat. Where a county publishes 1. inntak but admits through 3. inntak (Vestland), the application also shows the chance by the final intake,

with v_p the share of 1. inntak queues cleared by 3. inntak and δ_p the surviving-queue drop, both per programme from Section 8.1 where a programme has at least ten pairs, and the county-wide values otherwise. The interface is bilingual (Norwegian/English), phrased in the official Udir/vigo vocabulary throughout, and the compiled dataset is available from the page as JSON and in the repository as CSV and SQLite. The shipped model carries 2,040 programme forecasts, of which 164 are for series with no observed year; for the two held-out counties of Section 4.4 a separate fit on each county's own figures supplies 310 forecasts for their 21 schools, scored nowhere in this report except against the county's own years in Section 4.4. Two things it deliberately does not forecast: a series whose newest cell is utgått (discontinued; 65 series) gets no forecast, whatever the year before said, and a series with no observed year is tagged "ingen historikk" (no history) — with "lite historikk" (little history) at a single year — rather than handed a bare percentage. Those 164 forecasts are in the exported files, tagged by their zero history, so that a reader can exclude them. Nor does it forecast the schools of the two counties with history only (Section 4.4): the application shows their figures, and says beside them why there is no chance of admission.

10. Limitations

• Marginal, not individual. The model forecasts the queue's cutoff, not an individual's outcome. It ignores that applicants compete only at their highest surviving wish, tie-breaking, and seats consumed by priority and documentation quotas.
• Stationarity. The random walk extrapolates a smoothly moving market. Capacity changes, new schools, catchment redistricting, and reputation shocks are structurally unforecastable from thresholds alone; the model will be wrong at exactly the discontinuities families most want warned about, and the spread s prices this only on average. The panel's own window contains one such shock — the 2020–2022 examination cancellations moved the whole points scale (Section 2) — which the county walk absorbed as a level move; a future shock of that kind would first appear as a one-year bias.
• The published number, in the published intake. Where a county publishes 1. inntak, more applicants are ultimately admitted than the figure implies; Section 8.1 measures the gap where it can be measured, and equation (6) passes it on only for the county where it is measured.
• Short test window. Two held-out years (2,730 cells with a number, 3,943 that competed) from one country and fifteen county-years. The cluster bootstrap of Section 7.1 prices the within-school dependence but treats county-years as exchangeable; the by-county coverage of Table 4b, from 73% to 89%, is the honest size of what it leaves out. 2026 looks better than 2025 partly because most of its cells are drawn from the counties with the deepest histories (Appendix B).
• Cold starts. Trøndelag has one whole published year after one region's, and Buskerud three recent years after a decade's gap, so their forecasts rest mostly on borrowed effects; Buskerud's intervals cover 73% instead of 80%, and Trøndelag's 2026 forecasts cannot be evaluated until the county publishes that year.
• Copies of lost tables. 14 county-years in the panel rest on a newspaper's, a thesis's or a document archive's copy of the county's figures, typed by hand (Section 4.4). Where a copy overlaps another edition the two are compared cell by cell; a single-year excerpt has nothing to be checked against, and a newspaper may have selected, rounded or mistyped what it printed.
• Independence in the choice list. The at-least-one probability treats wishes as independent; a hard year hits several of them at once, so the true probability is somewhat lower.
• Catchment counties. A threshold binds only residents of the intake area (inntaksområde); the application states this where it applies, and the model does not know where an applicant lives.
• A proxy fill state in one county. Møre og Romsdal's "ingen venteliste" is the county's own display rule — a Vg1 figure of 25 or less — not an observed queue state (Section 4.4). A programme with a queue whose cutoff was 24.6 is labelled open, and one that admitted everyone with a weakest applicant at 27 keeps its number. The rule's held-out fill Brier (0.150 against 0.180 for the county's base rate, Section 7.5) says the labels are predictable, not that they are right; the county may link capacity data during 2027.
• Self-selected counties. The eight publishing counties chose to publish; nothing here is evidence about the seven that do not, and any future expansion inherits whatever made them different.
• Regime changes. A programme can move between the points queue and the quota states (F/D/U) between years; the hurdle predicts queue formation, not that kind of administrative transition.
• Mid-range calibration. The raw probability is within 0 points of the outcome below 30% and cautious by up to 8 points in the 60–70% bin (Section 7.4); the application's bands absorb both.
• Reflexivity. A public forecast can move where people apply, which moves the thresholds being forecast — a Goodhart-type feedback. At the current scale we judge this unmeasurable, but it is the reason the application shows bands and history rather than a single authoritative number, and the effect would first appear as a calibration drift in precisely the walk-forward record this report is built on.

11. Ethical considerations

The unit of analysis is the queue, never the applicant: the data contain no personal information — only the per-programme cutoffs counties already publish — and the application collects none; points entered by a user stay in the browser. The county publications are public records and the fylkeskommuner remain their controllers; the copies mirrored under sources/ are reproduced for reproducibility, with the document metadata that named individual officials stripped. The tool informs applicants; it does not gatekeep, rank, or score any person, and it must not be used by institutions for admission decisions. Miscalibration harms are asymmetric — an overconfident "you will get in" can cost a pupil their safety choice — which is why the fill probability is deliberately recalibrated downward (Section 6.3), why coarse bands absorb the optimistic low tail, and why every forecast is shown next to the threshold's actual history. The mitigation itself has a cost: a conservative "unlikely" can deter an application that would have succeeded, and for a pupil that chilling effect is as real as the false promise — the history shown beside every forecast, and the 3. inntak chance where it is measured, exist to let a family see past the band. The interface is Norwegian and English only; families most disadvantaged by the current PDF-fragmented practice include those reading neither, and closing that gap is distribution work the tool has not yet done.

Thresholds are also a ranking in disguise. A map of cutoffs can be read as a league table of schools, and league tables of admission scores are known to feed the sorting they measure: Serediak & Helland (2020) show how Oslo's points-based intake concentrates high-scoring pupils, and the free-choice literature Sandsør (2020) reviews finds the same pattern where scores decide places. We therefore do not present raw school means as rankings, label the mix-adjusted school effect a measure of demand, not quality, and decompose in Section 8.2 how much of the raw mean is timing and mix rather than contest. Finally, information tools tend to benefit the already informed; publishing the compiled dataset and the tool free of charge is an attempt to level an asymmetry that today favours families with the resources to collect and interpret eight counties' publications by hand.

12. Recommendations to publishers

Each of these is cheap, and each would have removed an entire error class from this work:

1. Publish machine-readable tables (CSV/JSON) alongside the PDF, at a stable URL per year, so that an edition is neither overwritten nor lost when a site is rebuilt.
2. State the intake in the table — it moves the numbers by 3–6 points and is today often omitted.
3. Standardise cell semantics: distinguish "filled at 0.0", "ingen venteliste", and fortrinnsrett / documentation admissions explicitly, and print the filled/not-filled state even where every programme carries a number.
4. Print Grep codes — programme names are today spelled differently across counties and across years within a county.
5. Publish history, not only the latest year — the uncertainty in any advice to a family cannot be quantified without it, and several of this panel's older years exist today only because a newspaper once reprinted the county's list. A national table, compiled once a year from figures every county already computes, would cost the publishers nothing new.

13. Conclusion

A censored panel of published cutoffs, a hurdle model with partial pooling, and uncertainty measured from the model's own out-of-sample record turn eight counties' publications into a calibrated answer to the question families actually ask. The same assembly measures the publication practice it depends on: the intake-round effect, the decomposition of school averages into demand, mix and timing, and the residual audit are of more lasting interest than any single year's forecast — and each points to a concrete, low-cost improvement the publishing counties could make.

Reproducibility statement

All code for data extraction, normalisation, model fitting, evaluation, and the figures in this report is available at github.com/avshalomd/poengkart; the version this report describes is tagged report-v1.32, and the numbers quoted here are from the build of 2026-10-28. The compiled dataset ships in the repository as CSV and SQLite (data/, including the paired-intake cells of Table 6 as alternate-rounds.csv) and from the application as JSON; the original county publications are mirrored under sources/, including the releases on request, with a hash manifest and their provenance in sources/README.md, and a copy in object storage for the day a county page disappears. tools/refresh.py regenerates the dataset, the model, every walk-forward forecast with its outcome (data/model-backtest.csv), and every figure from the sources; the only network steps are the geocoding and the photo lookups, which are cached, so the model and the report rebuild offline. The fit is deterministic; the cluster bootstrap uses a fixed seed. The whole pipeline runs in minutes on a laptop. tools/test_docs.py pins every number in this report and in docs/model.md to the shipped model file, so a refresh that moves a figure fails the build until the text is updated; validation further comprises 141 parser regression checks and 18,040 model invariants. The dataset is released under the Norwegian Licence for Open Government Data (NLOD 2.0) and the code under the MIT licence.

References

• Abdulkadiroğlu, A., & Sönmez, T. (2003). School choice: A mechanism design approach. American Economic Review, 93(3), 729–747.
• Arteaga, F., Kapor, A. J., Neilson, C. A., & Zimmerman, S. D. (2022). Smart matching platforms and heterogeneous beliefs in centralized school choice. Quarterly Journal of Economics, 137(3), 1791–1848.
• Azevedo, E. M., & Leshno, J. D. (2016). A supply and demand framework for two-sided matching markets. Journal of Political Economy, 124(5), 1235–1268.
• Bergmeir, C., & Benítez, J. M. (2012). On the use of cross-validation for time series predictor evaluation. Information Sciences, 191, 192–213.
• Brier, G. W. (1950). Verification of forecasts expressed in terms of probability. Monthly Weather Review, 78(1), 1–3.
• Chen, X., Peng, Y., Gao, Y., & Cai, S. (2022). A competition model for prediction of admission scores of colleges and universities in Chinese college entrance examination. PLOS ONE, 17(10), e0274221.
• Cragg, J. G. (1971). Some statistical models for limited dependent variables with application to the demand for durable goods. Econometrica, 39(5), 829–844.
• DeGroot, M. H., & Fienberg, S. E. (1983). The comparison and evaluation of forecasters. Journal of the Royal Statistical Society, Series D, 32, 12–22.
• Dempster, A. P., Laird, N. M., & Rubin, D. B. (1977). Maximum likelihood from incomplete data via the EM algorithm. Journal of the Royal Statistical Society, Series B, 39(1), 1–38.
• Efron, B., & Morris, C. (1975). Data analysis using Stein's estimator and its generalizations. Journal of the American Statistical Association, 70(350), 311–319.
• Fack, G., Grenet, J., & He, Y. (2019). Beyond truth-telling: Preference estimation with centralized school choice and college admissions. American Economic Review, 109(4), 1486–1529.
• Gale, D., & Shapley, L. S. (1962). College admissions and the stability of marriage. The American Mathematical Monthly, 69(1), 9–15.
• Gelman, A., & Hill, J. (2007). Data Analysis Using Regression and Multilevel/Hierarchical Models. Cambridge University Press.
• Gneiting, T., Balabdaoui, F., & Raftery, A. E. (2007). Probabilistic forecasts, calibration and sharpness. Journal of the Royal Statistical Society, Series B, 69(2), 243–268.
• Gneiting, T., & Raftery, A. E. (2007). Strictly proper scoring rules, prediction, and estimation. Journal of the American Statistical Association, 102(477), 359–378.
• Guo, C., Pleiss, G., Sun, Y., & Weinberger, K. Q. (2017). On calibration of modern neural networks. Proceedings of the 34th International Conference on Machine Learning, PMLR 70, 1321–1330.
• Hyndman, R. J., & Athanasopoulos, G. (2021). Forecasting: Principles and Practice (3rd ed.). OTexts.
• James, W., & Stein, C. (1961). Estimation with quadratic loss. Proceedings of the Fourth Berkeley Symposium on Mathematical Statistics and Probability, Vol. 1, 361–379. University of California Press.
• Janssen, M., Charalabidis, Y., & Zuiderwijk, A. (2012). Benefits, adoption barriers and myths of open data and open government. Information Systems Management, 29(4), 258–268.
• Kapor, A. J., Neilson, C. A., & Zimmerman, S. D. (2020). Heterogeneous beliefs and school choice mechanisms. American Economic Review, 110(5), 1274–1315.
• Kirkebøen, L. J., Leuven, E., & Mogstad, M. (2016). Field of study, earnings, and self-selection. Quarterly Journal of Economics, 131(3), 1057–1111.
• Lei, J., G'Sell, M., Rinaldo, A., Tibshirani, R. J., & Wasserman, L. (2018). Distribution-free predictive inference for regression. Journal of the American Statistical Association, 113(523), 1094–1111.
• Murphy, A. H. (1973). A new vector partition of the probability score. Journal of Applied Meteorology, 12(4), 595–600.
• Muth, J. F. (1960). Optimal properties of exponentially weighted forecasts. Journal of the American Statistical Association, 55(290), 299–306.
• Niculescu-Mizil, A., & Caruana, R. (2005). Predicting good probabilities with supervised learning. Proceedings of the 22nd International Conference on Machine Learning, 625–632.
• Platt, J. C. (1999). Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. In A. J. Smola, P. Bartlett, B. Schölkopf, & D. Schuurmans (Eds.), Advances in Large Margin Classifiers (pp. 61–74). MIT Press.
• Sandsør, A. M. J. (2020). Fritt skolevalg? En gjennomgang av relevant forskning (NIFU-innsikt 2020:4). Nordisk institutt for studier av innovasjon, forskning og utdanning.
• Serediak, O., & Helland, H. (2020). Inntak til Oslos videregående skoler: Analyse av simulerte inntaksmodeller (Skriftserie 2020 nr 1). OsloMet – storbyuniversitetet.
• Tashman, L. J. (2000). Out-of-sample tests of forecasting accuracy: An analysis and review. International Journal of Forecasting, 16(4), 437–450.
• Tobin, J. (1958). Estimation of relationships for limited dependent variables. Econometrica, 26(1), 24–36.
• Vovk, V., Gammerman, A., & Shafer, G. (2005). Algorithmic Learning in a Random World. Springer.
• Williams, W. H., & Goodman, M. L. (1971). A simple method for the construction of empirical confidence limits for economic forecasts. Journal of the American Statistical Association, 66(336), 752–754.

Official sources and data

• Akershus fylkeskommune. Forskrift om inntak til videregående opplæring og formidling til læreplass, Akershus (FOR-2025-02-13-354). Lovdata.
• Innlandet fylkeskommune. Forskrift om inntak til videregående opplæring og formidling til læreplass, Innlandet (FOR-2024-12-17-3529). Lovdata.
• Oslo kommune. Forskrift om inntak til videregående opplæring og formidling til læreplass, Oslo (FOR-2025-01-29-147). Lovdata.
• Opplæringslova. Lov om grunnskoleopplæringa og den vidaregåande opplæringa (LOV-2023-06-09-30). Lovdata.
• Statistisk sentralbyrå. Table 07495: Grunnskolepoeng, etter fylke og kjønn. https://www.ssb.no/statbank/table/07495
• Utdanningsdirektoratet (2025). Fortsatt økning i søkere til yrkesfag (søkertall videregående opplæring 2025–26). https://www.udir.no/tall-og-forskning/statistikk/statistikk-videregaende-skole/sokere-vgs/sokere-til-videregaende-opplaring/
• Utdanningsdirektoratet. Grep (the national curriculum and programme register), read through its open API. https://data.udir.no
• Utdanningsdirektoratet. Nasjonalt skoleregister (NSR). https://data-nsr.udir.no
• Kartverket. Address and place-name APIs. https://ws.geonorge.no
• The counties' threshold publications and the releases on request, listed document by document in sources/README.md of the repository.

Appendix A: The round bridge by programme (Vestland, 1. → 3. inntak)

Table A1: Paired within-year threshold change from 1. inntak to 3. inntak in Vestland (2017–2020, 2022–2026), by utdanningsprogram; sorted by the size of the drop. "Queues cleared" counts 1. inntak queues that no longer existed in 3. inntak.

Appendix B: Held-out results by year

Table B1: Threshold accuracy per held-out year. Most 2026 cells come from the counties with the deepest histories (Vestland, Møre og Romsdal — whose extract reaches 2012 — and Oslo, joined since v1.3 by Akershus and Innlandet), which partly explains the better figures.

Appendix C: Fill-event calibration

Table C1: Reliability of the recalibrated fill probability π′ on the held-out years (3,943 cells that competed on points, all eight counties, Møre og Romsdal's 732 proxy-labelled cells included; base rate 0.704). Held-out Brier 0.157 against 0.208 for the base-rate forecaster.

Appendix D: Version history

• v1.0 (August 2026). First report: six counties, the hurdle model, the walk-forward evaluation.
• v1.1. Revised after two independent review passes: the spread floor moved to the last pre-evaluation fit (Section 6.1), the Vestland 2023 cells excluded from scoring.
• v1.2. Innlandet 2020–2022 added, released under a freedom-of- information request; the level→fill coupling first wins the backtest.
• v1.3. Akershus 2024/25 and 2026/27 added, likewise released on request.
• v1.4. Møre og Romsdal 2012–2026 added from the county's dashboard extract; the coupled hurdle deployed; Innlandet 2026; and the review of 2 September 2026: discontinued programmes no longer forecast, Møre og Romsdal's fill probability fixed at 1, partial county-years excluded from the county walk, the decimal rule of Section 4.1.
• v1.5. After two further review passes and a data re-examination: every held-out comparison carries a cluster-bootstrap interval; a probabilistic persistence baseline; the admission probability scored with the deployed fill probability, and the fill calibration reported on the seven counties that carry a label; coverage by forecast level and county; the raw-mean decomposition of Section 8.2 replacing a cross-county re-ranking; the paired-intake cells exported; Section 2 sourced; related work extended to school-choice information, Norwegian school choice, conformal prediction and open data.
• v1.6. Møre og Romsdal's fill state, from the county's own dashboard rule (a Vg1 figure under 25 is published as "ingen venteliste", Section 4.4): the county enters the fill model, its fill probability is fitted rather than fixed at 1, and Section 7.5 measures the proxy against holding the county out. The rule removes 1840 cells from the level model and the county's low-side outliers with them; the admission probability's largest decile gap moves from the low bins to the 70–80% bin.
• v1.7. Three model changes, each judged by the backtest: the forecast spread conditioned on the forecast level as well as on history (Table 3b; the top band's held-out coverage from 95.6% to 80.9%); the admitted mean (Gjennomkar) carried through the dataset for Møre og Romsdal, with a backtest-chosen level-fit weight for the 14 thresholds one applicant set (it chose full weight); and exponential smoothing (Muth, 1960) as a third baseline in Table 4, which shows that most of the model's margin over persistence on long series is smoothing, not pooling.
• v1.8. A data release, no model change: Vestland's 2023 series switches from the county's 3. inntak file to its recovered 1. inntak file, so the intake exception of Section 5.4 is no longer exercised; the round bridge grows from four to nine years (2017–2020, 2022–2026: Hordaland's full 3. inntak table for 2017–19 and Vestland's unlinked 2022/23 3. inntak file; Table 6, Appendix A); Sogn og Fjordane's 2018–19 Vg1 table and Oslo's 2015/16 table join the panel; one Vestland programme spelt two ways and Rogaland's mislabelled year headers are repaired at the parser, retiring the hand-kept "uncertain years" flag. The level multiplier's monotone fit now separates the 40–45 band from 25–40 (Table 3b); held-out coverage is unchanged to the decimal.
• v1.9. A data release, no model change: Rogaland's 2026/27 thresholds, published on vilbli on 7 September 2026 as the rolling 2024–2026 edition, join the panel (413 cells, all Rogaland; every 2024 and 2025 cell the new edition reprints agrees with the previous one). The model is refit and the walk-forward backtest re-run on the enlarged panel, and the quoted numbers are re-pinned; the model's edge over exponential smoothing on two- to three-year series now has an interval that touches zero (Section 7.3), which the text says.
• v1.9.1. A refit on the maintainer's Mac, with no change to data or code. The v1.9 numbers came from a build on another machine, and a refit here moves the last displayed digit of many quoted values (the reliability bucket counts by one to eight cells, the Platt slope by 0.0006, no forecast by more than a tenth of a point) because the fit is resolved only to about 0.3 %: three outer passes of L-BFGS-B at scipy's default tolerance on a ridge objective that is flat around its optimum, so where it stops depends on the floating-point path. A one-ulp change in the inputs moves more values than the change of machine did. Every headline figure is unchanged (half-life 4.0, 80 % coverage 0.801, chance Brier 0.093, spread 4.56), and no conclusion moves; digits beyond that resolution are this build's, not reproducible ones. The quoted numbers are re-pinned to this build.
• v1.10. Telemark 2024–2026 added from the county's workbook, released on request on 17 September 2026: eleven schools, 166 Vg1 cells, a number for every offered programme and no fill state. The county trained the level model, sat outside the fill fit with its fill probability fixed at 1, and was scored in the backtest; every pinned number was re-read from that build.
• v1.11. Telemark is taken out of the model and out of this report, and stays in the published dataset and the application (Section 4.4). Without a fill state its figures cannot tell a cutoff from a programme where everyone got in, and v1.10's fit showed what mixing them in does: the held-out RMSE of series with one year of history rose from 5.88 to 6.20 points, the county's intervals covered 70%, and the county-year term's share of the variance between school means rose from 9% to 17%. The model, the panel and every quoted number are back to the eight counties of v1.9.1, and the refit reproduces that version's numbers. The county's schools keep a forecast in the application, from a satellite fit on its own figures that leaves the model untouched (Section 4.4).
• v1.12. The held-out county's forecasts stop borrowing the panel's spread. A QA sweep measured what the borrowed number was worth on that county: its 80% intervals covered 70% of the outcomes, because the panel's spread is measured on cells of a different kind. The satellite is now walked forward on the county's own years the way the panel is walked forward on the panel's (Section 4.4), and its forecasts carry that measurement — 7.8 points against the panel's 4.9 to 7.3 — with the coverage quoted in the application. Nothing in the panel, the backtest or any number of this report moves; this version also corrects Table 1b's year span for Oslo, which had omitted the 2015 table since v1.8.
• v1.13. Rogaland's figures are checked cell by cell against the county's own PDFs before the application is introduced to the county's lower secondary schools. Every published cell matched its source; the pipeline around them had six defects. The county reissued its 2024–2026 edition on 21 September 2026 with one more 2026 figure; an edition recovered through the Wayback Machine, kept under sources/ but never read, is read now and is the only print of eight 2024 cells; a level label printed as two words («Vg 4») and two programmes printed under two spellings are read as one. Rogaland goes from 3,801 to 3,808 cells. The Vg4 påbygging rows of Rogaland and Vestland carry vigo's code for that year instead of the Vg3 one (Section 4.2). Every pinned number is re-read from this build; no rate or error moves by more than its last digit, and no conclusion changes.
• v1.14. Vestland prints its påbygging after a vocational qualification at level 3 in two editions of three; those ten rows now carry vigo's Vg4 code as well, decided by the programme's name rather than the printed level, and one parser check guards it. Nothing in the panel, the backtest or any number of this report moves.
• v1.15. The same Vestland programme, printed at level 4 in 2022/23 alone, is one series at level 3, the level of the county's other editions and of its current table: eight series gain their 2022 figure and the dataset has eight rows fewer (2,308). Every pinned number is re-read from this build. The fit moves by hundredths of a point; the smallest calibration band (45 points and above, 67 held-out cells) moves most, from 80.3% to 79.1% coverage, and no conclusion changes.
• v1.16. Rogaland printed 3,0 for one 2026 cell, Bergeland's Vg2 Medier og kommunikasjon, below the lowest possible score; the county answered on 23 September 2026 that there were free places, and the cell now reads «ingen venteliste» through a correction recorded in the extractor with the county's answer (one more parser check guards it). The panel loses that one numeric cell. It was a held-out zero-history cell, so the zero-history row of Table 4 moves most: RMSE from 8.49 to 8.25 and the programme–county mean's from 10.14 to 9.36. Overall RMSE goes from 5.60 to 5.58, and no conclusion changes.
• v1.17. The backtest scores Vg1 and Vg2 and up apart (Table 4c) and adjudicates a Vg1-only fit against the pooled one (Section 7.5); the pooled fit wins on the calibration years and is kept. Vg1 intervals over-cover (84.6%) and the other levels' under-cover (75.4%). Table 4c and the ablation were measured on a refit whose headline scores match the build of 2026-09-23 to the displayed digit; no other number of this report moves.
• v1.18. Each level group carries its own factor of the forecast spread (Section 6.1): held-out coverage of the 80% interval goes from 83.5% to 81.4% on Vg1 and from 77.4% to 79.5% on Vg2 and up, and 80.6% to 80.5% overall; the admission probability's Brier score does not move. Every forecast's spread changes, so every pinned number is re-read from the build of 2026-09-25, the first built in a Linux container from the pinned tools/requirements.txt rather than on the author's Mac: numbers the change does not touch still move in their last digit (the fit's resolution, Appendix D v1.9.1), and no conclusion changes.
• v1.19. A series whose two newest figures are more than 8 points apart carries its own factor of the forecast spread (Section 6.1): held-out coverage after such a step goes from 72.2% to 80.3%. Section 7.5 adds the backtest of steadily rising series, which finds no case for a trend term. The point forecasts do not move; every spread does, so the pinned numbers are re-read from the build of 2026-09-26, made in the Linux container from the pinned requirements like v1.18's, and no conclusion changes.
• v1.20. Older years join the dataset: the predecessor counties' own tables (Akershus and Buskerud from 2012, Hedmark 2012–2018 into Innlandet, Hordaland and Sogn og Fjordane 2014–2019 into Vestland) and, where a county's own document is lost, copies of it in newspapers, a master's thesis or a document archive (Oslo 2012, 2013 and 2016, Rogaland 2012, Akershus 2015); Rogaland 2015 comes from the county's school portal, Rogaland 2017 from the county's own news article as Common Crawl captured it, and Trøndelag 2024 from the Trondheim region's table. Agder and Nordland join the dataset as history-only counties, shown in the application but neither fitted nor forecast (Section 4.4). The dataset grows from 228 schools in nine counties to 278 in eleven and from 12,993 cells to 17,615, and the level model fits 9,914 cells instead of 8,256. A county's year excerpts share one pooled level per county instead of a step of the walk each (Section 5.2). Spellings that differ between editions of one county's table join one series, and so do two programmes the register renamed under an unchanged code (Helse- og sosialfag, Produksjon- og industriteknikk). On the unchanged held-out years (2,193 cells) the RMSE moves from 5.58 to 5.54 and the 80% interval's coverage from 80.1% to 80.6%; the county-year step's spread grows from 0.93 to 1.01 points, and the admission probability's Brier score does not move. No conclusion changes.
• v1.32 (this version). Trøndelag's 2026/27 tables, published on vilbli by 28 October 2026 under new file names, join the panel: 178 cells over the five intake regions, every one checked against the list version vilbli serves beside each table. Five schools that had admitted everyone in their newest year now print a figure. The model is refit and the walk-forward re-run, and the quoted numbers are re-pinned: the level fit grows from 12,969 to 13,049 cells; on the held-out years the level RMSE moves from 5.45 to 5.47 points, 80% coverage from 80.4% to 80.2% and the admission probability's Brier score from 0.093 to 0.092. One reading changes: on 0-year series the model's edge over the programme–county mean, −1.25 points, now has an interval [−2.47, −0.13] that stays below zero, where v1.31's touched it (Section 7.2). Trøndelag is forecast a year ahead again, for the 2027 intake like every county.
• v1.31. Vestfold 2024–2026 added from the county's workbook, released on request on 28 October 2026: ten schools, 731 cells at Vg1–Vg3, the lowest grade points among those admitted for every offered programme area, no fill state and no intake round stated. Like Telemark's it is published and held out of the model and this report (Section 4.4); its schools get their own satellite fit, which over 498 cells of 2025–2026 scores an RMSE of 5.89 against 6.96 for persistence. The panel, every fit and every number quoted from them are unchanged.
• v1.30. On 21 October Akershus corrected its 2026/27 page: Nadderud's Vg1 Idrettsfag reads 46,30 after 2. inntak, as the workbook the county released on request had it, so the 48,0 that v1.29 kept is gone, and Nadderud's Vg1 Studiespesialisering reads 51,3 after 1. inntak where the page had printed 46,3. The round bridge keeps its 209 pairs and moves from −3.12 to −3.15 points (sd 3.2 to 3.1). The level fit stays at 12,969 cells; on the held-out years the level RMSE stays at 5.45 points, 80% coverage at 80.4% and the admission probability's Brier score at 0.093. No conclusion changes.
• v1.29. Akershus put its 2026/27 poenggrenser page up on 20 October, with 1. and 2. inntak side by side, and the live watch captured it. Its 2. inntak figures match the workbook the county released on request in one exception: Nadderud's Vg1 Idrettsfag reads 48,0 where the workbook had 46,3, and the published page is kept. Its 1. inntak column is new, so the round bridge (Table 6) now pairs two years, 2025 and 2026: 209 pairs instead of 101, a drop of −3.1 points (se 0.2) instead of −3.4, and 14% of 248 queues cleared instead of 16% of 124 (the abstract's range, which still read −3.2 to −3.4 and 16–32%, now quotes both rows of Table 6). The level fit stays at 12,969 cells; on the held-out years the level RMSE stays at 5.45 points, 80% coverage at 80.4% and the admission probability's Brier score at 0.093. No conclusion changes.
• v1.28. A data correction caught by the live watch: on 6 October Buskerud reprinted its 2026/27 table with Buskerud vgs's Vg1 Restaurant- og matfag as «Alle» where it had printed «4,0», a figure below 10,0 that was already kept out of the fit. The level fit stays at 12,969 cells; on the held-out years the level RMSE stays at 5.45 points and 80% coverage at 80.4%, the top band's held-out coverage moves from 94.0% to 95.5%, and the admission probability's Brier score stays at 0.093. No conclusion changes.
• v1.27. Møre og Romsdal's «ingen venteliste» now follows the dashboard as it is drawn: it stars an exact 25,0 as well as the figures under it, although its legend says «under 25», so 34 Vg1 cells that read 25,0 are now «ingen venteliste». The level fit drops to 12,969 cells; on the held-out years the level RMSE on series with four or more years of history moves from 5.15 to 5.14 points, 80% coverage from 80.5% to 80.4%, and the admission probability's Brier score stays at 0.093. Trøndelag's coverage (72 cells) moves by one cell, to 72%. No conclusion changes. In the app, a programme whose only figures are below 10,0 now counts as near-certain, as one with «ingen venteliste» in every year does.
• v1.26. Two more data corrections found by an independent re-extraction, no model change. From 2025/26 Møre og Romsdal's report labels Vg4 toppidrett «Vg3», so those rows overwrote or posed as the Vg3 ones (Molde 2026 Vg3 Realfag, toppidrett read 28,6 for 33,8); the level now comes from the Grep code. Høyanger's eight 2026 3. inntak cells, which the county files under a shortened school name, were lost and are back. The level fit is 13,003 cells, the held-out set 2,652, the Brier score stays at 0.093 and no conclusion changes. The parser suite gains two checks (139).
• v1.25. A data correction, no model change: where one Buskerud cell shared a figure among several queues and one of them was «Alle» (no waiting list), the parser now splits it into one series per programme variant instead of reading a single figure. Drammen's Idrettsfag toppidrett 2026/27 now has four series (one of them «ingen venteliste») and St. Hallvard's Musikk, dans og drama 2026 reads musikk 25,0 with dans and drama «ingen venteliste». The level fit is 13,001 cells instead of 12,999; the dataset holds 19,452 cells instead of 19,447, the fill model 18,616 instead of 18,611, and the held-out fill cells 3,763 instead of 3,758. The parser's regression suite gains one check (137). The Brier score stays at 0.093 and no conclusion changes.
• v1.24. A printed figure between 0 and 10 is read again, now with its meaning. Oslo's intake office explained that an applicant with fewer than seven counting numeric grades is scored as the sum of their grades divided by 16, times ten, so such a figure is the last admitted applicant's real score, not an error. The nine cells v1.23 dropped are restored as printed (Section 4.1, Table 1), and the model counts each as not filled, like "ingen venteliste": a full grade record scores at least 10,0, so everyone with one was admitted. The application shows the figure with a note and leaves it out of every average, as it does «ingen venteliste» and 0,0. The level fit is unchanged at 12,999 cells; the dataset holds 19,447 cells instead of 19,440, the fill model 18,611 instead of 18,604, and the held-out fill cells 3,758 instead of 3,755. The Brier score stays at 0.093 and no conclusion changes.
• v1.23. A data release, no model change: Buskerud's 2026/27 Vg1 thresholds, from the county's own page (which does not state the intake round, like the county's earlier years), join the panel — 78 cells. A printed figure between 0 and 10 is now dropped as an error in the county's table, since karakterpoeng start at 10,0: nine cells go, among them Kongsberg's 2025 and Buskerud videregående skole's 2026 «4,0». Buskerud holds 473 cells instead of 397 and the dataset 19,440 instead of 19,369. The model is refit and the walk-forward backtest re-run, and every quoted number is re-pinned: the level model fits 12,999 cells instead of 12,938, and the held-out years hold 2,648 cells instead of 2,583, with Buskerud's 2026 cells among them (RMSE 5.45, the 80% interval covers 80.6%; the admission probability's Brier score is 0.093). Buskerud, which no longer lacks the newest intake, is forecast one year ahead instead of two, so only Trøndelag carries the widened spread of Section 6.1; the county's held-out coverage moves from 69% to 74%. The model's under-forecast on steadily rising series (Section 7.5) shrinks to 1.23 points, and its interval now reaches zero. No other conclusion changes.
• v1.22. Every county is forecast for the same intake, the next that has not happened yet, moving on each 1 July (Section 6.1): Buskerud and Trøndelag, which have not published 2026/27, are forecast for 2027/28 instead of 2026/27, two years past their newest figure, with the spread widened ×1.032 as measured by a walk-forward that withholds the year before each test year. No other figure changes.
• v1.21. Møre og Romsdal's Vg2, Vg3 and Vg4 join the dataset, from the county's own Power BI dashboard, which the live source watch now reads (Section 4.4): 4,268 cells for 2012/13–2026/27, with the same * rule and the admitted mean, so the county holds 6,058 cells instead of 1,790 and the dataset 21,883 instead of 17,615. The level model fits 12,938 cells instead of 9,914, and 65 cells carry a threshold one applicant set instead of 14. On the held-out years (now 2,583 cells) the RMSE is 5.45 and the 80% interval covers 80.7%; the admission probability's Brier score is 0.093. Two readings move: the largest calibration gap is now the cautious 60–70% bin rather than the 70–80% one (Section 7.4), and Møre og Romsdal's proxy labels now cost the other counties' held-out fill Brier a trifle instead of sharpening it (Section 7.5); the jump factor is refitted at ×1.222. Figures quoted from model.json are re-read from this build; no other conclusion changes.