24 KiB
24 KiB
Changelog
All notable changes to this project will be documented in this file.
Unreleased
- Pre-registration of what is tested and what it takes for the hypothesis to hold (
FORHANDSREGISTRERING.md). - Source bank: 948 entries in 21 areas, each checked against its primary source (
kilder/). The second round (areas 13–21) covers the parameters that lacked a source when the model was specified. - Model specification, draft 3 (47 equations), after a pre-read of draft 2, two independent reviews of draft 3
and a control of the corrections (
modell/). - Parameters frozen before anything is computed: 124 parameters in 2025 kroner, one distribution each, with the
source entries they rest on (
modell/parametere.json). The checksum is recorded in the pre-registration. No code and no results yet. - Eight clarifications of the pre-registered criteria, approved by the project owner after the parameter freeze
and before anything is computed (
FORHANDSREGISTRERING.md, «Presiseringer»). The parameter file is unchanged. - First code (
src/tokenfabrikker/): one function per equation for the power plant, the data centre and the revenue (S1–S21, S36, S43–S47), a strict reader for the frozen parameter file with unit conversion, and the 10 000 Monte Carlo draws stored indata/. Python with pinned numpy and scipy, pytest, ruffALLand mypy strict. No result for the hypothesis has been computed. - An independent reader coded the key equations from the specification alone. The two implementations agree on
all 23 functions. The 19 ambiguities the reader found are settled in a new, dated chapter 16 of the
specification (
modell/); the parameter file and the pre-registration are unchanged. - Known-answer tests against Argonne, NOU 2026: 4, Epoch and NVE with the sources' own assumptions
(
tests/test_fasit.py). The misses are kept as documented expected failures: wind on land and offshore wind come out 5–6 % below NVE, and three rows of Argonne's table depend on a reading the source does not settle. - Build-out, power system and cash flow in code (
src/tokenfabrikker/): when each facility starts and which vintages of IT hardware it gets (S22, S33), the export rule with its own stored stream of random numbers (data/aarshendelser.npy), the fitted price lines and the power and capacity balances (S23–S25, S38), and the yearly cash flows with residual value, the additionality rule and net present value for Norway (S26, S34, S35, S40, S45). Tested on made-up numbers only; no result for the hypothesis has been computed. - A second independent reader coded these equations from the specification alone and found 28 ambiguities. They
are settled in clarifications 21–40 in chapter 16 of the specification. With the readings aligned, the two
implementations agree on 10 000 made-up build-out paths. The clarifications that are choices are marked and put
to the project owner (
modell/fagfellevurdering-utkast-2-og-3.md, part 5). - The price lines of S38 are held against NVE's twelve scenarios: within 5 % for Norway and three price areas, 6.0 % and 7.7 % off in one scenario for NO3 and NO4, kept as documented expected failures.
- Business economics, ownership, scale, environment and verdicts in code (
nivaaisert.py,eierskap.py,skala.py,miljo.py,dommer.py; S15, S19, S27 to S32, S37, S42), tested on made-up numbers only. A third independent reader coded the same chapters from the specification alone and found 60 points, 14 of them objections to the economics; a new reader then recoded from the clarifications in four more rounds (17, 5, 1 and 0 new points). The readings stand in clarifications 41–61 in chapter 16; the two implementations agree on 1 000 made-up paths. Corrected after the readers: the accident cost stays when the additionality rule binds, tax depreciation loses real value with price growth, and the special tax only meets return above the cost of capital. Weaknesses in the comparison of ownership forms that are not corrected are listed in clarification 61 and put to the project owner (modell/fagfellevurdering-utkast-2-og-3.md, part 6). No result for the hypothesis has been computed. - The comparison of ownership forms corrected before anything is computed (clarifications 62–67 in chapter 16): a residual value at the final year as a taxable sale, the electricity tax as the state's only levy with what displaced private data centres would have paid deducted, private owners who pay what their share is worth in expectation, the state's cash flow valued with the required return for compute, and the cost of capital before tax. An independent reader coded the corrections from the text alone in three rounds; its first-round finding that the private owners and the state were priced with different rates changed the correction. With it every ownership form has the same expected value in the main run, and they differ in risk and capital. The two implementations agree to within one krone on 1 500 made-up paths. The project owner approved the corrections before anything is computed.
- The direction of the project extended by the project owner (what can start quickly, a stepwise build-up, and
compute on the users' own machines), and four more clarifications of the pre-registered criteria (9–12), all
before anything is computed (
FORHANDSREGISTRERING.md). The sources' figures for the emissions test M5 (i) are set up and checked against the primary sources (clarification 68). The parameter file is unchanged. - Source bank extended with areas 22–26 for the addition to the direction (188 entries: fast start on the grid,
power with short lead time, stepwise build-up, compute on the users' own machines, avoided emissions and
cooling water), each checked against its primary source by an independent controller (
kilder/). - The addition to the direction written into the specification as chapter 17 (equations S48–S54) and into the pre-registration, before anything is computed: a new solution T, "stepwise", in two versions (modules of 100 MW on the grid with new renewable power, and the same with grid-connected nuclear power from when it can come), which A must also beat; no gas power; a threshold and two labelled scenarios for compute on the users' own machines; the calculation of emissions with a deduction; cooling water; and a timeline of megawatts in operation per solution. An independent reader coded the chapter from the text alone in three rounds; its findings changed the chip rule and the condition for nuclear power in the second version, the cap on early grid connections and the lead time. The parameter file is unchanged; the new quantities are candidates until the deviation from the freeze.
- One deviation from the parameter freeze, before anything is computed (
modell/parametere.jsonversion 1.1, new checksum in the pre-registration): thirteen new quantities for the addition to the direction, seven new sensitivities and three corrected lines in the list of sensitivities. None of the 124 frozen parameters got a new value, range or distribution. 137 parameters, 91 drawn quantities, new draws indata/. Before the deviation the project itself looked up the 30 source entries behind the new quantities in the primary sources; four things changed from the candidates (see chapter 17.8 of the specification). - The build-out of solution T in code (
src/tokenfabrikker/utbygging.py, chapter 17.1): when each module of 100 MW starts in T-F and T-K, with the grid balance of the first stage, the condition for nuclear power (S49) and the two stocks of chips, and the timeline of megawatts in operation per type of power (S50). The equations S48–S54 are pure functions, and the new quantities are read into the input data. Written from the text before the independent reader's code was opened, then run against it: 42 000 made-up build-outs and 20 000 made-up sets for S51–S54 without a difference. Six clarifications (70–75) in chapter 16 of the specification say how chapter 17 is read where the text was silent. Solution T is not taken through the cash flow or the verdicts yet, and no result for the hypothesis has been computed. - Solution T through the power system and the cash flow (
kontantstrom.py): everything per module, renewable modules with their own portfolio as in C, the nuclear modules of T-K on shared units as in B, and the fixed cost of a new nuclear country counted to the first unit. The loss of demand of chapter 17.3 enters the revenue and the disconnection arm of S47 (regn_program(..., bortfall=...)); without a loss every figure is unchanged. Emissions "with deduction" (S53) and cooling water (S54) are put together over a programme for all seven solutions (miljo.py). A new independent reader coded the same from the specification alone: 1 778 made-up paths through both codes without a difference, and the two sensitivities that are switches in both. Six clarifications (76–81) in chapter 16 settle what the text left open (modell/fagfellevurdering-utkast-2-og-3.md, part 11). The levelised cost, ownership, scale and verdicts refuse solution T and a loss of demand until chapter 17 is taken through them. No result has been computed. - Present values are now exactly rounded sums. As dot products the same computation did not give the same last digit every time, for every solution. A test runs the same computation six times and requires equal numbers.
- Solution T and the loss of demand through the levelised cost, the business economics, the ownership forms
and scale (
nivaaisert.py,eierskap.py,skala.py): the power cost per module, with a renewable module's own portfolio as in C and the nuclear modules of T-K on shared units as in B; tariff, building, IT hardware and residual value per module; and the factor of S51 per vintage in S19, S27, S31 and S42. For the five solutions of chapters 0 to 16 without a loss every figure is unchanged to the last digit. A new independent reader coded the same from the specification alone: 1 760 made-up paths through both codes, 6.6 million figures compared, without a difference. Six clarifications (82–87) in chapter 16 say how the text is read and name one property the reader found that is in favour of nuclear power on the grid in the business economics (modell/fagfellevurdering-utkast-2-og-3.md, part 12). M5 and the verdicts do not know solution T yet. No result has been computed. - The technologies of a facility are kept in a fixed order. As a set of names their order differed from run to run, and the last digit of the figures for solution C with it.
- M5 and the verdicts for five solutions (
miljo.py,dommer.py): T-F meets M5 (i) and (ii) by definition and T-K is judged as nuclear power; the ladder, part I and M7 cover A, B, C, T-F and T-K, and in M7 A must beat each of the four others that can be built. Three sensitivities of chapter 17.8 are switches: full displacement in the first stage (regn_program(..., full_fortrengning_i_forste_trinn=True)), new renewable power first from 2031 for C and T (lag_plan(..., ny_kraft_fra=2031)), and modules of 50 and 200 MW (Inndata.med). Next to S54 the withdrawal of water is also given with sea water only for the nuclear production the data centres themselves use. With the switches off every figure for all seven solutions is unchanged to the last digit. A new independent reader coded the same from the specification alone: about 400 000 comparisons of M5 and the verdicts on invented outcomes per draw, and 2 523 made-up paths with 10.1 million figures through both codes for the sensitivities, without a difference. Seven clarifications (88–94) in chapter 16 say how the text is read (modell/fagfellevurdering-utkast-2-og-3.md, part 13). No result has been computed. - Clarification 17 of the pre-registration, approved by the project owner 2026-11-13 before anything is computed: the conditions of a power solution can be named when every target it holds under conditions in has a condition that can be named, and when M5 holds under conditions the condition is that the sources in which nuclear power is not higher are the ones relied on, written out with the source. The independent reader found that the text did not say this. The project owner also decided that the grid connection of a nuclear plant stays out of the business economics, as the text says, and that the report states it (clarification 87).
- The run over the draws (
kjoring.py): for every draw the seven solutions on every step of the ladder in six scenarios (three price scenarios and the three marked scenarios for AI use on the users' own machines), with the outcomes the verdicts need; the verdicts per target with the shares they are set from, the step they are read on, part I and part II; the timeline (S50) as median, percentiles and shares over the draws; and the ownership forms on that step. Clarifications 95–103 in chapter 16 settle what the text did not say: what is computed per draw, that a path that cannot be built has no outcome while a set of values the build-out rejects stops the run, that the step is set once in «normalisering» and used in every scenario, which paths the ownership forms are compared on, how the percentiles are computed, and that part II has no verdict when no step can be built. The run gives the same figures to the last digit on one, three and four processes, with five different hash seeds and with solutions, steps and scenarios in reverse order (2 240 made-up paths). A new independent reader coded the same from the specification alone: 120 made-up draws through both codes (3 360 paths, 373 000 figures) and 600 runs of invented outcomes per draw (1.9 million figures and 545 000 verdicts, steps and years) without a difference (modell/fagfellevurdering-utkast-2-og-3.md, part 14). The run has only been given made-up draws. No result has been computed. - Clarifications 18 to 20 of the pre-registration, approved by the project owner 2026-11-13 before anything is computed: part II gets no verdict when no step of the ladder can be built, and the report says so; the step with a single facility cannot be the 25 % step, because the independent reader found that the rule let a weaker single facility give a better verdict (this makes part I easier to meet in that one case, and is changed now because it could not be changed credibly once the figures exist); and the ownership forms are compared for each of the five solutions on the step the verdicts are read on. The change of the step rule (clarification 104 of the specification) is not coded yet.
- The rule for the step is coded (clarification 104), and the rest of what had to be settled before anything
is computed is written into the specification and coded (clarifications 105–114): the price threshold (the
factor the price of compute must be multiplied by for the median net present value to be zero, found to
within 1e-6), its check against the contract prices, the threshold for lost demand (S52), the thresholds in
one drawn quantity behind «holder under betingelser», the rule for when the conditions of part I can be
named, first-order Sobol indices estimated from the stored draws, and the three threshold surfaces
(
terskler.py,sobol.py,kjoring.py). Indices, thresholds in a quantity and two of the surfaces are read from the same outcomes as the verdicts; only the price threshold runs the model again. A new independent reader coded the same from the specification alone: 383 000 comparisons on invented outcomes and 3 900 through both model codes on made-up draws without a difference, apart from one invented case no run can give. The reader found that the rule for naming conditions is too easy to meet when the share lies just below 80 %: quantities without any effect then give a threshold by chance. The project repeated the measurement. A stricter rule is proposed to the project owner and is not coded; neither is the reading that M6 needs no threshold approved yet (modell/fagfellevurdering-utkast-2-og-3.md, part 15). The sensitivity list is not coded through the run. No result has been computed. - Clarifications 21 to 23 of the pre-registration, approved by the project owner 2026-11-13 before anything is
computed: M6 is exempt from the requirement of a named condition (this makes part I easier to meet in one
case, and is said so); a threshold counts as a condition only when the quantity demonstrably affects the
outcome (a chi-square test at the 0.1 % level on the stored draws); and a threshold in the required return
for power must hold in both halves of the draws by the required return for compute, an assumption without
a source it is correlated with. The last two make the test stricter. Coded in
terskler.pyandkjoring.py, with the conditions of a solution taken together shown next to them (clarification 115 of the specification). - The sensitivity list is settled line by line in the specification before it was coded (clarifications
116–124) and coded through the run (
folsomheter.py): the 37 lines of the frozen list give 42 rows, each the main run with one change over the same draws, read on the main run's step and shown in one fixed table next to the main run. Eleven rows needed new switches in the model code (the markup for nuclear power, the unit size in the dedicated plant, the income factor by standard terms, lasting and fading scarcity, the stepped discount rate, the tax cost, the reserve above the dimensioning fault, efficiency by ownership form); with the switches off every figure is unchanged to the last digit. The state's cash flow next to M6 is coded. A new independent reader coded the same from the specification alone: 474 000 comparisons of the rules and the table on invented outcomes and 4 300 outcomes of the 37 run rows through both model codes on made-up draws without a difference. The reader found one place where the text gave two different numbers (which outage rates size the reserve units when the unit size is chosen freely), now settled, and that the row with standard contract terms can favour the dedicated plant; the project owner approved the readings 2026-11-13, and the row carries a note that says so. One figure differs between the two codes, by at most 6e-5: the reader does not scale decommissioning after the final year in the efficiency row, and the specification says it is scaled (modell/fagfellevurdering-utkast-2-og-3.md, part 16). No result has been computed. - The tools of chapter 12, fixed in text before they were coded (clarifications 126 to 133) and tried on
made-up numbers only: everything a run needs is one basis (
grunnlag.py); the main run writes six data files in a fixed format without rounding (datafiler.py); and the command line (verktoy.py) hashovedkjoring,kontroll,scenario,folsomhet,terskel,sammenlignandarbeidskopi, with input in JSON where an unknown field is an error.kontrollwrites the files again and says «identisk», «innenfor toleranse» or «endret». A made-up basis with 200 draws and the files written from it are stored with the tests; they are reproduced byte for byte with one and with eight processes and in a new process. The tools cannot yet run on the frozen parameters: the EEA assessment per ownership form is not written down. No result has been computed. - The map from equation to function (
modell/kart.md), generated from the code and held against it by a test: each of the 54 equations points to at least one function. - The project owner changed the hypothesis on 13 November 2026, before any result: it is that AI data centres
owned in principle by the state can become a source of income for Norway as oil has been. A dedicated
nuclear plant per data centre is one of the power solutions that are compared, on equal terms. The targets
M1 to M6 and their thresholds are unchanged. The texts and the code that give that solution a verdict of its
own are rewritten before anything is computed (
FORHANDSREGISTRERING.md, «Endringer»). - The power solutions are compared on equal terms (clarifications 135, 137 and 138), written before the code
and before any result: for each of the five solutions the share of the draws in which it has the highest
net present value alone and the share in which it shares the place, with no verdict of its own on dedicated
nuclear power, and no part II. A weakness in the first rule was found on made-up draws and corrected: the
two stepwise solutions are the same path where no module gets nuclear power, and those draws were counted
as «none best». An independent reader coded the rule from the text alone: no deviation in 25 000 made-up
runs. The surfaces now carry figures for the classes along their axes (clarification 134), and
kontrollprints a line per file as it goes. The EEA assessment of the ownership forms is proposed as «usikker» for all four (clarification 136) and awaits the project owner. - The project owner approved, before anything is computed, the new wording of the pre-registration after the changed hypothesis and the EEA assessment of the ownership forms (2026-11-14). The pre-registration now has one hypothesis with one verdict and, next to it, the comparison of the power solutions with no verdict; the old wording is kept under «Endringer». The assessment «usikker» for all four ownership forms is a fixed table in the code, so the EEA assessment takes no form out. A net present value that is not a number is rejected in the comparison (clarification 139). Nothing has been computed from the frozen numbers.
- Source bank extended with area 27 (90 entries in two files): what researchers and expert bodies say about series building and learning, a series across the Nordic borders, large against small reactors, the timeline for new nuclear power, long power purchase agreements and the state's role. One researcher and one independent controller per file (76 checked: 45 confirmed, 31 corrected, 14 added). Made after the parameter freeze and before any result; no parameter is changed. The review log (part 19) holds the findings against the frozen assumptions and names two that the frozen ranges do not cover.
- The main run on the frozen numbers was started 2026-11-14, the first time anything is computed. Before it:
the project owner decided that neither of the two findings from the source round on expert views gives a
deviation from the freeze (both stand as known limitations), and clarification 140 fixed how the run is made
(one command writing the files one at a time, from a separate copy of the code, after a technical trial on
16 draws) and a seventh data file for distributions over the draws (
fordelinger.json, not yet coded). Review log part 20. A first draft of the submission to the national AI strategy, with placeholders for every result, is ininnspill/. - A command for the two data files made next to the main run,
tilleggsfiler(clarifications 141 and 142):fordelinger.jsonwith the percentiles over the draws that the figures need, andmiljo.jsonwith the life-cycle emissions, the emissions "with deduction" and cooling water per year, which the pre-registration requires reported and no other file holds. Written and tested on made-up numbers while the main run was going, before any result was read; the six files of the main run are not touched. An independent reader held the written chapters of the report against the pre-registration (no verbatim text deviates; six sentences corrected), and the three knowledge gaps in the draft submission were held against the later source rounds. Review log part 21. - The main run on the frozen numbers is done: the six result files are in
data/resultater/(10 000 draws), with their checksums in review log part 22. The hypothesis does not hold in any of the six scenarios; the verdicts are read on one facility because the scale is not reached. The draft submission was held against the files once more (four figures corrected), shortened to under 1 000 words, and its wording approved by the project owner. The reproduction of the files (kontroll) and the two extra files are not yet made.