A ship's fuel figure for the year is a sum of readings nobody looks at between bunkerings. If a flow meter froze for a day in March, or read a little low all spring, the figure the verifier receives is built on it, and nothing in the figure shows it. EU MRV lets a company measure its fuel with flow meters, or with the bunker delivery notes and soundings of its tanks (Regulation (EU) 2015/757, Annex I). Either way, someone has to ask whether the numbers hold together. We asked that of three kinds of real ship and of a marine engine on a test bed, on data their publishers have made open, and wrote down what it finds and what it misses.
Three fishing ships, and the names they use
The first set is one month of three fishing ships, every ten seconds: the main engine's fuel meters, the shaft's power, the pitch, the temperatures, five generators (DataBio, Zenodo 3563390, CC BY 4.0). The publisher does not name the ships, so neither do we. Given only their column names and readings, Auge answered 83% of the 234 columns on the two ships we held out, and 191 of its 194 answers were right. Before this work it answered 24%. We wrote the new rules on the first ship, and the other two use almost the same names, so this part is close to true by construction. That is said in the report, not hidden in it.
A floor under the fuel
The check we had compared each day's fuel with the ship's own usual ratio of fuel to shaft work. It is good at a meter that freezes or jumps for a day. It cannot see a meter that is wrong by the same factor all month, because the usual ratio moves with it. So we added a floor that does not depend on the ship at all. The most efficient four-stroke diesel engine on record burns 165 grams of fuel per kWh at its best point (Wärtsilä, 2015). A meter that shows less than 160 g/kWh for the work the shaft measured is showing less fuel than the best such engine on record would have burnt.
318 of 320 day faults foundfrozen, dropped out, zero, scaled, offset and swapped meters planted on the two held-out ships; fixed limits on the same meters found 117
Over a whole month the floor found a meter reading half, or 60%, of the truth on both ships, a fault the usual ratio cannot see. It did not find a meter 15% low, and it does not see over-reading at all: an engine can burn more than its best for many honest reasons. Here fixed limits did better: the supply and return meters carry a nearly constant flow, so their narrow band catches any scale change on them. That is a cheap check worth adding, and the report says so. On clean days the check listed 1 of 15 days on one ship and 1 of 27 on the other, and neither was a meter fault. The preregistered verdict is SUPPORTED.
A tanker's fuel log, voyage by voyage
The second set is the fuel log a chemical and oil tanker's crew kept at every event, 2020 to 2022: the counters of each fuel meter, the fuel remaining on board, the fuel burnt and each bunkering (Ulutaş and Bilgin Güney, Zenodo 23089007, CC BY 4.0). We reconciled each voyage as a verifier would: the fuel by the counters against the fuel by the tanks and the deliveries.
0 of 61 voyages reconcilein 2021 and 2022, within 5% or half a cubic metre
The reason is in the log itself. On 38 of those voyages the main engine's counter does not move, and the diesel generators' and the boiler's stand still on nearly all of them, while fuel goes on being written down every row. And the fuel remaining on board is a ledger, not a sounding: on 1,649 of 1,665 rows it is simply the row before minus the fuel logged. So the tanks and the log are one figure here, and the counters, the one independent figure, stopped. A verifier would want to know that before reading the total. Auge lists it per voyage, with the reason.
We also planted errors in the counters, the log and the tank figures. Auge found 88 of 90. The preregistered verdict is still FALSIFIED: the naive check we had promised to beat flagged every voyage whose meter moved at all, because most voyages' meters did not, so there was nothing left to compare on. The rule was written before the run, and the verdict stands as it fell.
A training ship's hand-written rounds
The third set is a test of names nobody wrote our rules for: the hourly rounds of the training ship ARC Gloria, typed in from the crew's sheets (Llamas Reinoso, Martinez-Santos and Puertas, Zenodo 18307281, CC BY 4.0). A column called Fuel there is a fuel pressure, and Cylinder_1 is a temperature. Auge answered 13 of 73 columns and got 13 right, and it took none of them for a fuel meter. But a ten-minute word list answered more columns correctly, and with the units written in, our new ship rules placed fewer columns (9) than the general reader had (13). The verdict is PARTIAL.
We have changed that since. A name the ship rules have no word for now goes to the general reader, machine and all, and a unit written in the header overrules a guess from the name. On the same rounds with the units written in, Auge now answers 25 columns rightly, against 9 before, with 0 wrong. That is the ship the change was written for, so it proves little. The test is a ship it was not written for, and that is the next section.
A marine engine on a test bed
The fourth set is one none of the earlier work touched: a three-cylinder marine diesel engine on a test bed in Tokyo, run through a reference program and 15 runs with faults introduced on purpose, such as a clogged air filter or a blocked injector nozzle (BahooToroody and colleagues, Zenodo 19857425, CC BY 4.0). It is an engine, not a ship at sea, and its names are a laboratory's. The test was written down before we ran it.
93 of 93 answers righton the engine's names, as published and with the units in brackets
What Auge said was right. It said too little. With the units in brackets the ship rules were chosen, and they answered 50 of 143 columns, where the general reader on its own answers 61: the change above did not close the gap P62 found. A ten-minute word list got 93 right, because every name here says what it measures. And the row of units under the names was not read at all (0 units read, against 50 of 50 right when the same units were in brackets). We built those next; they are below.
99 of 102 planted meter faults foundon single test runs; fixed limits on the same meter at the same false-alarm budget found 35 of 114
The fuel check listed 2 of 19 runs with nothing planted, and both were the runs with a clogged injector nozzle, which burnt 28% and 41% more fuel per kWh than the other runs. Those are real engine faults, and on shift that is worth knowing the day it starts. The check cannot tell a meter that reads high from an engine that burns more, though, and the test counted them against it. The publisher's dictionary gives the fuel meter in cubic metres per hour, and its values only make sense in litres: taken at their word, every run sat at least 1,459 times over the floor, so no under-reading could fall below it. The check had no ceiling then, so it did not question the unit either.
The verdict is FALSIFIED: PARTIAL for the names, FALSIFIED for the fuel. It is in the report as it fell.
Three fixes, and a test we could not run
P74 named three things, and we have built them. A law that ties a column both to the main engine and to the shaft it turns no longer makes it a tie: it goes on the engine, and the reason says so. The row of units under the names is read. And there is now one generous ceiling, 1 litre of fuel per kWh at the shaft over the whole record, where the engines in these tests burnt 0.24 to 0.26 litres per kWh. It is our own margin, set far out of an engine's reach, not a published limit. A record above it is a question about the unit for a person, with the units that would make sense. Nothing is rescaled.
On the test bed's own export, which the fixes were written with in view, Auge now answers 77 of 143 columns from the header as published, against 43 before, with 0 wrong, and reads 77 of 77 units from the units row. The fuel meter is paired with the shaft in 2 of 2 files, and the ceiling names the unit: in cubic metres per hour the engine would burn 264 litres per kWh, and the values only make sense in l/h. That proves little, because the fixes were made for this export. Rerun after the fixes, P49, P52 and P60 give exactly the same results as before.
Not testedP78: the fixes on ship data none of the earlier tests touched
We wrote down the test before looking for its data: a real ship or marine engine, open, with a licence that allows commercial use, with a fuel meter and a measured power on one clock, that none of P49 to P74 had touched. We checked 12 candidates, and none qualified: the good ones are non-commercial, restricted, simulated or without a licence. So the fixes stand untested on new data, and we do not count them as proven. The test is written and waits for a ship's export, and a pilot's export would do.
What it gives the person who files the figure
In the console, a ship's export now has a Fuel and emissions tab, in Norwegian and English. It shows each day's fuel by the meter beside what the shaft's work implies and the floor, lists the days to look at with the reason, and explains a listed day from the readings around it. Below it is a draft of the period's fuel figure: what it rests on, and what is left empty because the export cannot give it, such as the fuel type and the density from the bunker delivery note. Nothing is filed and nothing leaves the machine. Whether a ship is in scope is not in the export either: EU MRV applies to ships above 5 000 gross tonnage and not to fish-catching ships (Article 2), so the fishing ships here are a test, not a customer's report. When the fuel per kWh over the whole record cannot be right in the unit the export gives, the tab says so before the figure, and leaves the figure as a question rather than a number to report.
The tab now also holds the supply and return flows to a narrow band. A circulating fuel supply carries a nearly steady flow while the engine runs, so a day when it moves more than a tenth is listed. On the fishing ships, fixed limits like this caught a whole month of every meter reading 15% low or 30% high, which the floor missed. Against the export's own usual day the band finds a day that changed; to find a whole period that changed, you type in the pump's flow or a period you know was good.
What this does not show
These are a handful of ships, one month to a few years each, and the faults are planted: they say what the checks can see, not how often real meters fail. No bunker delivery note or tank sounding was in any of the data, so a meter that over-reads by the same factor all year is still found only against those. The full reports, with every number and every limit, are P60, P61, P62, P74 and P78, which says why its test is not run yet.