Shipping: the names a ship already uses
Ships increasingly name their sensors by ISO 19848, with DNV's Vessel Information Structure. On 12,295 of DNV's published example names that it had not seen, the reader decoded 12,295, refused 8 of 8 malformed ones where a shape check alone refused 3, and placed 87% of the readings on their machine. The other 13% measure something it has no word for yet, and it says so instead of guessing. Decoding is close to certain by construction, because the name carries its own fields.
Before an EU MRV or FuelEU report, a fuel check compares each day's fuel meter with the load on the engine it feeds. On a simulated meter with faults planted in it, it found 142 of 145 and listed 0 of 29 clean days. On a real ship, the training ship ARC Gloria, it called no period ready, because the published data has no fuel meter. What that ship did show: its published record covers 51% of the period, and the main engine's oil pressure is missing for 142 hours while the engine ran.
P49, written down before the held-out run: SUPPORTED, within the limits above.
Then on real fuel meters: three fishing ships, a month of engine data each, published by the DataBio research project. On the ship held back, the check as built listed 2 of 27 clean days and found 149 of 150 gross faults planted into the real data (frozen, dropout, zero, a unit error, 30% off), where a rule on the day's total fuel found 35 of 138. Two things it got wrong. A meter 15% off went unseen (0 of 50): on real days fuel against shaft power varied by a robust spread of 1.2%, and its 25% margin is far wider; with a 10% margin, fixed on the other two ships, it found 52 of 52. And both days it listed were the engine idling with no load on the shaft, which it called fuel while the engine stood. That was wrong. Where the export has the engine's own speed, the check now uses it, with the 10% margin, and calls idling what it is. A person still reads the list before the report goes.
P52, written down before the held-out run: SUPPORTED.
District heating: can the meters be trusted, and can it find faults
The pack reads substations, heat exchangers and heat meters, supply and return, in Norwegian, Danish, Swedish, German and English, and checks each heat meter by the heat balance: power is flow times the temperature drop, and a meter's total is its power over time. It also checks the data behind the yearly figures a district heating company in Norway reports to SSB on form RA-0232: heat produced, heat delivered and the loss between them.
We tested whether its checks find faults the operator itself reported, on 46 substations held back from development. They did not. It found 9 of 33 reported faults; a plain rule that flags the weeks with the hottest return found 21, by flagging 47% of all weeks. Most reported faults were no heat or too little heat, faults in how the substation runs, where these checks are built for instruments that lie; that is our reading after the fact, not a result. What the pack does today: say whether the meters can be trusted. Finding faults in the substation itself would be a new experiment.
P50, written down before the held-out run: FALSIFIED.
Data centres: the energy report, and what it rests on
Data centres with at least 500 kW of IT power report their energy and water figures to the EU every year (Delegated Regulation (EU) 2024/1364). The pack reads a data centre's meters, cooling and air conditioners, here those of the Marconi100 supercomputer room in Bologna, whose operator published them. On a month it was not written on, it placed 97.6% of 2,169 documented readings right and 1.0% wrong, where a lookup by name alone placed 0.1%. That connector was written from this site's own documentation, so it shows what such a connector buys on the site it was written for, not how it reads a site it has never seen.
Before the report, it checks that the room's total equals the sum of its parts (99.95% of complete steps did) and finds a frozen total, a silent IT meter, a unit error or a meter reading high when they are planted (96% or more of each). And it says how much of the period the energy figure actually rests on: in that month, 25%, because one small meter in the sum logged only a quarter of the time. That is the number a person should see before signing the report. The ratio of facility to IT energy on those steps was 1.47, for one room of a larger site; it is not that site's official figure.
P51, written down before the held-out run: SUPPORTED, within the limits above.