Skip to main content
Case studya mid-sized Middle East carbonate operator

The Folder Said March 2020: Dating the First Renders of a Multi-Year AI Programme

An earlier version of this case study argued that our first renders of the operator's own borehole predated the contract by 21 months, because they sat in a folder named March 2020. The files say otherwise: the renders are dated 9 March 2022, the well beside them was logged in September 2021, and no client well data in the archive is dated before January 2022. The first render came three months after kickoff. This is the corrected record, and what it says about dating evidence.

18 Mar 20265 min read
The Folder Said March 2020: Dating the First Renders of a Multi-Year AI Programme

An earlier version of this piece told a good story. Sort the project archive by date, it said, and the oldest thing in it is a folder of eight raw FMI strip renders dated March 2020, 21 months before the December 2021 kickoff. From that it argued that the programme was won by technical courtship: unpaid work on the operator's own data, through Covid, long before a contract existed.

The story does not survive the files. "March 2020" is the name of the folder, not the date of anything in it. We have corrected the piece rather than quietly withdrawing it, because the way the mistake happened is more useful than the story it replaced.

What the files actually say

The folder holds two kinds of file. The first is the operator's data for one well: a processed FMI log, a conventional log, and two interpretation tables. The log's own header records the logging date as 25 September 2021, and the processed file carries a creation time of 26 September 2021. That data did not exist in March 2020. Its copies in the archive are dated 8 March 2022.

The second is the eight renders, FMI1 to FMI8. Each is dated 9 March 2022, the day after the data landed, and each was drawn with a release of the plotting library that was not published until June 2020. Nothing else in the archive points earlier. The first client well data anywhere in it is dated January 2022, a month after kickoff. The only project documents from 2020 are the non-disclosure and data-processing agreements, dated 16 July 2020, and the proposals that followed them.

Read by the files, the first renders of the operator's borehole were made in the first quarter of the paid programme, three months after kickoff. They are still a picture of the client's own rock, drawn the day after the data arrived. They are not evidence of anything that happened before the contract.

DATING THE FIRST RENDERS · THE FOLDER NAME VS THE FILES24 monthsbetween the two readingsA folder named March 2020 holds renders dated March 2022, beside a log recordedin September 2021. Read by the files, the first render came after kickoff, not before.EARLIEST RENDERS IN THE ARCHIVE8,4008,9009,4009,9008-pad FMI, no depth unit shownrendered from the operator's DLIS24-MONTH MISDATING2020202120222023Mar 2020folder nameJul 2020DPA signedSep 2021log recordedDec 2021kickoffMar 2022render file dates2-yr term60-day noticerendersREAD THE RENDER DATEby the folder nameone label, no file evidenceby the filesfile dates and the log headerrenders datedMar 2022against the DPA20 months afteragainst kickoff3 months afterevidencefile dates, log header
The dated spine of the engagement, read two ways. The earliest FMI strip renders in the archive sit in a folder named March 2020, and an earlier version of this exhibit took that name as their date, which put them 21 months before the December 2021 kickoff and four months before the July 2020 data-processing agreement. The files do not support it: the eight renders are dated 9 March 2022, the data files beside them 8 March 2022, and the log header of that well records 25 September 2021. Toggle between reading the date by the folder name and by the files; the render marker moves and the read-out restates its position against the agreement and the kickoff. The orange band is the 24-month span between the two readings, the one element that argues. The signed service agreement carried a 2-year term with 60-day either-party notice. All anchor dates are sourced from the project archive and the agreements, rounded to the month; the render's depth axis runs roughly 8,400 to 9,900 with no unit shown.

The instrument above sets the two readings side by side. Read by the folder name, the renders sit 21 months before kickoff and four months before the data-processing agreement. Read by the files, they sit three months after kickoff and 20 months after the agreement. The orange band is the gap between those readings: 24 months of story resting on one folder label.

How a folder name became a date

Nothing here was invented. The folder really is named March 2020, and an archive sorted by folder name puts it first. The error was treating a label someone typed as a timestamp something recorded. A folder name is metadata a person chooses, can mistype, and never has to update. A log header, a file's own creation time, and the version of the software that wrote it are recorded by machines with no stake in the story. When the two disagree, the recorded evidence wins.

The depth axis was a smaller version of the same mistake. The earlier piece said the render's axis, roughly 8,400 to 9,900, was in feet, and that this dated it to before the programme settled on metric plots. The axis carries no unit. Feet was an inference, and it was doing work the evidence never did.

What the contract looked like when it finally arrived

The contract terms the earlier piece cited are right, and they read the same way without the story around them. The service agreement, in the draft held in the archive dated 15 November 2021, carried a 2-year term with 60-day either-party termination notice. It was written as an effort obligation rather than a result obligation: the provider is not responsible for achieving a specific result. That framing suits a research programme on data neither side has fully seen.

The transferable lesson

Date evidence by what recorded it, not by what someone named it. Before a date carries an argument, check it against something a machine wrote at the time: a log header, a file's creation time, the version of the tool that produced it. If those are missing, say the date is unknown rather than borrowing one from a folder. The corrected record here is less flattering than the story it replaces, and it is the one we can defend.

Limitations

This correction rests on file dates and file headers in one archive. A file date records when a file was last written, so a render could in principle be older than its file date, but not older than the data it draws, and no client well data in the archive is dated before January 2022. We have not reconstructed when commercial conversations began or what was shown in them; the archive's proposals from August 2020 onward describe the approach in general terms. The 2-year term, 60-day notice and effort obligation are from the draft service agreement in the archive.

References

  1. Engagement archive, folder named "March 2020": one well's processed FMI and conventional logs (log header date 25 September 2021, file creation time 26 September 2021, archive copies dated 8 March 2022) and eight FMI strip renders dated 9 March 2022, drawn with a plotting-library release first published in June 2020. Internal materials, withheld under operator confidentiality.

  2. Non-disclosure and data-processing agreements dated 16 July 2020; draft service agreement dated 15 November 2021, with a 2-year term, 60-day either-party notice and an effort obligation. Internal materials, withheld under operator confidentiality.

Tarry SinghTannistha Maitiby Tarry Singh and Tannistha Maiti
Discuss a programme
Stay ahead

EarthScan insights, in your inbox.

Field-tested research on subsurface and energy-transition AI. About twice a month. No noise.

We use your email only for this newsletter. Unsubscribe anytime Privacy.