Craftpine

Craftpine

A research and development studio in Toronto. Every program is published in full, including the ones we close.

Since April 2026

The studio has run two programs. Corner ran 76 days, from April 30 to July 14, and closed. Its study of a store at 520 Barton, in Hamilton, was published the next day. Terrasope began in August and is paused. A preprint on folder embeddings followed on October 6.

AprMayJunJulAugSepOctApr 30First commitJun 10Destruction pass on themoney loop, all heldJun 30Unattended loopshipsJul 9First livecharge, $1.99Aug 12First spike,31 of 31 checksSep 10Bench workstopsCorner · 76 daysclosed Jul 14TerrasopepausedJul 15520 Barton studypublishedOct 6PreprintpublishedAprMayJunJulAugSepOctpausedApr 30Corner's first commitJun 10Destruction pass on the moneyloop, all heldJun 30Corner's unattended loop shipsJul 9First live charge, $1.99Jul 14Corner closed, day 76Jul 15520 Barton study publishedAug 12Terrasope's first spike,31 of 31 checksSep 10Terrasope bench work stopsOct 6Preprint published
Each program is a bar, drawn to scale by day. Terrasope's bar starts at its earliest dated event, the August 12 spike. Dates from the close-out ledger, the program reports and the preprint's Zenodo record.

Corner

Corner was the studio's first program. It built software for Toronto corner stores and planned an unattended store of its own at 520 Barton Street in Hamilton. On July 14 the store plan closed and the software was parked.

The Corner bench unit, a small touchscreen on a wooden bench, showing a tap-to-order drink menu

Program · Closed July 14, 2026

Corner

Payments and inventory for Toronto corner stores. It took a live charge, then closed on the numbers.

How it worked and why it closed
Aerial view of the corner of Barton Street and Dewitt Road: the two-storey building, its lot, and cars queued at the signal

Corner's store study · July 15, 2026

520 Barton

Could an unattended store with no staff pay for itself on one Hamilton corner? Eight concepts entered the study and none survived.

The study on one page
Read the 520 Barton study in full
On the web · 9 chapters, 4 appendicesDownload the PDF
78 pages · 11 MB

Terrasope

The second program began in August 2026 and is paused at the bench stage.

BoreSampleReadDecideRecordbuilton paperbuilt

Program · Paused September 2026

Terrasope

Critical-mineral sensing for ground with no GPS, no satellite link and no people. Only the record layer is built so far.

Where it stands

Preprint

A folder's metadata is read before its files and holds a few kilobytes. A vector kept there can route a query without opening the folder, and the cheapest one to keep is the mean of the files' embeddings.

When is that mean enough? Where a folder mixes images and texts, the mean loses its minority modality, the kind of file there is less of.

textsimagesmeanheld-out imagea few vectorsper typeRecall@5 lost by the meanA held-out minority file, Zenodo folders00.10.20.30.40.5Two vision-language embedders0.16 to 0.25Two CLIP-family dual towers0.35 to 0.50textsimagesmeanheld-out imagea few vectorsper typeRecall@5 lost by the meanA held-out minority file, Zenodo folders00.10.20.30.40.5Two vision-language embedders0.16 to 0.25Two CLIP-family dual towers0.35 to 0.50
A mixed folder's mean sits between its texts and images, far from the minority. A held-out minority file therefore finds its folder less often through the mean than through a few vectors per type. The drawing is a schematic, not data. Ranges from the preprint's abstract, in recall@5: how often a file's own folder ranks in the top five.
A few vectors for a mixed folder
Version 1 · October 6, 2026