The model signs nothing

Every lab report ends with a signature: somebody at the lab puts a name under the results. A brand that imports a dresser certifies it meets the federal rules, and the report is its evidence. At the brand in this story, no report reaches the records until a person approves it. The model signs nothing.

This month we released our first open-weight model. We call it Vulcan 2.5. It reads the paper behind product compliance and puts what it finds on the record. Then it stops, and a person decides.

Here’s one lab report, followed from the email to the approval. The report is invented. So are the brand, the factory and the dresser. But the steps are the ones our product compliance pass runs for a home furnishings retailer.

The pass matches every SKU on the open purchase orders to its lab report, fills in the CPSC’s own workbook and logs weak matches. This brand chose to have each report read the night it lands, not in a monthly pass. The model doing the reading is ours.

11:48 p.m. The email.

At the factory outside Ho Chi Minh City, it’s 1:48 on Tuesday afternoon. At the brand it’s Monday night, and the compliance inbox is the only thing awake. One email comes in with a thirty-one-page PDF from the factory’s testing lab.

The subject line is an item number, and the body is two words: see attached.

11:49 p.m. The reading.

Vulcan 2.5 opens the report. The cover page gives the lab, the report number and three dates: sample received 04/09/2026, tested 04/09/2026 to 11/09/2026, issued 12/09/2026.

The lab writes the day first. So the sample arrived on September 4, the tests ran for a week, and the report came out on September 12. But read the American way, the lab received the dresser on April 9, tested it for seven months and issued the report on December 9. It’s September 21. A person would laugh, and a parser would file it.

Vulcan 2.5 writes September 12 and cites page 1, where the date came from. Every value it takes carries its page. Page 2 describes the sample: supplier item DR6, a six-drawer dresser in two finishes, Walnut and White.

On page 6, the dresser passes the stability test under 16 CFR part 1261, the federal rule meant to keep a dresser from tipping onto a child.

Page 9 is the lead test under 16 CFR part 1303. That rule holds the paint on a dresser to the same limit as the paint on a toy: ninety parts per million. It doesn’t ask whether anyone plans to chew the dresser. Walnut and White both pass.

11:51 p.m. The match.

The brand’s open purchase orders list the dresser three times: Walnut, White and Sage, a new finish on its first order. The report covers two of them.

At the end of a long day, you might match Sage anyway. It’s the same frame, the same drawers and the same tip test, so it’s probably covered. Probably isn’t a field.

The lead test is a test of the paint, and Sage is a new paint the report never names. So Vulcan 2.5 matches Walnut and White and marks each one Update, a new report for an item the brand already carries. Sage goes in the log with its reason, because the pass never files a weak match.

Then the model stops. It doesn’t write to the factory, and it leaves even the two matches it’s sure of as drafts for the morning.

8:10 a.m. The desk.

The compliance lead opens her day to two SKUs ready for review and one in the log. Say her name is Dana. She clicks the issue date, and the report opens at page 1 with the date marked where the model found it. She checks the two results the same way, on page 6 and page 9.

The model left her a reading she can hold against the paper, line by line. It didn’t leave her a decision.

8:26 a.m. Approved.

Dana approves Walnut and White, and their rows go into the CPSC’s workbook. For Sage, she writes one line to the factory: please send a lead test for the Sage finish. This brand doesn’t sell a finish its reports don’t name. So Sage waits.

The approval files a receipt beside the report.

• Source: the PDF, all thirty-one pages, stored exactly as it arrived.
• Read: the lab, the report number, three dates, two finishes, two rules, two results, each with its page.
• Matched: Walnut and White, Update. Sage, not in this report, logged.
• Approved by: Dana, 8:26 a.m.
• Model: Vulcan 2.5, and the exact file it ran.

Why the weights are open.

The model signs nothing. But its name is on the receipt, one line under hers. A lab report is worth something because you can check it. It names the rule and the method, so another lab can run the same test on the same dresser. A model should work the same way.

A model is a large file of numbers learned from examples, and those numbers, called weights, decide how a page becomes a field. If you rent a model, its owner can change it or retire it. The receipt would still carry the name of something nobody can run anymore.

We’ve published Vulcan 2.5’s weights. The version on Dana’s receipt is a file, so next year the brand can run it on the same report and check what it read.

Why it does one job.

Ask a general model on Monday which way 12/09/2026 goes, and ask it again on Tuesday. You can get two answers, both confident.

Vulcan 2.5 reads lab reports, spec sheets and vendor packets. That’s the whole list, and nobody asks it for a sonnet. It started from an openly licensed base model and was trained on proprietary data for this one job. No client’s report, supplier, SKU or price is in the weights.

Give it the same report twice and you’ll get the same reading twice. And it gets more fields right than a general model does. We’ll publish the test.

Dana will never open the weights. She reads the receipt, and she needs its last line to mean the same thing next year that it means this morning.

The model still signs nothing. Now anyone can check the one that read the paper.

Bring us the documents your team retypes, and we’ll show you what the model reads and where a person signs. Get in touch. Vulcan 2.5 is the first of our models, and you’ll find the others, one per industry, under Models.