Can Jev pick emission factors as well as an LLM?

Key finding. On 142 procurement lines, Jev put the right emission factor first on 58.5% of lines, against 64.8% for the LLM it replaced. The gap is not statistically significant, and it sits entirely at the last step: Jev found the right material family as often as the LLM, then chose the wrong variant inside it. The whole matching run cost 57% less.

What is Jev? TypeSafe launched Jev on 15 September 2026 as a "System One" model. It is not a language model: it writes no text. You give it a state and a set of typed questions, and it returns a probability for every answer, in one pass. TypeSafe positions it for classification and routing, and charges for input only (USD 0.042 per million input tokens at launch; output is free, since there is none).

Why we tested it. Matching a procurement line to an emission factor ends in a choice: our pipeline finds ten candidate factors, and a model picks one. That is a classification task, the kind Jev is built for. If a cheaper model makes the same choice, every inventory we match gets cheaper.

The test: one step swapped, everything else fixed#

1. We took 142 procurement lines from real customer files, each with a reviewed reference factor in ecoinvent. Most are materials and components: 81 raw materials (steel, plastics, chemicals, concrete), 25 assembled products and 14 electronic parts. The lines are as messy as the files they came from: part numbers, abbreviations, and descriptions in Finnish, German, Spanish and other languages.
2. Our pipeline matched each line twice. It reads the line, writes a search, retrieves candidates, and passes the ten best to a model that picks one. A geography check and an audit step follow.
   • The LLM arm: GPT-5.6 luna at medium reasoning effort, the model and the prompt our pipeline used in production at the time, tuned over several months.
   • The Jev arm: jev-latest on 21 September 2026, with instructions we wrote for it that day.
3. Only the picker changed. Both arms ran fresh, with no cached results. The retrieval step returned the same candidates in both: it put the reference factor first on exactly the same 61 lines, and among the ten on exactly the same 113.
4. We compared the two arms line by line. For each line we recorded where the picker ranked the reference factor. "Top 1" means it chose the reference factor; "top 2" means the reference factor was its first or second choice.

On 29 of the 142 lines, the reference factor was not among the ten candidates. No picker can get those right, so both arms are capped at 79.6%.

The results: the same family, a different variant#

Counted only on the 113 lines where the reference factor was among the ten, the LLM picked it 81% of the time and Jev 73%.

Line by line, both picked the reference factor on 77 lines and both missed on 44. The LLM was right where Jev was wrong on 15 lines, and Jev right where the LLM was wrong on 6. A 15 to 6 split on 21 lines is not significant at the usual 5% threshold (exact McNemar test, p = 0.078). After the geography check and the audit, which repair some picks, the final result was 63.4% for the LLM and 59.2% for Jev (10 lines to 4, p = 0.18).

Where Jev went wrong. On 11 of its 15 misses, the reference factor was Jev's second choice, and on 3 more its third. Jev read the line correctly and found the right family of factors. It then chose a neighbouring variant:

The variants carry different factors, so in an inventory this choice moves the number. Some of these are close calls: a one-word "steel" line does not say which steel it is, and a reviewer could argue for either choice. Jev itself was unsure on most of them: on 8 of its 15 misses, it graded its own match "acceptable" rather than "high".

The 6 lines Jev won look like the 15 it lost, in reverse. On printed advertising material, Jev chose printed paper and the LLM plain coated paper; on a packaging line whose reference factor is PET labels, the LLM also chose coated paper.

Confidence: no reasons, and grades that bunch in the middle#

Our pipeline grades every match, so that a person reviews the doubtful ones and not the rest. The LLM writes a grade and a reason in words. Jev returns a probability for each grade and writes nothing, so its arm had no reasons to show.

Jev moved lines out of the two clear grades, "high" and "review", into the two in the middle. The share of lines sent to review stayed about the same (47.9% with the LLM, 46.5% with Jev), but the flags were a little less useful. Of the picks that were materially wrong, the LLM flagged 75% and Jev 71%. Of the lines flagged, 65% and 61% did not carry the reference factor.

The missing reason matters more than the grades. A reason in words is what lets an auditor check a match without redoing it: "the line says gunmetal, which is a bronze, so market for bronze". A probability does not explain itself.

Cost: 57% less for the whole run#

These are the costs of the whole matching run, every step included, at list prices on the day. Only the picker changed, so the whole saving comes from that one step. The 45,000 output tokens in the Jev arm come from the other steps, which still use an LLM. The run was not slower: 3 minutes 51 seconds for the Jev arm against 4 minutes 54 for the LLM, with 50 lines in parallel.

We did not measure energy use, so we make no claim about it.

What it means#

• For the choice between close variants, the LLM is better here. Low-alloyed or unalloyed steel, nylon 6 or 6-6: that last step is exactly where a factor choice moves an inventory, and it is where Jev lost.
• Jev is close on everything before that step. From the top 2 onward it matched or beat the LLM. On choices where the options are far apart, it may well be good enough, for less than half the cost. We have not tested that.
• The test favours the LLM. Its prompt was tuned for months on lines like these; Jev's instructions were written in a day. Instructions written for the variant choice are the obvious next test. If we run it, it will be a new dated study, and this one stays as it is.

We kept the LLM as the picker.

The limits of this test#

• A small sample. 142 lines, one run per arm. The gap is not significant, and on this suite a second run of the same arm can move the final result by several points.
• One library, mostly materials. All reference factors are in ecoinvent, and 81 of the 142 lines are raw materials. Services, energy, transport and spend-based lines are barely represented.
• Our own reference factors. Each reference factor was reviewed, but some lines are ambiguous, and on a few of Jev's misses a reviewer could defend its choice.
• Unequal preparation. A prompt tuned over months against instructions written in a day.
• One version, one date. jev-latest and GPT-5.6 luna as served on 21 September 2026, at the prices of that day.

Try the matching on your own lines#

Our matching agent runs the pipeline tested here on your raw procurement export:

• No cleaning, no template. Send the file as your ERP or accounting tool exports it, in its own columns and languages.
• Every match graded and explained. Each line comes back with its factor, library, year and licence, a confidence grade and the reason for the choice.

Send us your procurement export and get every line back with a cited emission factor. No account needed.

Related#

• How wrong are spend-based GHG estimates? We measured it