karpo-decision-31B is second on Decision Index 0.4.1 at 73.86; REAX 1.3 stays first
Source · Model card: karpo-decision-31B
The Decision Index space file data/index.json was generated on 11 Nov 2026 at 14:28 UTC. It has 297 entries. The file of 6 Nov had 274. Full score is still balanced_skill. Jev 1.13 in this file is still 59.22, which remains the catalog reference. This board is not mixed with Artificial Analysis or JevBench.
Strict sort of that file:
| Rank | Model | Full |
|---|---|---|
| 1 | REAX 1.3 | 73.98 |
| 2 | karpo-decision-31B | 73.86 |
| 3 | GLiDE 27B | 73.37 |
REAX 1.3’s Full score is the same number as the 6 Nov news. It is still not a catalog page. GLiDE’s 73.37 is the same number as the 6 Nov row; the rank is now third. The GLiDE capability score stays 8.6, because accuracy 9 comes from two boards above Jev, not from being second.
karpo-decision-31B (MachinePulse, weights first published 5 Nov 2026) is a Gemma 4 31B fine-tune with a readout over option codes. The card documents choice, noul and score, text only, and a /v1/systemone server in machinepulse-ai/karpo-decision. From 12 Nov 2026 the card’s licence for the weights is MachinePulse Community: commercial use under US$10 million revenue, with an attribution line for products offered to third parties. Copies taken earlier under Apache-2.0 keep that grant. On 12 Nov the repo had 26 downloads and 3 likes.
The card calls the model #1, tied with GLiDE and REAX, because the board UI groups scores within 0.9 points and the chart they show hides inference-technique rows. That grouping is the author’s claim (UnverifiedClaim). The file’s three Full scores are 73.98, 73.86 and 73.37.
Catalog page: karpo-decision-31B. On 12 Nov the scores were 4.8 / 7.4 / 2, from this 11 Nov file, where karpo was strict second and accuracy stayed at the cap of 8.
Update, 16 Nov 2026. The file generated 2026-11-12T12:34:44Z (300 entries) still has karpo at Full 73.86 and puts it strict first. REAX 1.3 in that file is 72.64. Untied first place lifts accuracy from the cap of 8 to 9. Capability is 7.6. Maturity 4.8 and adoption 2 stay. Calibration is still 4.
Other new rows in the same file, not catalogued this round: Logits Sigma-1 Duo 72.28, Copernicus-31B 72.11, OneJev1.6g-31B 70.86, Jeffan-27B-1211 70.30, eshbar-lulilla 70.30. EmbeddingGemma 2 System-One is still Full 7.41.
