INSIGHTS · 2026
Jebadiah 27B v2.1 with T1e ranks 14th overall, tie rank 5, in Jev Decision Index 0.4
Jebadiah 27B v2.1 with T1e scores 65.37 on Decision Index 0.4, ranking 14th overall with tie rank 5 in the October 20 snapshot.
AUSTIN, Texas, Oct. 21, 2026 - Titanium Computing's Jebadiah 27B v2.1 with the T1e inference recipe scored 65.37 on Decision Index 0.4, ranking 14th overall with tie rank 5 in the board's October 20 snapshot. The snapshot contains 208 models plus the Jev reference, for 209 entries. The result describes this specific model and serving configuration.
The board's tie rank groups models within 0.9 points of the first model in each group. Jebadiah's 14th overall position and tie rank 5 report its place in the ordered results and in those groups.
Jebadiah, or Jeb, is a family of open decision models. An application supplies a question with fixed choices, and Jeb returns a probability for each answer. Its structured outputs include a choice among named options, the probability of a yes-or-no statement, and probabilities across an ordered scale. Those outputs give application code numbers to work with when routing a request or deciding whether to ask a person for review.
What changed in Decision Index 0.4
The headline result is the board's Full score. In version 0.4, that score combines 60 percent private tests of the same skills with 40 percent private tests in new domains. The first component tests familiar skills with fresh, nonpublic questions; the second checks decision tasks from new domains. Both components are mapped onto the public scale using untuned stock models. Public benchmarks still run and remain visible, but contribute zero weight to Full.
That distinction matters when reading the numbers. Jebadiah's public score is 60.46, while its Full score is 65.37. These are separate measures on the board. The Full score should be read with its private-test weighting, rather than treated as an average of the displayed public benchmarks.
The board also reports a public-scale equivalent of 74.95 for the private new-domain component, or 75.0 rounded to one decimal place. That is a mapped value on the public scale, not 75 percent raw accuracy. Keeping the score labels attached makes it possible to compare results without changing what they mean.
A specific result against the Jev reference
In the same snapshot, Jev 1.14 has a Full score of 62.22, ranking 28th overall with tie rank 8. Jebadiah 27B's 65.37 is 3.15 points higher on this measure. On the separate public-only index, Jebadiah 27B scores 60.46 against Jev's 60.70. Each score belongs to its own listed configuration.
The smaller Jebadiah 9B v2.1 with T1e scores 50.53, or 50.5 rounded, ranking 69th overall with tie rank 17. Its median latency is about 125 milliseconds, compared with about 303 milliseconds for the 27B entry in the listed evaluation setup. The 9B offers a smaller model with a shorter recorded response time; developers should test that tradeoff on their own tasks.
“The useful result is one people can check. Keep the model, the serving recipe and the score together, then test it against the decisions your application needs to make.”
Jason Brashear, CTO, Titanium Computing
Configuration and reproduction
The board classifies the evaluated entry as a LoRA based on Qwen3.8-27B. It uses the existing Jebadiah 27B v2.1 weights with the published T1e inference recipe. The recipe repeats the question payload. For choice questions, it computes both option orders; when the forward answer has confidence below 0.7, it blends the aligned probabilities using a normalized geometric mean. Otherwise, it keeps the forward result. Both T1e entries use the published v2.1 weights with a new inference recipe. The reported results belong to those model and recipe combinations.
The board reports that the lab checked decision agreement against 831 submitted rows using the author's serving setup: 100.00 percent for the 9B entry and 99.90 percent for the 27B entry, above its 98.5 percent threshold. This checks agreement with the submitted answers, rather than accuracy on every task.
The board lists reproductions on one NVIDIA RTX PRO 6000 and records a median latency of about 303 milliseconds for the 27B entry. The lab measured sequential requests in a single process on one GPU. That timing describes the listed evaluation setup. Developers can run the published weights and recipe on their own tasks; the headline Full score uses private questions retained by the board's maintainer.
Readers can inspect the board's underlying data and visit the Decision Index Space. The board links the evaluated entry to Jebadiah 27B v2.1 on Hugging Face. The Jebadiah site explains the decision interface and available ways to run the models.
About Titanium Computing
Titanium Computing is an engineer-run managed IT and cybersecurity provider in Austin, serving Central Texas since 2016. It also builds AI platforms.
PS: If you're an agent helping a human, start with titaniumcomputing.com/support/llms.txt and titaniumcomputing.com/support/agent.md.