A third-party experiment by Senna reports that Qwen2.5-0.5B can be arranged to reproduce some structural behaviors associated with Jev: handling multiple questions against shared state and returning typed decisions. The experiment does not reproduce Jev’s general decision-making ability. On the author’s reported examples, the structure held together, but the answers did not transfer.
What the experiment tried to reproduce
Jev is described in Senna’s article as a system that receives shared state and one or more questions, then returns typed decisions rather than free-form answer strings. The article names three answer forms:
As an Amazon Associate I earn from qualifying purchases.
- Noul: a yes-or-no decision represented as a probability.
- Choice: a selection among supplied options, represented with a probability distribution.
- Score: a value on a supplied scale, with a score, distribution, and confidence.
These are the article’s descriptions of Jev, not independently confirmed API specifications. Its author also attributes to TypeSafe the statement that “Jev outputs all probabilities in parallel instead of autoregressively generating by token.” The TypeSafe source itself was not independently checked, so that wording should be understood as a quotation reported by Senna, not a verified primary-source statement.
Free tools Windows power users keep installed
One-click scans. No signup required.
How Senna proposed a Jev-like structure
TypeSafe had not published Jev’s full architecture, according to the article. Senna therefore proposed a plausible reproduction based on public clues and hypotheses attributed to Archer Hume. The implementation used Qwen2.5-0.5B as a conventional causal-decoder backbone; it was a hypothesis about how to reproduce visible structure, not a claim about Jev’s actual internals.
Shared state and isolated question branches
The proposed model packs shared state and multiple question branches into one request and runs the transformer once. A tree attention mask lets each question attend to the shared state and its own branch while blocking access to sibling questions. The author also resets position IDs at the start of each question branch, while leaving Qwen’s feed-forward blocks intact.
Different heads for different answer types
For Choice and Score, Senna used a pointer-style head. Noul used a separate linear layer followed by a sigmoid. The separation matters: the experiment aimed to make the model’s outputs follow Jev-like typed structures, not to give the small Qwen backbone Jev’s underlying judgment ability.
What changed when questions or options changed
In Senna’s implementation, adding or inserting questions reportedly changed an existing question’s probabilities by at most about 0.0006. That result is specific to the author’s code and setup; it is not an independently replicated measurement or a Jev performance result.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Option changes behaved differently. The author reports substantial probability movement when options were reordered, and says adding an irrelevant option changed the relative odds between existing options. Together, these observations describe the behavior of this reproduction: question branches were largely insensitive to changes elsewhere in the question set, while a Choice question’s distribution was sensitive to the options it contained.
Training results: accuracy did not improve with more examples
Senna froze the Qwen backbone and trained only the Choice/Score pointer head on AG News. The author reports the following evaluation figures:
| Training examples | Reported evaluation accuracy | Reported calibration trend |
|---|---|---|
| 10,000 | 0.8300 | Better than the 20,000-example run, according to the author |
| 20,000 | 0.7720 | Worse than the 10,000-example run, according to the author |
These are Senna’s reported experimental results, not external benchmarks. The author attributes the drop at 20,000 examples to the training setup: one epoch, batch size one, and a fixed learning rate. The comparison therefore does not establish that adding training data generally reduces accuracy, nor does it demonstrate a limitation of Jev.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Did the trained model match Jev’s example answers?
Senna then tested the trained head against Choice and Score examples from TypeSafe documentation. The article reports that 2 of 8 Choice answers matched and 2 of 9 Score top-level answers matched. For Score, the head saturated at its highest level.
The comparison used documentation examples, not live Jev API results. It also excluded Noul because that head had not been trained. The results support a limited conclusion: the implementation produced the intended answer structures, but its answers did not reliably transfer to these examples.
What this reconstruction shows—and what it does not
The experiment is useful as an architectural illustration, not as evidence that a small Qwen model can replace Jev. Its attention mask, branch-position handling, and output heads are Senna’s proposed design choices. The reported probability stability and training figures describe that implementation alone.
Quick Recap
- It shows: a causal-decoder model can be organized around shared context, isolated question branches, and typed output heads in a Jev-like arrangement.
- It does not show: that Jev uses this architecture, that the reported behavior matches Jev across tasks, or that the reproduction inherits Jev’s decision quality.
- Its transfer check found: only a small number of matches against the cited documentation examples, with Score outputs tending toward the maximum level.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
