Research brief / AI systems
Laya and Jev Decision Probe
A small typed-decision comparison and a test of prompt routing for arithmetic, with the limits kept visible.
The research question
Can a local decision model improve on a specific arithmetic miss when a hook supplies an explicit operation order?
Inside the work
Three synthetic choice questions compared a pinned local Laya checkpoint with Jev 1.13.0. A separate arithmetic prompt revision and a per-call Laya hook tested the same left-to-right guidance. The hook changed the model-facing instructions while keeping the caller's original question intact. These are prompt-routing observations, not training results.
Research into compiling human intent into a structure an agent can actually execute. The ALX project offers a concrete way into this field of research.
Explore ALXObserved results
September 26, 2026. Three fixed synthetic choice cases ran once through Jev 1.13.0 and a pinned, offline Laya checkpoint. The two arithmetic follow-ups were targeted interventions, not new baseline cases. Percentages below are the models' probability for their selected choice.
| Condition | Jev | Laya |
|---|---|---|
| App crash routing | technical · 100% | technical · 92.71% |
| 12 − 5 + 3, broad question | ten · 86% | four · 38.24% |
| Age absent from facts | unknown · 100% | unknown · 65.31% |
| Arithmetic with explicit PEMDAS guidance | ten · 100% | ten · 52.77% |
| Broad question through Laya math hook | not called | ten · 52.77% |
Jev matched the expected choice in all three original cases; Laya matched two. Laya changed from four to ten when the instructions said addition and subtraction have equal precedence and should be evaluated left to right. The hook reproduced that model-facing wording, so its ten answer confirms routing to the tested prompt path. It does not show unassisted arithmetic improvement.
What to test next
The model-gym hypothesis is to freeze natural questions and expected answers, vary one instruction or routing intervention at a time, add different arithmetic forms and non-math controls, repeat runs, and evaluate held-out cases before considering training. Exact calculations in an application should also have a deterministic check.
What this brief establishes
This is one small synthetic probe set, not a model benchmark or a fine-tuning result. Selected-choice probabilities are not measured accuracy, and the models' separate confidence fields are not directly interchangeable. The local hook is not installed in a product. No real financial records were used.
In the same field
ALX Semantic Compiler
Source-preserving semantic compilation, execution contracts, and bounded repair with explicit evaluation limits.
Protocol Compiler Framework
A repeatable method for compiling research and operating rules into agent-usable protocols.
ALX LoRA Evaluation Track
Evaluation plans for whether specialized training can preserve ALX structure and reversibility.