JEV made fast, structured model decisions a first-class software primitive. Geltre is built for the step immediately before that decision: selecting the evidence that actually matters.
The first Orienting Model. Geltre selects the task-relevant evidence from large, noisy state before JEV, another decision model, or an LLM sees it—without generating or inventing evidence.
Decision quality depends on the state a decision model sees. Production applications rarely have one neat paragraph of context. They have thousands of possible facts, most of which are irrelevant to the decision in front of the system.
Geltre treats that upstream step as its own model primitive. In the paired API, the developer supplies state plus the typed JEV question once. Geltre uses that same question to select exact source references, then JEV receives the reduced state with the question unchanged. It does not replace JEV or require a duplicate task.
Geltre orients. JEV decides. LLMs deliberate when needed. Code acts.
Held-out ServiceOps v0.3 result only. This is not a universal cross-domain claim.
The earlier v0.8 CodeOps cohort failed its frozen recall gate at 0.808219 and remains published as a negative result. The fresh v0.13 H256-R4 cohort passed its evidence/reliability gates; across 197 paired live JEV records it caused 0 damage, rescued 1 Raw → JEV error, and reduced JEV input/cost by 64.5582%. These results support the explicit CodeOps alpha contract, not a universal selector claim.
Read the benchmark evidence →Orient logs, status, policy, and recent actions before JEV chooses ACT, RETRY, RESTART, ESCALATE, or ASK.
Select mutation and verification evidence from long agent traces before a VERIFY, REVISE, or ACCEPT decision.
Reduce noisy tool history at deterministic checkpoints before a bounded control decision. Each new domain still needs its own validation.
Pass a compact provenance-preserving evidence set to a slower generative model only when open-ended deliberation is actually needed.
Select and bind task-relevant evidence from noisy state.
Return a typed probabilistic decision software can use directly.
Escalate only when generation or deeper deliberation is actually useful.
Execution, permissions, retries, state changes, and receipts stay deterministic.
JEV is a product of TypeSafe AI. Geltre is a product of Nichols SI. Nichols SI is not affiliated with, endorsed by, or partnered with TypeSafe AI. References to JEV describe an independent integration pattern and public benchmark work.