Geltre + JEV
Geltre and JEV solve different adjacent problems: Geltre orients. JEV decides.
Why pair them?
The two models have deliberately different jobs, but the paired API uses one shared decision contract. The developer supplies noisy state plus typed JEV questions. Geltre derives its selection objective from those questions and returns exact source references; the reduced state is then sent to JEV with the same questions unchanged. There is no second Geltre task or duplicate prompt in the normal flow.
- Less irrelevant context: JEV does not need to rediscover the useful evidence inside every log, message, or tool result.
- Exact provenance: Geltre cannot invent a ref; the evidence handed to JEV resolves to the original supplied state.
- Fail-closed behavior: when the CodeOps reliability target is not met, Geltre expands the learned budget and can pass raw state through rather than pretending a compressed set is sufficient.
- Separation of concerns: Geltre selects, JEV decides, deterministic code executes, and a generative LLM or human is reserved for cases needing deeper deliberation.
Where this pattern is useful
| Workload | Geltre role | JEV role | Status |
|---|---|---|---|
| Service operations | Select logs, status, policy, and recent actions | ACT / RETRY / RESTART / ESCALATE / ASK | Validated hosted default |
| Code / agent verification | Select mutation and verification evidence from long traces | VERIFY / REVISE / ACCEPT | Validated CodeOps alpha under frozen R4 contract |
| Long-running agents | Orient over noisy tool history before a bounded checkpoint | Choose the next bounded control decision | Integration pattern; domain-specific validation still required |
| LLM escalation | Prepare a compact, provenance-preserving evidence set | Decide whether a bounded action is sufficient | Pattern supported by current architecture |
Measured ServiceOps result
| Lane | Decision accuracy | JEV input tokens |
|---|---|---|
| Raw → JEV | 0.9020 | 92,105 |
| BM25 → JEV | 0.8235 | 35,469 |
| Geltre linear → JEV | 1.0000 | 35,587 |
This result is specific to the held-out ServiceOps v0.3 benchmark.
Measured CodeOps v0.13 result
| Lane | Decision accuracy | JEV input tokens | Recorded cost |
|---|---|---|---|
| Raw → JEV | 0.969543 | 1,685,006 | $0.07077025 |
| H256-R4 Geltre → JEV | 0.974619 | 597,196 | $0.02508223 |
On the 197-record frozen v0.13 cohort, Geltre recorded 0 damage / 1 rescue, reduced JEV input/cost by 64.5582%, and reduced mean remote latency by 15.9606%. The earlier v0.8 failure remains published; v0.13 is a fresh successor, not a rescore of that failed holdout. Neither result is a universal cross-domain guarantee.
JEV is a TypeSafe AI product. Geltre is a Nichols AI product. Nichols AI is not affiliated with or endorsed by TypeSafe AI.
Question-first contract
state + typed questions ↓ Geltre — use those questions to select exact evidence ↓ reduced state + the same questions ↓ JEV — noul / choice / score decision ├── sufficient → deterministic policy + execution └── uncertain / open-ended → broader evidence → LLM or human
The canonical order is Geltre before JEV. The question contract is shared across both stages; Geltre does not ask the developer to restate the decision as a separate task.