Build blueprint
LLM Output Escalation Mesh
A mesh of cheap per-claim decisions selecting accept, recheck, regenerate, retrieve, or ask a human.
- Problem
- Systems apply one verifier to all generated outputs or trust them uniformly.
- Why Jev
- A generated response can require dozens of independent confidence and policy decisions.
- Architecture
- Parsed output units → parallel verifier/router questions → selective expensive checks → response.
- Current alternative
- One LLM-as-judge pass or universal retrieval.
- Jev advantage
- Spends expensive verification only where cheap decisions indicate risk.
- Unknowns
- Verifier correlation with the generator and claim segmentation quality.
Why this confidence: Promising cascade architecture; independent verifier performance is unknown.
This is an authored hypothesis derived from the research corpus. Nothing here demonstrates product demand, or that Jev performs well on this particular workload. Run the validation experiment before building past the MVP.
Sources
Limitations
- Verifier correlation with the generator and claim segmentation quality.
- This is a research hypothesis, not evidence of product demand or Jev performance in this workflow.
Supporting research
- Probabilistic predicate + deterministic action PlausibleJev supplies fuzzy predicates while TypeScript, policies, and workflows execute constrained actions.
- MAGI System on Jev ObservedAn open-source three-sage voting experiment inspired by Neon Genesis Evangelion.
- Routing, classification, verification, and workflow control are the dominant early mental models. PlausibleThose categories recur in the collected launch discussion and align with the documented output primitives.
- Cascade router PlausibleA cheap decision chooses whether to use rules, a small model, a premium model, a specialist, or a human.