Decision boundaries in Jev workflows
A decision boundary in a Jev (TypeSafe System One) workflow is not a hyperplane inside the model. It is the closed answer space you wrote plus the thresholds your code applies to probabilities, confidence, or noul.
This explainer is for practitioners who already know Jev is not a chat model and want to place the cut between auto-act, review, and abstain. Independent unofficial notes — not TypeSafe docs. Confirm primitives on docs.typesafe.ai. We do not issue API keys.
Three surfaces that actually draw the cut
- Schema boundary — Choice options, Score levels, or a Noul yes/no. The model cannot emit a fourth department you did not list. That is TypeSafe’s type-safety claim: out-of-schema, not “true.”
- Distribution boundary — Choice/Score
confidenceis derived from how peakedprobabilitiesare. Official confidence page: a flat distribution means “I don’t know.” Noul has no confidence field; the cut is distance from 0.5 (you choose). - Policy boundary — your
ifstatements. TypeSafe’s examples use a ~0.5 confidence floor and a higher bar for irreversible actions. Those numbers are illustrations. Fit them on your labels.
Label-in-request (Jev) moves surface (1) without retraining. Train-your-own BERT moves it by changing weights. See Jev versus classification models.
Worked boundary: act / review / abstain
def route(choice_ans, *, read_floor=0.5, write_floor=0.9):
if choice_ans.confidence < read_floor:
return "abstain" # human or safe default
if choice_ans.choice == "approve_transfer" and choice_ans.confidence < write_floor:
return "review"
return "act"Do not copy floors across primitives. Official jaggedness: a Noul and a yes/no Choice on the same sentence need not match; a question and its negation need not sum to 1.
Where the boundary lies to a trained classifier
| You need | Prefer |
|---|---|
| Weekly label edits, no training team | Jev Choice criteria |
| Token spans, embeddings, on-prem only | Fine-tuned encoder |
| Hierarchical taxonomy with abort | Cascade of Choices + confidence abort |
| Exact counts, dates, money | Code, then Jev on the remainder |
jev-1.13 is literal. If the instruction says “refund mentioned” it will not infer “unhappy about fit.” The missing half of the sentence is the boundary.
FAQ
Is low confidence the same as class “other”?
No. other is a label you added. Low confidence means the mass is spread — including across other. Handle both.
Should I threshold the winning probability instead of confidence?
TypeSafe says you may use any statistic of probabilities. confidence is the convenience default.
Does pinning jev-1.13.0 freeze the boundary?
It freezes the weights behind the alias. Your thresholds still need a re-fit if you change criteria text.
Limits
No unpublished ECE curves. No claim that Jev’s boundary is “better” than BERT. Schema-safe ≠ correct.
Hub: Explainers. Siblings: classification comparison, confidence threshold glossary. Official: docs.typesafe.ai.
Sources
Public TypeSafe or adjacent documentation only. No private claims.