HYBRID DECISION BENCHMARK

Rules for the obvious.
Jev for the ambiguous.

Eight labeled visitor journeys. One transparent baseline. Expected answers are withheld from Jev until scoring.

Run live with your own key ↗
FUNCTION ONLYOne fixed priority tree
63%

5 of 8 matched the labeled rubric

VS
FUNCTION + JEVRules route; Jev resolves conflict
100%

8 of 8 matched the labeled rubric

API EFFICIENCY4/8

cases handled by function without Jev

4 judgments · 254 ms total
THE TEST SET

Every answer is inspectable.

The baseline always prioritizes page-count first, then UTM/referrer, then default. The hybrid uses functions for clear cases and sends only conflicts to Jev.

Visitor journeyExpectedFunction onlyHybrid result
01
Launch click after casual reading

Active Product Hunt campaign vs. three editorial page visits

CampaignLabeled rubric
Returning×
Campaignjev
02
Developer evaluation

Deep history normally means returning, but every signal is technical

BuilderLabeled rubric
Returning×
Builderjev
03
High-intent return

No acquisition signal and repeated commercial exploration

ReturningLabeled rubric
Returning
Returningfunction
04
Internal campaign tag

UTM exists, but it marks internal navigation—not acquisition

ReturningLabeled rubric
Returning
Returningfunction
05
Fresh paid-search visitor

Strong acquisition signal and no meaningful history

CampaignLabeled rubric
Campaign
Campaignfunction
06
Low-context direct visit

No evidence for a specialized experience

BuilderLabeled rubric
Builder
Builderfunction
07
Tagged GitHub README

A UTM-tagged link that is still clearly developer traffic

BuilderLabeled rubric
Campaign×
Builderjev
08
Newsletter click during evaluation

Fresh email click vs. an established pricing and security journey

ReturningLabeled rubric
Returning
Returningjev
MVP ARCHITECTURE

The function is the frame.
Jev fills the gap.

01

Function extracts facts

Normalize referrer, classify UTM, count meaningful paths, and detect obvious cases.

02

Jev resolves conflict

Only ambiguous combinations become typed, three-way Jev choices.

03

Function enforces safety

Validate the answer, restrict it to three variants, and fall back on failure.

FAIRNESS CHECK

Recorded live Jev run from September 20, 2026. Expected labels were used only for scoring after the response and were not included in Jev state or questions. This proves behavior on the labeled MVP set, not conversion lift.

OPEN SOURCE

Verify it on your machine.

Clone the repository, add your own TypeSafe key locally, and rerun every case.

View on GitHub