Small decisions. Interesting possibilities.Submit contentSubmit
evoke preview

evoke

A reflex runner that asks Jev which small program a spoken sentence should run and refuses when the weakest answer is unsure.

Added to Jevfast

How it uses Jev

The sentence, the reflex catalog and the values each reflex accepts form the state. A Choice question picks the reflex, further questions pick its arguments, and the weakest probability across those answers is compared with the confidence bar the user keeps in a file. Values must come from the user's words or from a list the user owns, and an action that cannot be undone asks for confirmation before it runs.

What Jev decides

Reflex choice for one sentence (illustrative)Example answers · not a recorded Jev response · Source ↗
Question 1 · reflex
YOUR APP
INSTRUCTION

Which installed reflex should answer this sentence, or none of them?

STATE

The state carries the utterance, the installed reflexes and the values each accepts. The documented session for "make it cosy" returned lights 0.45, none 0.40, timer 0.10 and volume 0.05, and the 0.45 top answer stayed under the 0.50 route floor, so nothing ran.

JEV · CHOICE
  1. lights
  2. timer
  3. volume
  4. route
  5. none
Question 2 · room
YOUR APP
INSTRUCTION

Which room does the sentence refer to?

STATE

The state carries the utterance, the installed reflexes and the values each accepts. The documented session for "make it cosy" returned lights 0.45, none 0.40, timer 0.10 and volume 0.05, and the 0.45 top answer stayed under the 0.50 route floor, so nothing ran.

JEV · CHOICE
  1. den
  2. kitchen
  3. bedroom
  4. office

App workflow

  1. Read the sentence

    The CLI, the SDK or a reflex picked from the package manager passes the utterance and the installed reflex catalog into the core.

  2. Ask the closed questions

    A Choice question selects the reflex, and follow-up questions select the values it accepts, with Jev answering each with a calibrated probability.

  3. Apply the bar

    The weakest answer is compared with the user's confidence line; below it, nothing runs and the ranking is printed.

  4. Check the values

    Every value must come from the user's own words or from a list the user owns, and an out-of-range value is asked for again rather than clamped silently.

  5. Run or confirm

    A non-destructive reflex at or above the bar executes its program; a destructive one always asks first, and the confirmation is recorded with the arguments it will use.

Why it is interesting

The gate is layered rather than a single score: a warm sentence with a flat ranking runs nothing, a value outside its range never reaches a program, and a destructive reflex asks even at 0.97, which turns calibrated probabilities into a permission system a user can inspect.