Workflow guide

Decision-Native Computing: Why Tomorrow's Software Separates Judging from Generating

Why the future of software has two distinct AI calls: Generation Calls (to produce text) and Decision Calls (to choose which code branch to run). How HA-Jev, Kev (0.5B on-device), and the Decision Runtime pattern turn sensor states into real-world automated actions.

1. Step 1

   Two kinds of AI calls: Writing essays vs. flicking switches

   Most people think AI is only for drafting emails, synthesizing documents, or generating code patches—these are open-ended Generation Calls where heavy models spend seconds producing prose. But production software and hardware (IoT, CRM, factories, trading engines) rarely need prose; 99% of the time, they only need a Decision Call: should we notify the user? Is this sensor anomalous? Which branch should the code execute?

   

2. Step 2

   The washing machine loop: Real-world HA-Jev home automation

   The open-source project HA-Jev provides a textbook proof-of-concept for Home Assistant: device telemetry is fed as Jev state—'power = 0W, cycle completed 45 minutes ago, door = closed'. Instead of an expensive LLM writing a chat response, Jev answers a single Noul question: 'Did the user forget the laundry?' Jev returns a decisive 0.92 probability in milliseconds, triggering an instant mobile push notification.

   {
     "model": "systemone",
     "state": "Appliance: Washing Machine. Power: 0W. Cycle status: finished 45 minutes ago. Door state: closed.",
     "questions": [
       {
         "type": "noul",
         "id": "laundry_forgotten",
         "instruction": "Based on the idle power and closed door after cycle completion, did the user likely forget the laundry?"
       }
     ]
   }

3. Step 3

   The critical safety gate: Never let AI unlock doors unattended

   If an AI errs on forgotten laundry, the inconvenience is trivial. If an AI errs on unlocking a physical door or transferring funds, the impact is catastrophic. HA-Jev demonstrates the vital Safety Gate rule: low-risk actions execute automatically at high confidence; but high-risk operations (e.g. door locks, alarms, file deletions) are strictly intercepted to require explicit human PIN confirmation.

   {
     "model": "systemone",
     "state": "Event: Visitor detected at front door. Recognized face: family member (confidence 0.81). Action requested: unlock_front_door.",
     "questions": [
       {
         "type": "choice",
         "id": "lock_action_policy",
         "instruction": "Evaluate security risk for this automated door unlock request.",
         "options": [
           {
             "id": "auto_allow",
             "label": "Low risk, allow automatic unlock"
           },
           {
             "id": "require_pin",
             "label": "High security asset, require user passcode verification"
           },
           {
             "id": "block",
             "label": "Unrecognized or dangerous, block completely"
           }
         ]
       }
     ]
   }

4. Step 4

   Offline on-device: 0.5B tiny models (Kev) make decisions local and free

   Jev's System One interface is becoming a software paradigm. Projects like Kev (by Jared Palmer) package Qwen 0.5B/0.8B with a custom readout head that runs on a Mac, iPhone, or edge device without internet. In SemIf benchmarks on an RTX 3090, reading option logits directly without token generation delivers a 5.2x systems speedup over traditional JSON-generating models.

5. Step 5

   The Decision Runtime pattern: Pluggable backends with policy gates

   Modern software architecture is adopting the Decision Runtime: upstream sensors and events feed into an engine that dynamically routes questions to hosted cloud Jev, local edge Kev (0.5B), or deterministic code rules. A unified Confidence Gate and Security Policy layer evaluates every verdict—high-confidence safe actions execute instantly, while uncertain or high-risk actions escalate to heavy LLMs or human reviewers.

   

6. Step 6

   Universal applicability: Swap the washing machine for your business

   The washing machine loop is not just a smart-home novelty. Replace the appliance telemetry with CRM customer churn signals, server latency spikes, construction site camera alerts, or e-commerce refund requests. The architecture remains identical: Sense state -> Decide branch -> Enforce gate -> Act. Decoupling fast decisions from slow generation makes your entire product lightning fast.

All guides · Request builder