Community · E2 · artifact verified

Drive the Aside browser with typed decisions

bside pilots the Aside browser with Jev instead of a chat LLM: each tick answers which action, which element, and whether the goal is met, over an action schema the pilot cannot hallucinate outside of.

01 · Role in the system

What Jev does here

Each tick the harness asks Jev typed questions - which action, which element, is the goal met - answered with calibrated probabilities in about 0.3 to 1 second at the author-quoted fraction-of-a-cent input pricing. Actions come from a schema code enumerates, so the pilot cannot emit a tool call that does not exist. The README publishes a same-browser, same-loop benchmark against two LLM engines: a 5-step search-and-navigate task where Jev took 52.8 seconds over 7 steps for $0.0018, against 72.9 seconds and $0.039 for a frontier model and 68.9 seconds for a light model that also clicked elements that were not on the page. The event log shows that failure signature directly - generated refs resolving in 16 milliseconds against 0.7 to 1.8 seconds for real clicks - and a replay player renders the runs side by side.

02 · Control boundary

Where Jev sits

Enumerated action schema to one typed decision request per tick; the aside CLI observes and acts while Jev only decides, and a bench harness replays every event.

Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.

03 · Known limits

What this evidence does not prove

  • The cross-engine benchmark is author-reported from the author's own runs; the harness and event logs ship in the repository for reproduction.
  • Requires the aside CLI signed in plus a TypeSafe key; the comparison engine needs a separate OpenAI-compatible endpoint.
  • One multi-step navigation task measured, not a suite.

04 · Attribution

Public sources

This is a Community record: the project was published by a third-party community author.