Browser skill ecology

Use case

A browser skill ecology runs a small set of local specialists for privacy-sensitive or low-latency tasks. It can route simple tasks locally, escalate complex tasks to a server, and collect local evaluation traces without centralizing raw user data.

Components

Flow

FUNCTION browser_skill_request(input, local_registry, policy)
    contract <- CLASSIFY_REQUEST(input)
    candidates <- local_registry.FIND(contract)
    candidates <- FILTER_BY_LOCAL_BUDGET(candidates, policy.device_budget)

    IF candidates IS EMPTY
        RETURN ESCALATE_TO_SERVER(input, reason: "No local candidate")
    END IF

    selected <- SELECT_LOWEST_COST_VIABLE(candidates)
    output <- RUN_LOCAL(selected, input)
    RECORD_LOCAL_TRACE(selected.digest, input.summary, output.summary)
    RETURN output
END FUNCTION

Breeding loop

Do not mutate models directly in the browser at first. Use the browser to collect consented traces and failure summaries. Generate descendants offline or on a controlled server, then ship signed packages back to clients through the registry.

Constraints

Browser deployments should overweight memory, download size, startup time, battery, and privacy. A model that is impressive in the cloud may be unfit locally. The browser ecology is the clearest example of why frugality is not optional.

Safety

Local does not automatically mean safe. Validate package signatures, deny unexpected network access, expose model identity to the user, and provide a clear reset path that removes local packages and traces.