Skip to content

Uncertainty and gates

Aionis should not always turn memory into an immediate first action.

Sometimes the strongest runtime behavior is to say:

there is not enough evidence yet; inspect, widen, or rehydrate first

That is what the uncertainty layer and gate surfaces are for.

What this layer changes

Uncertainty is not just a score. In Aionis, it can suppress an overconfident task start, escalate planning, and tell the host what to do next before execution begins.

low / moderate / highrecommended actionsgate actionplanning escalation

What uncertainty looks like

The runtime exposes:

  • level
  • confidence
  • evidence_gap_count
  • reasons
  • recommended_actions

When the runtime needs a stronger intervention, planning surfaces also expose a structured gate:

  • gate_action
  • instruction
  • primary_reason
  • preferred_rehydration

The four gate actions

These are the important gate types:

GateWhat it means
inspect_contextRead the current planner-facing context before acting
widen_recallPull a broader recall surface because the current retrieval is too thin
rehydrate_payloadBring colder payload detail back into active use before acting
request_operator_reviewAsk for explicit human or host review before execution proceeds

What changes at task start

Without gates, a runtime can sound more certain than it really is.

With gates, Aionis can:

  1. suppress an unsafe or weak first_action
  2. escalate from kickoff to planning context
  3. expose the right next move for the host or operator

That makes the startup path more honest and usually more useful.

Minimal planning example

ts
const planning = await aionis.memory.planningContext({
  tenant_id: "default",
  scope: "repair-flow",
  query_text: "repair billing retry serializer bug",
  context: {
    goal: "repair billing retry serializer bug",
    task_kind: "repair_billing_retry",
  },
  tool_candidates: ["bash", "edit", "test"],
});

Read these fields first:

  1. planning.planning_summary.action_retrieval_uncertainty
  2. planning.planning_summary.action_retrieval_gate
  3. planning.kickoff_recommendation

If the gate says rehydrate_payload, the runtime is telling you that colder memory exists and should be restored before you trust a direct first step.

Why this matters

This is one of the clearest differences between a memory layer and a decision layer.

A memory-only system can retrieve context.

A gate-aware runtime can decide that:

  • the current context is not enough
  • more recall is needed
  • colder payload should be restored
  • a host or operator should review the move

That is the difference between sounding informed and behaving responsibly.

  • memory.actionRetrieval(...)
  • memory.taskStart(...)
  • memory.planningContext(...)
  • memory.contextAssemble(...)
  • hostBridge.planTaskStart(...)

Deep dives

Self-evolving continuity runtime for agent systems