Operational Runtime Model

Reference documentation for Operational Runtime Model in Adaptive Skills.

Adaptive Skills should define runtime contracts before it builds runtime machinery.

This document describes how a consumer runtime may use the capability layer. It is not an implementation of a runtime engine.

Lifecycle

1. Intake
2. Context classification
3. Capability routing
4. Depth selection
5. Execution
6. Checkpoint or gate
7. Validation
8. Handoff or closure
9. Execution record
10. Evolution signal

Runtime context

A runtime context should be small enough to hand off and explicit enough to review.

{
  "task_id": "optional-local-id",
  "goal": "Turn vague feature request into a small executable slice",
  "risk_level": "medium",
  "ambiguity_level": "high",
  "selected_capabilities": ["workflow", "feature-planning"],
  "activated_modules": ["ambiguity-check", "scope-boundary"],
  "mode": "workflow/extended",
  "human_approval_required": false,
  "evidence_expected": ["clarifications", "smallest-slice", "acceptance-evidence"]
}

Checkpointing

A checkpoint should happen when continuing by momentum is riskier than pausing.

Typical triggers:

• context is too large;
• task shape changed;
• proof is unclear;
• next step crosses ownership;
• human approval may be needed;
• execution mode needs to deepen or shrink.

Use checkpoint-review for local pauses. Use AletheIA when the pause is a macro gate or governance decision.

Resumability

A resumable execution should leave:

• current goal;
• selected capabilities and modules;
• what was proved;
• what remains unresolved;
• next recommended step;
• risks, blockers, or human decisions.

Use handoff-summary when the next round needs a clean restart.

Observability

The first layer of observability is an execution record, not telemetry automation.

The record should help decide:

• did the selected capability fit?
• was the selected depth proportional?
• did evidence support closure?
• did a routing hint fail?
• should the Evolution Layer receive an observation?

Approval boundaries

Risky or state-changing actions belong to the runtime and governance layer, not the capability metadata.

Adaptive Skills may say:

human decision gate required

But the actual approval mechanism belongs to AletheIA or the consumer runtime.