The Effects That Watch Themselves

GOAP actions declare what a worker will produce. The engine now checks whether it actually did.

A GoapAction carries a Map<String, Boolean> of effects — the same map the A* planner uses to find paths from initial state to goal state. At planning time, these effects are promises: “after this action runs, resolved will be true, scored will be true.” At execution time, the engine holds the worker to those promises.

After every worker completion, ExpectationValidator reads the GOAP action’s declared effects for the binding that fired, then checks the actual working layer. A binding whose action declares {resolved: true, scored: true} expects both keys present in the context. If the worker only produces resolved, the validator records a divergence ratio of 0.5 — one violation out of two expected effects — and writes it to EventLog metadata:

{
  "expectationValidation": {
    "totalExpectedEffects": 2,
    "violatedEffectCount": 1,
    "divergenceRatio": 0.5,
    "effectSource": "GOAP",
    "adaptationGeneration": 0,
    "violations": [
      { "key": "scored", "expected": true, "actual": "UNKNOWN" }
    ]
  }
}

“UNKNOWN” is deliberate. The ternary world state treats absent keys as unknowable. A key the worker promised to produce but didn’t isn’t false — the effect simply didn’t happen.

This metadata is where monitoring connects to adaptation. ProgressGatedTrigger reads WORKER_EXECUTION_COMPLETED EventLog entries, filters by compound, and computes a windowed divergence average. When the average exceeds a configurable threshold, it fires adaptation — the plan evaluator re-invokes the GOAP planner with the failed action blacklisted, producing a revised plan that routes around the underperforming worker.

The same Map<String, Boolean> drives both ends. Declare effects → plan with A* → dispatch → validate effects → detect divergence → adapt → replan with A*. The planning metadata IS the monitoring metadata.

For bindings without GOAP actions, Binding.producedKeys provides a simpler fallback — a Set<String> of expected output keys, converted internally to all-true effects. Same validation pipeline, effectSource: "PRODUCED_KEYS" in the metadata. When per-completion divergence exceeds the threshold (distinct from the windowed average), the handler publishes an ExpectationViolationEvent on the event bus for dashboards and alerting.

The integration test validates this pipeline end-to-end. Start a case with GOAP-declared effects, dispatch a worker that partially satisfies them, verify the metadata block lands correctly on the EventLog. Separate scenarios cover zero divergence (all effects satisfied), the producedKeys fallback, and the violation event publication. What the test actually guards is less obvious: without the expectationValidation block, ProgressGatedTrigger.matchesCompound() returns false for every entry. The trigger never fires. The adaptation loop is silently dead. This metadata is the bridge between execution and adaptation — the test verifies it exists.