The neocortex cognitive subsystem has had a half-built feedback loop for weeks. GoalPrioritizationPhase detects when a goal’s confidence has dropped — sets a decay-signal property on the MindMap node. CognitiveGoalOrchestrator reads those signals and produces GoalRevision records. But nothing consumed them. The revisions accumulated every tick, the eidos goals stayed ACTIVE, and the decay-signal never cleared.

The interesting part wasn’t the code — it was mapping the data flow across three repos to find exactly where the loop broke. Five gaps, three repos:

  1. Nobody called pendingRevisions() on the orchestrator
  2. No code transitioned an AgentGoal.lifecycleState in eidos
  3. No GoalLifecycleProvider implementation read from eidos (only a no-op existed)
  4. AgentRegistry had no partial-update method — only full descriptor register()
  5. No idempotency guard existed for re-consumption between ticks

The fix is three pieces that each do one thing. A default method on AgentRegistry.updateGoalLifecycleState() does targeted read-modify-write without rebuilding the entire descriptor graph. SocialAvatarCognition.consumeGoalRevisions() maps “dormant” to DORMANT and “abandon” to ABANDONED after each tick. And EidosGoalLifecycleProvider bridges eidos goal states back to neocortex so GoalResolutionPhase.sync() can clear the signal.

The part I didn’t expect: neocortex’s sync mechanism was already complete. GoalResolutionPhase.sync() calls GoalLifecycleProvider.getLifecycleStates(), imports any returned status, and removes decay-signal. The only thing missing was a real provider to call. One @ApplicationScoped bean in blocks, and the loop closes.

The updateGoalLifecycleState default method is worth noting as a pattern. Adding it to the interface rather than doing the read-modify-write at the call site means InMemoryAgentRegistry can eventually override with a computeIfPresent and JpaAgentRegistry with a direct SQL UPDATE — but the default works correctly now via the existing findById + register path. The user wanted this scoped in rather than deferred, and it turned out to be the right call — the API is cleaner and the implementations can optimize independently.

<
Previous Post
Catching Foreign Exceptions Without a Shared Type Hierarchy
>
Next Post
The Invisible 66%