Layer 2 of the autonomous intelligence stack gave agents the ability to propose their own goals. It deliberately left every self-generated goal at SECONDARY priority — behind case-assigned work, never competing for the agent’s attention when real tasks exist. D6 in the Layer 2 spec was blunt about why: a drive that’s running hot — say, accumulated curiosity from dozens of low-retention memories — could produce a PRIMARY goal that preempts a case worker’s assigned caseload. That’s not governance. That’s a runaway feedback loop with production consequences.

Layer 3 narrative identity changed the equation. An agent that has built a self-narrative — “I am the person who helps in crises,” backed by episodes and themes with measurable salience — has something Layer 2 lacked: provenance for priority. The escalation isn’t “my curiosity score is high”; it’s “this goal aligns with who I understand myself to be, and that understanding has been confirmed across multiple independent narrative synthesis cycles.”

Where the SPI Lives

The original design sketch (D11 in the #142 spec) placed the escalation policy in engine-api — a GoalPriorityEscalationPolicy SPI that DefaultGoalFormationService would call at registration time. The idea was clean: engine governs priority, blocks provides the implementation.

The dependency graph disagreed. The escalation decision needs NarrativeState, DriveProfile, and DriveGoalProposal — all blocks types. Engine-api can’t depend on blocks. Putting the SPI there would mean either generic Map<String, Object> attributes (losing type safety at the governance boundary, which is exactly where you want it most) or polluting eidos-api with types it shouldn’t own.

The fix was straightforward: the policy lives in blocks, where its inputs live. GoalProposalOrchestrator evaluates escalation before converting proposals to GoalFormationProposal.ProposedGoal, setting suggestedPriority = PRIMARY when warranted. The engine already respects suggestedPriority — no engine change needed for the escalation path itself. For defence-in-depth, DefaultGoalFormationService gains a one-line guard: reject PRIMARY on drive-sourced goals without an escalatedBy attribute. The policy produces the provenance; the engine verifies it exists.

Synthesis Cycles as the Governance Clock

The hardest constraint in the governance contract is sustained alignment. A single tick where a theme’s salience is high and its axis weight matches the goal — that’s noise. The narrative could shift next cycle. Wall-clock time is also wrong: two hours of silence don’t mean the narrative was re-evaluated and the alignment confirmed. They mean nothing happened.

The right clock is the narrative synthesis cycle. Each time NarrativeSynthesiser runs its LLM and produces a new NarrativeState (gated by the composite trigger — enough new reflections, sufficient novelty, or quiet period bypass), synthesisedAt advances. That’s a genuine re-evaluation. If the theme-goal alignment survives two of those, the escalation has earned its rank.

Demotion works the same way in reverse. A PRIMARY goal whose escalation theme fades — the “crisis-helper” theme drops below salience threshold or its curiosity axis weight goes negative — starts a demotion counter. Two synthesis cycles of sustained misalignment, and the goal returns to SECONDARY. Theme-specific demotion: we check the original escalation theme, not whether any theme happens to support the axis. The provenance is the contract.

Cross-Axis Composition

The other Layer 2 deferral was cross-axis goals. Per-axis mappers produce one goal per drive: “explore knowledge gaps” from curiosity, “reconnect with subject X” from affiliation. But a DerivedTheme with significant positive weights on both curiosity and affiliation suggests a compound goal — “learn about X by engaging with Y.” The theme is the compositional signal that single-axis mapping can’t see.

Detection is heuristic: scan themes for ≥2 positive axis weights above a configurable threshold. The dominant axis becomes the proposal’s axis() (for deduplication and escalation alignment checking), with the full composition preserved in crossAxisWeights attributes. An optional CrossAxisGoalEnricher SPI — backed by LlmCrossAxisGoalEnricher via AgentProvider — replaces the mechanical “Compound goal: connector across CURIOSITY and AFFILIATION” with something contextually rich. The heuristic detects; the LLM describes. Same layering pattern as the rest of the stack.

The Pipeline

GoalProposalOrchestrator grew from a two-phase pipeline (map drives to goals, rank by intensity) to five phases. The new middle three — cross-axis composition, escalation evaluation, demotion evaluation — all gate on narrative availability. When NarrativeOrchestrator isn’t wired in, they silently skip. Every drive-sourced goal stays SECONDARY. The agent works, the goals form, the capacity limits hold — exactly as Layer 2 designed it.

When narrative IS available, the same orchestrator becomes a governed escalation pipeline. A GoalProposalTick.Changes — renamed from Proposed to reflect the four signal types it now carries — bundles proposals, abandonments, priority adjustments, and governance attribute updates into a single atomic output. The scheduler applies all four. The compositor guarantee holds: no side effects, no direct writes, just signals.

What This Opens

The GoalEscalationPolicy SPI is a @FunctionalInterface. NarrativeGoalEscalationPolicy is the first implementation — narrative alignment as the escalation evidence. A future HumanApprovalEscalationPolicy could gate PRIMARY on explicit human confirmation via the oversight system. The SPI doesn’t care what the evidence is. It cares that evidence exists and that the engine can verify it.

The deeper question is whether sustained narrative alignment is the right bar for priority escalation, or whether it’s just the first bar we can measure. An agent whose “crisis-helper” theme has persisted across five synthesis cycles has something an agent with a momentary curiosity spike doesn’t. Whether that something is sufficient to preempt case-assigned work — that’s a deployment decision, not a platform one. The thresholds are configurable. The governance is structural.

<
Previous Post
Org Annotations — Closing the Loop
>
Next Post
Type Safety as Documentation