Vocabulary-Aware Compliance — Delegation and Escalation Observation Semantics
Delegation and escalation were the compliance dimensions I’d been avoiding.
Not because they’re hard to implement — the implementation turned out to be small. Because they’re hard to define. The #85 behavioral contracts spec explicitly deferred both: delegation needs “observation window semantics” and escalation needs “vocabulary-aware interpretation.” Both sounded like they’d require new infrastructure.
They didn’t.
The observation window question dissolves once you realise the existing TTL + threshold model already provides windowing. VIOLATED signals expire after 90 days by default. The probe threshold is 3 per-dimension. So “3 delegation failures within 90 days → BehavioralViolation” — that’s a window. No new concept needed.
The real question for delegation was never how to window observations but what constitutes a single observation. And that’s the engine’s call, not eidos’s. The engine sees task characteristics — complexity, decomposability, whether delegation was warranted. Eidos says “this agent expects to delegate.” The engine decides when to check.
Escalation was the interesting one. AgentDisposition.autonomy() is an open String — “directed”, “semi-autonomous”, “autonomous” from ConscientiousnessTerm, or completely different values from domain vocabularies. BehavioralExpectations.escalationExpected() needs to interpret that string to determine “should this agent escalate when uncertain?” But eidos-api can’t depend on casehub-eidos-vocab. The vocabulary terms are in a different module.
The solution: add impliesSupervision() as a default method on VocabularyTerm itself. Returns false by default. Vocabulary authors override it for terms that indicate supervised behavior. ConscientiousnessTerm.DIRECTED returns true. AUTONOMOUS inherits false. DISC terms get their own overrides — STEADINESS (maps to DIRECTED) returns true, DOMINANCE (maps to AUTONOMOUS) returns false.
escalationExpected() resolves the autonomy string through VocabularyRegistry, gets the term, calls impliesSupervision(). If the vocabulary isn’t registered or the value doesn’t resolve — false. Fail open. An agent without vocabulary-grounded disposition shouldn’t be penalised for not escalating.
The design review caught something I’d missed: DISC terms need their own impliesSupervision() overrides, not reliance on cross-vocabulary equivalence. The axisExactMatch bridge declares term equivalence, not behavioral property inheritance. STEADINESS being equivalent to DIRECTED doesn’t mean it automatically inherits DIRECTED’s impliesSupervision() return value. Each vocabulary independently declares the behavioral properties of its own terms. A cross-vocabulary consistency test validates they don’t drift.
The review also pushed for a convenience overload — escalationExpected(AgentDescriptor, VocabularyRegistry) that internalises the vocabUriForAxis(AUTONOMY) call. The 3-param version stays for callers with pre-resolved URIs. Both overloads exist; the engine picks whichever fits its observation path.
Five commits, 384 insertions across api, vocab, and runtime. The probe pipeline didn’t change — Step 6 already counts all VIOLATED signals regardless of dimension key. Adding ComplianceDimension.DELEGATION and ComplianceDimension.ESCALATION as string constants was enough to make the probe handle both dimensions.
The engine-side observation logic — deciding what constitutes a delegation opportunity or an escalation trigger — is tracked in engine#645. That’s where the policy decisions live: task complexity for delegation eligibility, uncertainty detection for escalation triggers. Eidos defined the contract; the engine implements the observation.