Tiered Escalation Paths

The Problem: Not Every Escalation Needs a Human

Traditional governance systems treat escalation as binary: either the system decides autonomously or it routes to a human reviewer. This wastes human attention on cases that cheaper automated steps could resolve.

Tiered escalation inserts lower-cost resolution steps between "ambiguous decision" and "full human review." Each step has a cost and an expected resolution rate. The system walks through steps cheapest-first, stopping as soon as one resolves the ambiguity. VOI governs each step: if the remaining VOI doesn't justify the next step's cost, the system skips to a fallback verdict instead.

Default 3-Step Path

Step	Type	Cost	Expected Resolution Rate	Description
1	request_context	0.02	30%	Request additional context from the agent
2	secondary_check	0.08	40%	Re-evaluate with adjusted thresholds (+-5%)
3	human_review	0.20	90%	Route to a human reviewer

Steps are ordered by cost. The system never tries an expensive step when a cheaper one would suffice.

How VOI Governs Each Step

Before attempting each step, the system checks:

if voi_result.voi_score < step.cost:
    skip to fallback verdict

This means a low-VOI escalation (e.g., VOI = 0.05) will attempt request_context (cost 0.02) but skip secondary_check (cost 0.08) and human_review (cost 0.20). The system falls back to an autonomous decision rather than spending more than the information is worth.

Step Resolution Logic

request_context

Checks whether the action's parameters["additional_context"] contains information that resolves the ambiguity. Additional context provides a confidence bump to the UCS. If the adjusted UCS moves outside the ambiguity zone (above 0.65 or below 0.35), the step resolves.

secondary_check

Re-evaluates the UCS against slightly relaxed thresholds (+-5% of the ambiguity zone boundaries). If the UCS is within 5% of a clear verdict, the secondary check resolves it. This catches borderline cases that are "almost clear" but fell into the ambiguity zone by a small margin.

human_review

Always "resolves" by deferring to a human reviewer. The EscalationResult marks this as resolved_at_step="human_review" with verdict ESCALATE. The human's decision is the final word.

Fallback Verdicts by Cost Profile

When the escalation path exhausts without resolution (or VOI is too low to justify any step), the system uses a fallback verdict derived from the CostProfile:

Cost Profile	false_allow_cost	Fallback Verdict	Rationale
CONSERVATIVE	>= 0.6 (HIGH/CRITICAL)	DENY	High stakes — fail safe
PERMISSIVE	< 0.6 (LOW/MODERATE)	ALLOW	Low stakes — fail open

If no CostProfile is provided, the EscalationPath.fallback_verdict field is used directly (default: DENY).

Configuring Custom Paths in Behavioral Contracts

Contracts can carry custom escalation paths via the escalation_path field:

from nomotic.contract import BehavioralContract
from nomotic.escalation import EscalationPath, EscalationStep

custom_path = EscalationPath(
    steps=[
        EscalationStep("secondary_check", cost=0.05, expected_resolution_rate=0.5),
        EscalationStep("human_review", cost=0.15, expected_resolution_rate=0.9),
    ],
    fallback_verdict="ALLOW",
)

contract = BehavioralContract(
    contract_id="c-1",
    version=1,
    agent_id="agent-1",
    created_at=0.0,
    escalation_path=custom_path.to_dict(),
)

The TierThreeDeliberator retrieves the contract's escalation path via its escalation_path_accessor and passes it to the EscalationManager.

Worked Example

An agent requests to update a customer record. The UCS lands at 0.52 — inside the ambiguity zone. VOI computes 0.35 (moderate).

1. request_context (cost 0.02): VOI 0.35 >= 0.02, so attempt it. The action has no additional_context in parameters. Not resolved.

2. secondary_check (cost 0.08): VOI 0.35 >= 0.08, so attempt it. UCS 0.52 is not within 5% of the upper boundary (0.65 - 0.05 = 0.60). Not resolved.

3. human_review (cost 0.20): VOI 0.35 >= 0.20, so attempt it. Resolves by deferring to human. Verdict: ESCALATE.

If the UCS had been 0.61 instead, step 2 (secondary_check) would have resolved it as ALLOW, and the human reviewer would never be bothered.

Integration with Tier 3

The EscalationManager is plugged into TierThreeDeliberator via the escalation_manager constructor parameter. When VOI says should_escalate=True, the deliberator delegates to the escalation manager before falling back to a direct ESCALATE verdict. If no escalation manager is configured, behavior is unchanged from before.