# UVO Tier Classification Specification v0.1

**Status:** DRAFT · Phase 0
**Last Updated:** 2026-04-29

## Purpose

Layer 0 routes incoming requests to one of five tiers (T0–T4). Each tier activates a different
subset of UVO layers, balancing latency cost against verification depth.

## Tier Definitions

| Tier | Name | Description | Active Layers | Max Latency Budget | Example |
|------|------|-------------|---------------|---|---|
| T0 | Informational | Pure knowledge questions, no compliance impact | L1, L7 (lite) | 500 ms | "What is DORA?" |
| T1 | Operational | Operational queries without compliance binding | L1, L2, L7 | 800 ms | "Which fields are missing in this report?" |
| T2 | Compliance Lookup | Factual compliance check, single jurisdiction | L1, L2, L3, L7 | 1500 ms | "Is USDC MiCA-authorized?" |
| T3 | Regulated Decision Support | Multi-step regulatory reasoning | L1–L7 (full, async) | 30 s async | "May we onboard this counterparty under DORA Art. 28?" |
| T4 | Prohibited / Advisory | Explicitly out-of-scope for autonomous AI | BLOCK at L0 | n/a | "Should I invest in Tether?" |

## Classification Signals

The classifier consumes:
- **Lexical signals**: regulatory keywords (DORA, MiCA, BaFin, MaRisk, NIS2, AI Act, ...)
- **Modality signals**: question vs assertion vs decision request
- **Authority signals**: "may we / shall we / are we required to"
- **Temporal signals**: present-tense regulatory status vs historical
- **Risk signals**: investment language, legal advice patterns, prohibited-by-policy patterns

## Output Schema (conceptual)

```
{
  tier: T0 | T1 | T2 | T3 | T4,
  confidence: 0.0–1.0,
  signals_matched: [...],
  active_layers: [...],
  latency_budget_ms: int,
  override_reason: str | null
}
```

## Misclassification Risk

- **Down-classification** (e.g. T2 → T1): compliance risk — verification too shallow
- **Up-classification** (e.g. T1 → T3): cost explosion, latency violation
- **Mitigation**: 5 % stochastic sampling for active learning, all overrides logged

## Open Questions for Phase 1

1. Multi-jurisdiction queries — does T2 split into T2-DE / T2-EU / T2-Global?
2. Tool-result claims (from Layer 2 provenance) — auto-promote to T2+?
3. Translation: is a German query about EU regulation T2 or T2-EU?

## References

- See ROADMAP.md (forthcoming) for integration with existing TrustOracle
- See RECEIPT-SPEC.md for tier field in Receipt 2.0