# 03 — Notification Timing & SLA Model

The 7 May 2026 incident's defining failure was a **~40-minute** gap between a
visible air threat and the public cell-broadcast reaching Rezekne. This package
treats notification speed as a **first-class, modelled quantity** rather than an
implicit expectation.

## The SLA chain

1. `Decision_NotificationUrgency` takes `threatLevel` and the estimated minutes
   until the object could reach people, and outputs:
   - `notificationTier` — `immediate | priority | standard | none`
   - `broadcastSlaSeconds` — the **maximum** time allowed from authorisation to
     confirmed dispatch.
2. `Task_DispatchBroadcast` (VUGD) carries a **non-interrupting boundary timer**
   armed at `broadcastSlaSeconds`.
3. If dispatch is not confirmed before the timer fires, `Task_EscalateSla`
   runs **in parallel** — it does not cancel the broadcast, it escalates the
   delay to the Crisis Management Centre so a human intervenes.

## Proposed SLA defaults

| Threat level | Lead time | Tier | `broadcastSlaSeconds` | Rationale |
|--------------|-----------|------|----------------------:|-----------|
| critical | any | immediate | 120 | Object inside airspace, closing on people. |
| high | ≤ 15 min | immediate | 180 | Little lead time; alert must precede impact. |
| high | > 15 min | priority | 300 | Lead time exists; brief verification window. |
| elevated | any | priority | 600 | Threat present, not yet tracking toward people. |
| low | any | standard | 1800 | Advisory only. |
| none | – | none | 0 | No public broadcast. |

Against these defaults, the Rezekne broadcast (~2400 s) breached even the most
lenient non-zero SLA **by an order of magnitude**. Under this model the boundary
timer would have fired at 120–180 s and escalated to KVC ~38 minutes earlier.

## Guardrail

`broadcastSlaSeconds` is a **hard floor for escalation, not a target for
dispatch** (guardrail GR-3). Dispatch should normally be far faster; the SLA is
the line past which the absence of a broadcast becomes its own incident.

## Open for validation

The table values are **modelling proposals**. They need calibration against
real NBS classification timelines and VUGD cell-broadcast dispatch latencies
before the package can leave `draft`.