GUIDES

Scoring Engine

How Aether ranks runtimes — weights, intent multipliers, and explain output.

Formula¶

Each allowed runtime receives four dimension scores (0–1):

Dimension	Default weight	What it measures
Cost	0.25	Resource size vs runtime economics
Performance	0.25	CPU/memory/GPU fit
Reliability	0.25	Health, restart, isolation
Availability	0.25	HA, scaling, runtime uptime profile

Total score = weighted sum × intent multipliers × workload-class adjustments.

Weights are configurable via EngineConfig / ~/.aether/config.toml.

Workload classification¶

The engine classifies specs before scoring:

Class	Triggers
Stateless	Default web/API
Stateful	PVC / persistence block
GpuCompute	`requirements.gpu` set
Batch	Cron/batch patterns
General	Fallback

Classification filters ineligible runtimes (e.g. GPU → KubeVirt bias).

Intent multipliers¶

When intent: is present in the spec:

Goal (low-latency, cost-optimized, high-throughput, balanced) shifts weights
SLA latency/availability thresholds filter or penalize runtimes
Budget caps favor lower-cost runtimes
Resilience High boosts Kubernetes HA
Compliance isolation/encryption flags favor KubeVirt/K8s

Run aether intent --spec workload.yaml to compare with and without intent.

Explain mode¶

CLI¶

aether decide --spec examples/workload.yaml --explain

Prints score table plus per-runtime + reasons and - warnings.

API¶

curl -X POST http://localhost:5090/api/ai/recommend \
  -H "Content-Type: application/json" \
  -d '{"yaml":"<workload yaml>", "explain": true}'

Dashboard¶

AI Engine page and Intent Debugger call explain: true and show all runtimes.

Confidence¶

Confidence reflects score separation between first and second ranked runtimes. Close scores → lower confidence; update spec constraints or intent for a clearer winner.

Examples — sample traces
AI Purpose — what is rule-based vs roadmap
Code: src/ai/scoring.rs, src/engine.rs