Availability

Metric	Target	Measurement
API availability	99.9% uptime	Health endpoint returns 200
Transform endpoint	99.9% success rate	Non-5xx responses / total requests
Rehydrate endpoint	99.95% success rate	Higher -- critical path

These are SLO targets (objectives a deployment aims to meet), not measured results. For indicative measured figures see Performance and Benchmarks. The NER targets are deliberately looser than typical measured NER latency to leave headroom for hardware, model, and payload variation.

Builtin-Only Mode (detector.mode: builtin)

Endpoint	p50	p95	p99	Max
POST /v1/transform	Under 5ms	Under 10ms	Under 20ms	Under 50ms
POST /v1/rehydrate	Under 0.1ms	Under 0.5ms	Under 1ms	Under 10ms
POST /v1/detect	Under 5ms	Under 10ms	Under 20ms	Under 50ms
GET /v1/health	Under 1ms	Under 1ms	Under 2ms	Under 5ms

Builtin + NER Mode (detector.mode: both)

Endpoint	p50	p95	p99	Max
POST /v1/transform	Under 200ms	Under 500ms	Under 1s	Bounded by the NER timeout
POST /v1/rehydrate	Under 0.1ms	Under 0.5ms	Under 1ms	Under 10ms
POST /v1/detect	Under 200ms	Under 500ms	Under 1s	Bounded by the NER timeout

Throughput

Mode	Single Instance	Horizontal Scaling
Builtin-only	>1,000 req/s	Linear with instances
Builtin + NER	20-50 req/s	Limited by NER sidecar

Payload Limits

Limit	Default	Configurable
Max request body (non-file routes)	10 MB	`server.max_body_size_bytes`
Max file/image upload	50 MB	`file_upload.max_size_bytes`
Max batch items	100	`limits.max_batch_size` (lower only; capped at 100)
Max session TTL	3600s (1 hour)	`session.ttl_seconds`
Max concurrent streams	Unlimited	OS/runtime limits
SSE heartbeat interval	15s (proxy only)	Proxy SSE keep-alive; the server stream endpoints send none

Error Budget

Error Type	Budget (per 1000 requests)
5xx errors	Under 1 (0.1%)
Detection false negatives	Under 50 (5%) for regex, under 100 (10%) for NER
Detection false positives	Under 30 (3%) for regex, under 150 (15%) for NER
Token repair failures	Under 10 (1%)
Session expiry (expected)	N/A -- by design

Degraded Mode SLO

When NER sidecar is unavailable (mode=both):

Metric	Guarantee
Availability	No hard dependency on NER -- the request still succeeds via builtin regex
Latency impact	One NER-timeout wait per request (`detector.timeout_secs`), then builtin-only
Entity coverage	Builtin types only; person, company, and location unavailable
Data safety	Unaffected -- PII protection maintained

Session Security

Property	Guarantee
Encryption	AES-256-GCM (AEAD)
Key strength	256-bit minimum
Nonce uniqueness	Random 12-byte per seal (cryptographically random)
Tamper detection	Authentication tag verified on every unseal
Replay protection	Enforced for transform continuations via a per-session-id strict-monotonic counter store (`replay_backend`: `memory` default, or `redis` for HA). Rehydrate and RAG context are exempt by design and remain TTL-bounded bearer tokens. Memory backend has a cold-start window on restart; Redis closes it across replicas. Running `redis` replay across replicas is a multi-replica deployment, so it also requires `revocation_backend`: `redis`, otherwise a value revoked on one replica stays restorable through another and the server refuses to start
Cross-tenant isolation	Tenant ID validated during unseal

Revocation

Property	Guarantee
Revocation latency	Immediate (in-memory + file)
Revocation persistence	Survives server restart (file backend)
Revocation consistency	Eventual (file backend)
Future restore suppression	100% -- revoked values always return [DELETED]

Monitoring

Required monitoring for SLO compliance:

# Prometheus alerts (see deploy/prometheus/alerting-rules.yml)
- guardai_transforms_total          # request rate
- guardai_errors_total              # error rate
- guardai_transform_duration_seconds # latency
- guardai_rate_limit_rejections_total # capacity
- guardai_prompt_security_triggers_total # security

Service Level Objectives