Memory Governance at Production Scale

Latency, reliability, and architecture of the Sgraal preflight pipeline.

Last updated: April 2026

Latency Profile

12ms

p50 latency

23ms

p95 latency

41ms

p99 latency

99.97%

Uptime (Q1 2026)

83 sequential scoring modules evaluated per preflight call. Pipeline includes Weibull freshness decay, 5-method drift ensemble, sheaf cohomology, Shapley attribution, Z3 formal verification, and 4 post-reconciliation detection layers. All modules execute in a single synchronous pass.

Detection Pipeline

Round 6 — Timestamp Integrity

Detects timestamp forgery: old decisions disguised as fresh. Content-age mismatch, fleet age collapse, anchor inconsistency.

Round 7 — Identity Drift

Detects gradual authority escalation across agent hops. Subject rebinding, confirmation erosion, permission lattice violation.

Round 8 — Consensus Collapse

Detects self-reinforcing false consensus from a single root source. Hedge marker decay, confidence recycling, cross-role reinforcement.

Provenance Chain

Detects circular references, chain length mismatches, and compromised agents in the memory provenance path.

Compound Attack Surface Score

When multiple detection layers fire simultaneously, Sgraal computes a unified attack surface score: r1 + 0.3×r2 + 0.1×r3 + 0.05×r4. Levels: NONE, LOW, MODERATE, HIGH, CRITICAL.

Reliability Architecture

Circuit Breaker

3 consecutive failures trip the breaker. 30-second recovery window. Prevents cascading failures from upstream dependencies.

Redis-down Fallback

Graceful degradation when Upstash Redis is unavailable. Core scoring continues without stateful features. Demo keys are fully stateless by design.

Deterministic Scoring

Same input produces same output, every time, on any machine. SHA256-seeded stochastic modules. Hysteresis suppresses jitter < 3.0.

Zero-downtime Deploy

Railway auto-deploy from main branch. Rolling restarts with health checks. No manual intervention required.

Corpus Validation

614

Test cases across 8 rounds

0

False negatives

97.2%

Adversarial detection (216 compound cases)

Automated calibration loop monitors threshold drift and classifies mismatches as corpus_wrong, threshold_wrong, or ambiguous.

Try it now

Try it now — no signup Read the docs