System status · Beta
The infrastructure we sell is the infrastructure we run.
This is the operational record. Latency and throughput figures come from the same system processing grāmatr's own requests in production. Uptime is shown as our SLA target — live per-day availability monitoring is being wired to that production telemetry now (see below).
grāmatr Cloud: available now. Private cloud and on-premises: committed September 1, 2026. Fully air-gapped: committed January 1, 2027. SOC 2 Type I: targeting October 15, 2026 — not certified today.
Live status
Request path — all five layers
grāmatr operates as five functions in a single request path. All five must be operational for sessions to be classified, governed, and attributed. The status of each layer is reported independently.
Classification Engine
Intent pre-classification before model response
Context Delivery
Organizational context assembly in the request path
Policy Engine
Governance evaluation and enforcement layer
Cost Attribution
Per-session attribution at the moment of inference
Session Intelligence
Compounding classifier training from routed sessions
grāmatr API
Request routing and model endpoint orchestration
Service level commitment
The contract grāmatr holds itself to
The 99.9% SLA applies to the combined request path: classification, context delivery, policy enforcement, cost attribution, and session intelligence. All five layers. Not each layer independently.
Uptime commitment
99.9%
Measured against the full request path. No partial credit for individual-layer availability.
Pre-classification latency
<100ms
End-to-end across all five layers before model response. Measured at the request boundary.
Failure behavior
Fail-open
If grāmatr is unreachable, requests pass through to the model. No single point of failure added to your AI estate.
Uptime history
90-day record
Live per-day uptime monitoring is being instrumented now. Until it is connected, the 99.9% figure is grāmatr's SLA target — not a measured rolling average — and the bars below represent the target operational state, not live telemetry.
Incident history
Every incident, on the record
When a layer degrades or fails, the entry goes here: what happened, which component, how long, and how it resolved. The incident log is not curated for presentation — it is the operational record.
The dogfood record
Real applications run on grāmatr in production.
Real applications run on grāmatr in production — Anneal first, the CPP (Covington Place Partners) financial-governance application in active development as the second. Every request they make runs through the same classification, context delivery, governance, attribution, and intelligence layers reported on this page.
Before any external customer deployment, grāmatr was already running real applications under production load. The system on this status page is not a demo environment maintained separately from what is being sold — it is the same system. The full production record, with the classification, memory, and learning counts, is on the proof page.
What that means for the operational record: grāmatr's classification latency is measured from the platform's own production traffic — ~26ms typical for intent classification, ~135ms for the full contextual intelligence packet. The 99.9% uptime figure is our SLA target; the live availability measurement is being instrumented now. Anneal runs on grāmatr first, and this status board is being wired to that production telemetry.
Infrastructure design
How grāmatr is built to stay up
Failure behavior
Fail-open by architecture
If grāmatr becomes unreachable, requests continue to the model without governance applied. No interruption to your AI estate. The governance gap is logged and resumes on reconnection — it does not become a service failure.
Latency
Sub-100ms classification
Intent classification completes in ~26ms (p50) / ~37ms (p95). The full contextual intelligence packet — classification, enrichment, and contract assembly — returns in ~135ms typical, in the request path before the model responds.
Data isolation
Row-level security at the database layer
Each organization's session data is isolated by architecture, not by application-level policy. RLS is enforced at the database layer. No cross-tenant data is accessible at the application layer — the boundary is enforced before any query runs.
Session retention
60 days by default · configurable
Session content — the context delivered and the interaction payload — is retained for 60 days by default, configurable to your application's needs. The governance audit trail (classification outcome, policy decision, cost attribution, and intelligence signal) is retained separately and longer, satisfying the EU AI Act log-retention floor.
Model endpoints
Model-agnostic routing
The request path is independent of the model endpoint. Claude, GPT, Gemini, and Llama are all supported. Classification, context delivery, governance, attribution, and audit trail are consistent regardless of which model processes the response.
Deployment
Every deployment tier
Multi-tenant grāmatr Cloud, private cloud, on-premises, and air-gapped on-premises. The operational architecture — and the SLA commitments — are scoped to grāmatr Cloud. Private cloud and on-premises SLAs are defined in the deployment agreement.
Questions about an incident or SLA applicability?
If you are evaluating grāmatr infrastructure for a regulated deployment — financial services, healthcare, government — and need SLA evidence for your compliance documentation, book a conversation. We will walk through the operational record with you directly.
Book a Conversation →