<15ms
end-to-end Gateway latency, p95. Measured; per-component budgets in docs.
5
LLM providers supported on day one
Five capabilities

Each is a deployed product, not a roadmap claim.

Five capabilities, delivered by seven components that all ship today. They deploy together as the BladeRun Platform, or individually, in the order that fits your AI roadmap. Most teams start with Gateway and Time Machine, add Overseer + Kill Switch in the second pilot phase, then opt into Federation.

01VerifySee and attest every agent action.Every prompt, response, tool call, and sub-agent spawn is captured and verifiable before it executes. Delivered by the Gateway in the AI call path, the Sentinel SDK inside your agent frameworks, and BladeRun.js on customer-facing surfaces.
02BindBind the agent to the mandate a human authorized.Cap, payee, purpose, expiry, aggregation: every action is checked against the mandate the human actually granted. Delivered by the Gateway policy plane, the mandate engine.
03EnforceAllow, step up, or block at the action boundary.Allow, step up, or block at the action boundary, and contain on anomaly. Two planes work together. The mandate engine is deterministic: same input, same verdict, every time, auditable line by line. Overseer AI is statistical: baselines and drift detection that trigger step-up and containment, never silent allows. On breach, the Kill Switch isolates the session in milliseconds, all at the Gateway boundary.
04ProveSigned, replayable evidence under your keys.An immutable, cryptographically signed record of every agent action, written to your S3 bucket under your KMS keys and your retention policy. Delivered by the Time Machine.
05FederatePrivacy-preserving cross-institution threat intelligence.Threat signals shared across institutions as rules every member receives, with no raw data ever leaving any member. Delivered by the Federation Network.
Where each component lives

Inside your perimeter. Outside it. And in the call path between.

Sentinel SDK runs inside your agent process. Gateway sits at your perimeter. Overseer, Kill Switch, and Time Machine run in your environment (SaaS-managed VPC, your own VPC, or air-gapped). The Federation Service runs in BladeRun infrastructure as the rule distribution channel, but all data published into it is one-way, hashed, and pre-image resistant.

In the agent's action path. Out of your payment authorization path. Fail-open or fail-closed, your policy.

Your environment self-hosted VPC · your own VPC · or air-gapped, under your keys
In the call path
Your AI application Sentinel SDK Gateway
Runs alongside, never gating the response
Overseer AI Kill Switch Time Machine Your S3 · KMS · IAM
↓ outbound LLM call, leaves your perimeter
External · model providers
OpenAIAnthropicAzure OpenAIAWS BedrockGoogle Vertexself-hosted
Federation Service · the one BladeRun-hosted piece one-way · hashed · pre-image resistant
Threat ResearchHoneypot NetworkCross-Member Correlation

The Federation Client pulls signed rule updates every 60 seconds and publishes only pre-image-resistant hashes upward. No raw data (prompts, tool calls, evidence) ever leaves your environment.

Architectural invariant

No generative LLM has authority to block, isolate, or escalate.

Two planes. The mandate engine is deterministic: same input, same verdict, every time, auditable line by line. Overseer is statistical: baselines and drift detection that trigger step-up and containment, never silent allows. The autoregressive LLM that explains decisions runs after both planes have acted, never in front of them. No safe harbor: SR 26-2 leaves agents out of scope, examiners still act on unsafe practice, and this architecture gives you the evidence trail they follow.

Principle 01

Deterministic mandate engine

The mandate engine and its rule set are deterministic: same input, same verdict, every time. Overseer's statistical baselines (z-scores, n-grams, drift) run as a separate plane. Each is independently auditable with documented inputs, transformations, and FP-rate metrics.

Principle 02

Generative LLM as narrator only

Small autoregressive model (under 8B params) writes the human-readable post-incident narrative. Cannot fire the Kill Switch. Cannot modify a detection. Zero enforcement authority by architecture.

Principle 03

Maturation, not magic

Every detection threshold ships in shadow → alert-only → enforcement progression with explicit SOC sign-off at each stage. Every transition is logged to Time Machine.

Latency budget

What runs on the call path. What runs alongside. What runs after.

The Gateway latency budget is the only thing your users feel. Detection, scoring, logging, and federation contribution all run alongside or after; they do not gate the response. We publish per-component p95 and p99 budgets in the documentation.

L1
Identity verification. Signed agent session validated against active certificate.
<1ms
L2
Tool-permission check. Requested call matched against agent's signed tool certificate.
<1ms
L3
Injection classifier. Structural classification on prompt content. CPU-bound; runs in parallel with L4.
~6ms
L4
DLP redaction + endpoint registry. 12 entity classes; deterministic per-session tokenization. MCP endpoint hash lookup.
~7ms
async
Overseer scoring. Behavioral baseline scoring runs alongside the call. Anomaly threshold breach fires Kill Switch via Redis pub/sub.
off-path
async
Time Machine write. Full session record written to your S3 bucket; signed and indexed.
off-path
async
Federation contribution. Hashed signal extracted, k-anonymity check, batched for upstream publish.
off-path

On-path total: under 15ms p95. Detection, logging, and federation all run alongside the call without blocking the response.

Two topologies, one envelope

Bank-built agents calling out. External agents calling in.

Both directions are real production patterns at the Tier 1 banks we've spoken with. Both are inside BladeRun's coverage envelope. They look different in the call diagram and different in what we can see.

Topology A · Outbound

The bank built the agent

Prompts, tool definitions, orchestration code all run inside your perimeter; the agent calls outbound to an external LLM. We see everything: prompt, tool calls, sub-agent spawns, response.

Bank Agent → Sentinel SDK → Gateway →
External LLM (OpenAI / Claude / Bedrock / …)
Coverage: full call shape, prompt, tool calls, response
Topology B · Inbound

An external agent calls in

Claude, ChatGPT, Operator, Gemini, or any third-party agent calls inbound into your MCP servers or APIs. We see what we see, and our cross-institution correlation matters most.

External Agent (Claude / Operator / Gemini / …) →
Gateway → Bank MCP / API surface
Coverage: call shape, signed identity, mandate scope, behavioral baseline
Deployment

Run it as SaaS, or in your own environment.

Same platform, same detection, same SLA. You choose where it lives. Merchants and teams that want speed take the hosted SaaS; banks and sovereign workloads run it on-premises.

SaaS · managed cloud

We run it in a dedicated VPC.

BladeRun operates the platform in a single-tenant VPC colocated with your model providers. Nothing to run, live in days, and your evidence still lands under your keys.

Where it runsA dedicated, single-tenant BladeRun VPC, colocated with your LLM providers.
Your keysYour S3, your KMS, your IAM. Signed evidence never leaves your control.
ComplianceSOC 2 Type II in scope; Type I available now under NDA.
Time to liveDays. Turn it on with one routing rule; reversible the same way.
FederationFull rule channel; signal contribution is opt-in.
On-premises · your environment

Your VPC, or fully air-gapped.

The entire platform runs inside your environment under your own IAM. No agent traffic or evidence ever leaves your perimeter, including a fully air-gapped option for sovereign workloads.

Where it runsYour VPC, your own data center, or fully disconnected (air-gapped).
Data flowNothing leaves your perimeter. Traffic, prompts, and evidence stay in.
ComplianceMeets OCC and Federal Reserve constraints on data flow.
Air-gappedOffline, signed rule updates; Federation runs local-only.
Runtime dependencyNone on BladeRun infrastructure. The platform runs standalone.
Common questions

What architecture leads ask first.

Can we deploy components individually?
Yes. Most banks start with Gateway + Time Machine in the first pilot. That gets you full visibility and audit-grade evidence with the minimum surface area. Overseer + Kill Switch follow after the baseline has matured. Federation is opt-in. The components are designed to deliver value individually and compound when combined.
Does Sentinel SDK work with our agent framework?
SDKs ship for Python and TypeScript, covering LangChain, LangGraph, LlamaIndex, AutoGen, CrewAI, Semantic Kernel, and the OpenAI/Anthropic native agent runtimes. Java and .NET ship in 2026. Integration is OpenTelemetry-native; if your framework already emits OTel spans, integration is configuration-only.
What's the failure mode if the Gateway is unreachable?
Your policy decides. BladeRun sits in the agent's action path, not in your payment authorization path, so you set the behavior: fail-closed rejects agent calls when the Gateway is unreachable, fail-open lets them proceed. Regulated workloads typically choose fail-closed; lower-risk workloads often choose fail-open. The Gateway is stateless behind a load balancer, so rolling deploys and node failures are handled without traffic loss.
How does this integrate with our existing SIEM?
Time Machine emits structured detection events to your SIEM in OCSF format. Splunk, Sentinel, Chronicle, and Datadog connectors are out-of-box. The Gateway also exposes a streaming feed for real-time SOC dashboards. The Time Machine record itself stays in your S3; the SIEM gets the index, not the payload.
Can we run BladeRun behind our existing API gateway?
Yes. BladeRun Gateway sits between your application and the LLM provider. It does not replace your edge gateway (Apigee, Kong, AWS API Gateway, etc.). The deployment topology supports Bank App → Edge Gateway → BladeRun Gateway → LLM. Or Bank App → BladeRun Gateway → LLM, with the edge gateway only on the inbound API surface.
Where is BladeRun's own infrastructure hosted?
For SaaS deployments, AWS us-east-1 and us-west-2 are GA; eu-central-1 and ap-southeast-2 follow in 2026. The Federation Service runs in the same regions, behind a separate isolation boundary. On-premises deployments do not depend on BladeRun infrastructure for runtime, only for periodic Federation rule pulls (which are signed and cacheable for air-gapped environments).
See the platform in your environment

One routing rule. Seven components. Visible in 48 hours.

Pilot the Gateway and Time Machine in one workflow. The other four components light up incrementally as the baseline matures. Reversible by routing rule.