Two runtimes. One control plane.

Chronicle splits enforcement from intelligence. The enforcement runtime runs where your decisions are made — on your infrastructure. The intelligence layer runs in shared cloud, aggregating signals across your estate.

Split-plane design

The critical enforcement path never leaves your environment. Intelligence aggregates separately, with full evidence sovereignty.

Customer Infrastructure

Chronicle Rust Runtime

:8300

Enforcement, LLM gateway, tool gateway, HTTP sensor

Vault

:8200

AppRole auth, Transit key wrap, secrets KV

PostgreSQL

:5433

Local receipt ledger, Merkle anchors

Redis

:6379

Event stream, session state, flush queue

MinIO

:9000

Encrypted evidence object store

Shared Cloud (STOE-managed)

Chronicle Server (FastAPI)

:8080

Decision ingest, tenant query API, auth middleware

Analytics Rollup Worker

5s cadence, computes approval rates, ML coverage, risk scores

Drift Worker

Hourly baseline comparison, behavioral deviation scoring

Alert Worker

Hourly policy breach surfacing, daily intelligence briefing

PostgreSQL (multi-tenant)

:5432

RLS-enforced canonical tables per tenant

Events flow: Client Runtime → Redis stream → batch flush → Server API → ingest table → materializer → canonical tables → query API

The hot path — built for milliseconds.

Layered verdict engine

Four parallel verdict layers: rule engine, policy engine, structural checks, and out-of-distribution signals. Each layer produces a trace. The combiner selects the authoritative verdict and retains all rejected paths as negative lineage.

Fire-and-forget evidence

/enforce returns in milliseconds. Evidence encryption and event emission happen synchronously but inline. The heavy persistence path (Redis → Postgres) is async via the flusher worker — never on the critical path.

Session causality

Every request carries a session ID. The session store in Redis chains decision IDs, enabling a full decision DAG across agents and tool calls. You can trace multi-hop agent behavior from a single root action.

Identity binding

JWT extraction middleware (v1.2) validates tokens via JWKS, classifies actors as human / service / agent, and injects identity into enforcement context. Every decision knows who authorized it.

LLM gateway

Transparent OpenAI/Anthropic passthrough. Normalizes streaming SSE. Emits pre-call / post-call bookend events. Tracks token counts, latency, and model identity. Drop Chronicle in front of any LLM endpoint.

MCP tool gateway

Proxies MCP upstream tool calls. Enforces pre- and post-call policy on every tool invocation. Emits tool call attestation events. Works with any MCP-compatible tool server.

We hold encrypted blobs.
You hold the keys.

Chronicle generates a per-call data encryption key (DEK). Evidence is encrypted AES-256-GCM on the client side before the event is emitted. The DEK is wrapped using Vault Transit — only the wrapped key envelope travels to the server.

Without your Vault, no one can read your evidence — including us. Revoke your Vault keys and the data becomes permanently inaccessible. You own your entire decision history.

01

Decision made

Agent action triggers Chronicle enforcement endpoint

02

Evidence built

Request context, identity, verdict, and traces assembled

03

DEK generated

Per-call AES-256 data encryption key created client-side

04

Evidence encrypted

Evidence blob encrypted with DEK — never leaves unencrypted

05

Key wrapped

DEK wrapped via Vault Transit; key_envelope stored with event

06

Event emitted

Encrypted blob + key envelope sent to server — plaintext never exposed

Four levels of decision replay.

Chronicle doesn't just record what happened — it records what you need to reconstruct, verify, and challenge it.

L0

Signal plane

live

Intent and outcome captured. Sufficient for compliance reporting and trend analysis.

L1

Evidence replay

live

Stored request replayed against stored rules. Verify the verdict was correct at the time of the decision.

L2

Receipt verification

live

Merkle tree + inclusion proof. Cryptographically prove a decision record has not been altered since capture.

L3

Deterministic replay

in progress

Bit-perfect input snapshots. Full environment reconstruction for forensic analysis of edge cases.

L4

Counterfactual sim

roadmap

Policy variant testing. Answer 'what would have happened if the policy was different' for any historical decision.

Ready to see the architecture in action?

We walk through a live Chronicle deployment with every interested team.

Request a demo