Sentinel is the Trace-to-Trust kernel that turns EU AI Act compliance from a blocker into a baseline. Every agent decision is traced, attested, and sealed at runtime — so your teams stop fighting audits and keep shipping.
In most regulated organisations the AI stack is already assembled from third-party tools of different vendors: governance tools, observability platforms, LLM providers, identity systems. Each layer lives with its own vendor.
What is missing is the layer that turns this heterogeneous stack into a regulatorily usable whole — the knot-resolver that produces the single audit package a regulator can verify, without replacing any of your existing investments.
Every agent decision becomes a signed, hash-linked attestation. Every evidence pack PDF carries a PAdES signature with EU-sovereign RFC-3161 timestamp. Every chain of decisions can be verified independently by your auditor — air-gapped, offline, no vendor lock-in.
Sentinel maps directly to EU AI Act Art. 12 (record-keeping) and Art. 17 (quality management). It supports BSI IT-Grundschutz preparation, BaFin and MaRisk requirements, and any regulator who asks the fundamental question: “Can you prove what your AI did, and when?”
Langfuse, Datadog, Arize, and LangSmith give you performance, cost, and drift signals. Sentinel gives you legally durable evidence of what each agent decided.
Microsoft AGT, OPA, Cedar, and Bedrock Guardrails enforce policies at runtime. Sentinel seals the cryptographic proof that those policies were applied.
When an auditor asks for courtroom-defensible proof, that is the question Sentinel answers — and the question no observability or governance tool was designed to solve.
Scaffold a local pilot, run ten decisions through @sentinel.trace, write a signed PDF evidence pack, score yourself against EU AI Act Art. 12.
[pdf] pulls reportlab + pyhanko for evidence-pack PDFs; [pades] is [pdf] without reportlab; [pqc] adds ML-DSA-65 post-quantum signing via oqs-python.
Sentinel automates EU AI Act Art. 12/13/14/17 — the logging, transparency, oversight, and quality-management obligations. Other articles require organisational action. We mark the split honestly.
sentinel audit-gap to see the exact split for your setup.
EU AI Act enforcement applies to decisions that touch rights, access to services, safety, or meaningful financial outcomes. The architecture is technology-agnostic; the sectors below are where procurement is already active.
A German bank runs credit decisioning, fraud scoring, AML, and transaction approval through Sentinel. When BaFin requests evidence on model drift under BAIT §6.3, they deliver signed evidence packs in minutes — not weeks of log reconstruction.
Procurement, logistics, dual-use assessment. Air-gapped and classified deployments with VS-NfD path. Cryptographic evidence that survives supply-chain scrutiny and export-control audits.
Enterprise software firms embed AI into products sold to regulated customers. Sentinel gives their customers the evidence packs needed to deploy those products under the EU AI Act — without reinventing compliance infrastructure per vendor.
Benefit eligibility, permit approval, critical-infrastructure AI. Statutory transparency under NIS2 and sector law. Evidence packs that regulators, courts, and citizens can independently verify.
Quality control, predictive maintenance, robotic decisioning. Standards-aligned retention across plant lifetimes — 15+ year evidence archival in formats that outlive every vendor swap.
Sentinel wraps each agent decision as it happens. Input, policy, output, and jurisdiction are bound into a signed attestation before the next call begins. No log collection, no post-hoc reconstruction, no manual mapping under audit pressure.
An auditor can only trust evidence as far as they trust whoever produced it. Sentinel runs as an independent layer — not owned by the operator, not owned by the LLM vendor, not owned by the cloud provider. The signature chain holds whether anyone trusts you or not.
Sentinel does not replace your governance tools, observability platform, or LLM provider. It sits between them — receiving policy results, enriching with traces, emitting signed evidence. Bidirectional by design: allow signals feed innovation, kill signals gate risk.
Regulatory retention runs ten years or longer. Your LLMs and agents will not. The evidence layer must outlive every model, every framework, every vendor swap — stable signature format, stable storage interface, stable regulatory mapping.
We are building Sentinel the way HashiCorp built Terraform: primitive hooks in the kernel, the ecosystem grows through the community. Three stages, honestly labelled.
Production-ready, Apache 2.0, 923 tests passing. Install with pip install sentinel-kernel.
[pdf] extra)
Optional ML-DSA-65 post-quantum signing ([pqc] extra)
LangChain · CrewAI · AutoGen · Haystack · FastAPI · Django integrations
SQLite · PostgreSQL · Filesystem storage backends
OpenTelemetry span export
Air-gapped / VS-NfD / EU-sovereign deployment paths
BSI IT-Grundschutz preparation scaffolding
Four architectural items raised by design partners. All four have committed design docs under docs/architecture/. Each item re-ships to production only after passing our fresh-venv E2E verification harness — discipline over shipping speed.
The devil sits in the details of complex enterprise landscapes. These integrations we build with the community, not for it.
Sentinel is in design partnership with a select group of regulated enterprises building production-grade agentic AI under the EU AI Act. Request a working session with our team, or explore the code on GitHub.