Strategic reference. Last updated for v3.4.0. Drawn from the Sentinel positioning memo (April 2026).
In April 2026 the European regulated economy has a structural problem that has built up over two years and becomes an existential question in the next eighteen months.
The American AI platforms — OpenAI, Anthropic, Google, Microsoft — have won the frontier-model layer. That is irreversible. The Chinese open-weight models — DeepSeek, Qwen, GLM, Kimi — have broken the cost layer. The cost difference between open-weight and proprietary APIs sits at ten-to-thirty times. Europe did not win the model layer and will not. Mistral is real and significant, but an order of magnitude smaller than Anthropic and structurally capped. Aleph Alpha has repositioned as an orchestration stack. European AI investment is a fraction of American investment.
What Europe has is the regulatory layer. The EU AI Act enters full force on 2 August 2026. DORA is in force. NIS2 is in force. BaFin BAIT 8 and MaRisk AT 9 have been substantially tightened. BSI IT-Grundschutz is binding for critical infrastructure. This is not a burden. It is the only strategic position Europe structurally still holds in the AI era, and it will be copied worldwide the way GDPR was copied.
Between these three realities — American models, Chinese cost, European regulation — sits the European regulated institution and does not understand what to do. It pilots. It does not scale. Every quarter it loses strategic time against greenfield startups that compete without its constraints. That is not a tool problem. That is an infrastructure problem.
A platform-engineering lead at a German tier-1 bank put it in six sentences in April 2026: "Agents is a huge topic for us, governance and security included. Feels completely unplannable — you run a workshop and three new tools have shipped before it ends. Microsoft and Entra-ID look most promising. Including Copilot as an agent platform. Everything in motion. We're starting with small PoCs."
Six sentences that contain the entire strategic dilemma. The institution sees the problem, has run the workshops, is piloting small, and is on the way to adopting Microsoft — not because Microsoft is the best solution but because it is the only one that feels as if it would work. The buyer is not careless; the buyer has no better option.
The same pattern shows up across the top 150 DACH enterprises in banking, insurance, public-sector IT, defence logistics, and industrial manufacturing. They see the pattern. They pilot. They do not scale. They slow down. Greenfield startups that do not carry the regulatory load win accounts at their competitors. The accumulated IT landscape turns from an advantage into a structural disadvantage.
Meanwhile, European startups build modern AI applications — autonomous systems, defence reconnaissance, precision industrial — that regulated integrators would like to procure but cannot, because the compliance cycles are incompatible with the startups' development speed. In the US this integration problem is solved by capital. In China it is solved by state coordination. Europe has neither mechanism. The integration between European innovation and European institutional infrastructure does not happen, and that is why Europe strategically falls behind.
European regulation is read by most decision-makers today as a blocker. That is wrong. It is the way out.
The EU AI Act forces American and Chinese AI systems deployed in Europe to demonstrable decision records under EU jurisdiction. That is not only a compliance burden. It is the legal grounding on which a European CISO can, for the first time since the cloud era, say "no" to American standard platforms — without having to argue against the business case. "BAIT 8.2 requires evidence under our jurisdiction" is not the burial of an idea. It is the opening of an alternative.
The EU AI Act and its successor regulations convert regulatory requirements into strategic levers — if the infrastructure exists to make them operational. "We satisfy the regulation" is defensive. "We use the regulation to scale faster than our competitors" is offensive and commercially stronger.
Regulation, operationalised as infrastructure, makes established enterprises competitive again against greenfield startups. Not because the regulation goes away — it does not go away, it will increase — but because it can be transformed from process overhead into infrastructure overhead. Process overhead slows things down. Infrastructure overhead speeds things up, because every future regulation builds on the same base instead of being redistributed.
Sentinel is the evidence infrastructure for the regulated AI era — for regulated European enterprises, enabling them to move like startups and prove like banks.
This thesis has three operative levels that engage in sequence and build on each other.
Individual value. Sentinel is a tool for the single institution. It lets a regulated enterprise deploy AI systems faster because the regulatory provability arises automatically instead of being distributed across processes. The six-week compliance review every bank knows becomes a three-day routine check. The PoC that otherwise dies in pilot scales to production.
Ecosystem value. Sentinel is a bridge in mixed ecosystems. When an established integrator and a fast-moving startup must work on the same system, Sentinel sits between them. The startup develops quickly; Sentinel records every decision in a format the integrator's compliance processes accept. The integrator does not need to match the startup's pace. The startup does not need to learn the integrator's bureaucracy. The infrastructure layer mediates.
Structural value. Sentinel is Europe's technical answer to the integration problem the US solves through capital and China solves through power. Europe has neither the capital for acquisition-driven consolidation nor the political structure for mandated integration. So it must happen through infrastructure — through a common language for AI decisions that both fast-moving startups and established institutions can speak.
These three levels are not three products. They are three zoom levels of the same thing. The infrastructure that delivers level one is exactly the infrastructure that enables level two, which is exactly the infrastructure that carries level three.
Four verbs. A causal chain. Together the complete lifecycle of an AI decision in a regulated environment.
Trace. Every AI decision is captured with context,
input hash, timestamp, and outcome. Append-only, tamper-
resistant, privacy-by-default (SHA-256 hash instead of raw
input, unless explicitly configured), stored under the
operator's jurisdiction. Sentinel's @sentinel.trace
decorator delivers this today. Apache 2.0. One-line install.
from sentinel import Sentinel
sentinel = Sentinel() # local storage, zero config
@sentinel.trace
async def approve(request: dict) -> dict:
return await your_agent.run(request) # unchanged
await approve({"amount": 50_000})
# → DecisionTrace persisted: data_residency=local,
# inputs_hash=sha256:..., policy_result=ALLOWPolicy enforcement runs in the same critical path — every AI decision passes through a policy layer before it executes. Policy-as-Code — OPA/Rego or Python. Versioned, testable, CI/CD-deployable. The EU AI Act Art. 14 kill-switch primitive is first-class. Multi-party approval gates are available.
Attest. Every decision trace is cryptographically attested — SHA-256 over the canonical payload, signed with the operator's key, hash-chained to the prior attestation within its agent namespace, optionally RFC-3161 timestamped via EU- sovereign TSAs (DFN-CERT, D-Trust). Post-quantum ML-DSA-65 (FIPS 204) is available for long-term retention scenarios. Attestations are portable, JSON-serialisable, and independently verifiable.
Audit. Every attestation is independently verifiable
without vendor dependency. Integrity checks, hash-chain
verification, and counterfactual replay are deterministic and
offline. The auditor reproduces verification from the
artefact alone — no call-home, no cloud, no lock-in. The
sentinel.audit API exposes decisions for review: policy
compliance, counterfactual analysis, and regulator access.
Comply. Every record, every policy decision can be aggregated into an evidence pack — cryptographically signed, mapped to EU AI Act, DORA, NIS2, and BSI IT-Grundschutz, in a format an auditor accepts. The basic generator is open source. Enterprise features (HSM signing, RFC-3161 timestamping via EU TSAs, long-term retention of ten-plus years, automated BaFin reporting templates, legal-hold management) are commercial.
This formula is distinctive. Dome Systems uses "Connect. Secure. Operate." as its agent-governance lifecycle. That is good but general — it could apply to IoT or classical microservices equally. Trace. Attest. Audit. Comply. is specific to the AI-decision context and leads directly to the evidence artefact European regulation requires.
Sentinel is one product with four functional modules. The modules share a single codebase, a single data model, a single CLI, and a single installation. They are not four products. They are four faces of the same product, each serving a different buyer archetype.
Your business logic
│
▼
┌─────────────────────────────────────────┐
│ SENTINEL │
│ Evidence infrastructure for AI │
│ │
│ ┌────────┐ ┌────────┐ ┌────────┐ │
│ │ TRACE │ │ POLICY │ │EVIDENCE│ │
│ │ ✓ │ │ ✓ │ │ ✓ │ │
│ │ Trace │ │Enforce │ │ Attest │ │
│ └────────┘ └────────┘ └────────┘ │
│ │
│ ┌───────────────────────────────┐ │
│ │ FEDERATION → roadmap │ │
│ │ Multi-institution aggregate │ │
│ └───────────────────────────────┘ │
└─────────────────────────────────────────┘
│
▼
DECISION LAYER (your choice)
LLMs · ML classifiers · Rule engines · Robotic systems
Switch anytime. No lock-in.
│
▼
LOCAL STORAGE (your infrastructure)
SQLite · PostgreSQL · NDJSON
Sentinel Trace is the entry point for developers. The
@sentinel.trace decorator, install via python3 -m pip install --user sentinel-kernel, value in five minutes: the
developer sees what their AI agent decided, with which inputs,
under which policy evaluation, in what time. This module stays
Apache 2.0 permanently. It is the distribution flywheel. The
developer does not pay, but they carry Sentinel into their
team, and the team into the organisation.
Sentinel Policy is the entry point for platform- engineering teams. OPA/Rego-based or Python-rule-based policy enforcement on every traced call. The base engine stays Apache 2.0. Commercial are preconfigured industry policy libraries — BaFin catalogues, KRITIS rulebooks, defence use-case templates, EU AI Act Art. 5 prohibition lists. The value is consistency across models and systems. Deploying a new AI system does not mean "new compliance processes" but "apply existing policies."
Sentinel Evidence is the entry point for CISOs, DPOs, compliance officers, and internal audit. Evidence packs as PDF and NDJSON, cryptographically signed, mapped to EU AI Act, DORA, NIS2, BSI IT-Grundschutz. The default generator stays Apache 2.0. Commercial enterprise features are HSM integration, RFC-3161 timestamping via EU timestamping authorities, multi- party signing for critical decisions, long-term retention of ten-plus years (BaFin retention obligation), automated BaFin reporting templates, legal-hold management.
Sentinel Federation is the entry point for multi- institution scenarios: holding-group structures, industry consortia, supervisory oversight. This is the future module, not shipping today but architecturally anchored and publicly committed as roadmap. It will enable aggregated compliance visibility across federated institutions without individual data disclosure. Relevance: holding-company subsidiaries, supervisory oversight over banking groups, KRITIS sector coordination, the bridge between fast movers and slow institutions at larger scale.
The modules are not invented by marketing. They correspond
exactly to the structure that already exists in the code:
sentinel/core/ has the tracer and storage backends;
sentinel/policy/ has the evaluator with OPA integration;
sentinel/compliance/ has the checkers for EU AI Act, DORA,
NIS2, BSI, plus the evidence-pack generator; sentinel/core/ attestation.py has the signing infrastructure;
sentinel/federation/ exists as a roadmap placeholder with a
defined interface.
Sentinel has always been four things. Until v3.2 we described the first three as a two-and-a-half-layer kernel (Trace + Govern, with Evidence bundled under Govern) and used Route / SovereignRouter as a placeholder for the fourth layer. From v3.3 onward we use four explicit names. Evidence stands alone because it is what regulated buyers pay for. Federation replaces Route as the roadmap slot because multi-institution aggregation is the larger strategic question; routing between models is a narrower capability that may re-emerge as a Federation sub-feature, not an independent layer.
Errors in positioning arise from scope creep, not from under-positioning.
Sentinel is not an AI platform. Sentinel hosts no models, calls no LLMs, runs no inference. Sentinel is the layer that traces, attests, audits, and delivers compliance evidence for the decisions of other AI systems.
Sentinel is not a complete compliance solution. EU AI Act compliance spans Articles 9 through 17 and more. Sentinel primarily addresses Articles 12, 13, and 14, and partially 9, 11, 16, 17, 26, and 72. The rest organisations must cover themselves or with other tools. This honesty builds trust; the alternative (overclaiming) loses it.
Sentinel is not an observability tool. LangFuse, LangSmith, Weights & Biases do observability. Sentinel tracks decisions and policies. Complementary.
Sentinel is not a security platform. NeMo Guardrails, LLM Guard, garak filter prompts and detect jailbreaks. Sentinel comes after the security layer: what did the system that passed security decide, and is the resulting artefact provable.
Sentinel will not pursue a closed-source strategy. The kernel stays Apache 2.0 permanently. Enterprise features are always add-ons to a permanently open core. This commitment is not negotiable — it guarantees the BaFin BAIT 8.2 exit capability regulated buyers require.
Sentinel will never be US-incorporated. The company stays Swentures UG under German jurisdiction. An American subsidiary for US market access can exist later, but the core entity and the evidence infrastructure stay EU-incorporated.
Dome Systems — US-incorporated, closed beta, building an agentic infrastructure platform with significant venture backing. Intellectually strong positioning through three essays. Dome will take the US market and western Fortune-500. Sentinel does not compete directly with Dome; the markets are geographically and jurisdictionally separated. Dome cannot effectively serve regulated European institutions because any US-incorporated entity falls under the CLOUD Act (18 U.S.C. § 2713), regardless of data-centre location. BaFin BAIT 8.2 exit capability is not credibly demonstrable under closed SaaS, even with escrow.
Microsoft Agent Governance Toolkit (AGT). US-incorporated, MIT-licensed. Covers OWASP Agentic Top 10, framework-agnostic with LangChain/CrewAI/Dify integration. Microsoft AGT wins every Microsoft-shop without a sales call. Sentinel is not a direct competitor in Microsoft-first shops without an explicit EU-jurisdiction requirement. Sentinel is the answer for enterprises that must work model-agnostically or must hold evidence artefacts explicitly under EU jurisdiction. This is often a co-existence in the reality of European banking, not a replacement.
Cylake. US-incorporated, founded by Nir Zuk (Palo Alto Networks). "Sovereignty is the next firewall." Targets nation states and tier-1 banks with hardware + software. Cylake is not a direct competitor but is currently claiming the word Sovereignty in the cybersecurity context with billions of marketing weight. The semantic shadow is real. That is why Sentinel does not lead with Sovereignty — it leads with Provability, operationalised as Trace. Attest. Audit. Comply., with Sovereignty as a consequence rather than a thesis.
Native cloud governance — Azure AI Foundry, AWS Bedrock Guardrails, GCP Vertex AI. Deeply integrated. Excellent when an organisation stays on one cloud. Loses value the moment multi-cloud or on-premise scenarios appear. Sentinel is the multi-cloud-plus-on-premise clasp.
Native AI-provider governance. Anthropic, OpenAI, Google will build proprietary governance features into their models. These will work excellently — per model. They will not be structurally model-spanning, because that contradicts their business model. They will also not be under EU jurisdiction, because the firms are US-incorporated. Sentinel sits one layer above and abstracts.
LangFuse, LangSmith, Weights & Biases. Observability. Measure performance, tokens, latency. Complementary, not competing. LangFuse ships a Sentinel-compatible integration.
Sastrify, Enactia, Kovrr, ADOGRC — GRC platforms. Not competitors. Partner candidates. They operate on the inventory-and-management layer. Sentinel operates on the runtime layer. The layers complement each other.
The gap in this landscape is observable. No other solution simultaneously offers: model-spanning consistency, EU jurisdiction for evidence, open source with exit capability, cryptographic provability in regulation-mapped format. Sentinel positions precisely in this intersection.
Three independent clocks align in 2026.
The regulator. EU AI Act Annex III high-risk enforcement begins 2 August 2026. Penalties up to €15M or 3% of global annual turnover. Every regulated European buyer must demonstrate automatic tamper-resistant decision logging from that date. The procurement conversations for Q4 2026 are happening now.
The sovereignty moment. After two years of CLOUD Act headlines and Schrems-era jurisprudence, "EU-jurisdictional" has stopped being marketing and started being a purchase criterion. BSI publishes reference profiles. Linux Foundation Europe exists specifically to give open-source projects an EU-governed home. The infrastructure to be credibly operated under EU jurisdiction exists for the first time.
The empty field. Nobody is building this exact thing. Large US observability tools are observability tools, tied to LLMs. Cloud providers are cloud providers. The EU-operated projects we respect — Mistral, Aleph Alpha, Haystack, Nextcloud — are model vendors, orchestration stacks, or collaboration platforms, not decision infrastructure. Sentinel is the only open-source, EU-operated, technology- agnostic decision infrastructure we can find. That will not last forever. It should last long enough to become the default.
Sentinel will not be Mistral. Sentinel will not be Anthropic. Sentinel will not be HashiCorp in the pre-IBM era. Those scales require venture capital in US or Chinese ecosystems and conditions Europe structurally does not offer.
What Sentinel can realistically become, at a good outcome over five-to-seven years: a category that must exist because regulation forces it to, occupied by an EU-incorporated vendor because US competitors structurally cannot occupy it.
This is not "change the world." It is "occupy a structural position in Europe that a US competitor would otherwise take, and give Europe an option it otherwise would not have." That is an honest, realistic, achievable ambition. It is smaller than the maximum formulated in session-high moments. It is bigger than most European infrastructure firms have ever managed.
Scale what you can prove. Move like a startup. Prove like a regulated bank.
The rest is implementation.
Full competitive detail: docs/landscape.md · Phase plan: docs/roadmap.md · Provability conditions: docs/provability.md · What Sentinel is not, extended: docs/commercial.md.