EcoKure DTL is an enterprise verification and evidence platform for systems that use probabilistic AI to produce recommendations, decisions, content or actions.
The platform separates two responsibilities:
- AI systems interpret, reason, generate and explore.
- DTL determines what may be accepted, served, recorded or acted upon.
The language may vary. The accepted semantic state does not vary when the canonical task, evidence, policy and state are identical.
request or system event
-> probabilistic interpretation
-> canonical task and state
-> deterministic DTL lane
-> specialist gate and policy checks
-> evidence pack and attestation
-> ALLOW / BLOCK / ABSTAIN / HUMAN REVIEW
-> natural-language or system rendering
This repository is the public EU AI Act application pack and a dependency-free reference verifier for the broader EcoKure DTL architecture. It is not a legal certificate, conformity assessment, notified-body service or substitute for a company's legal, safety, security and governance responsibilities.
Organisations are putting probabilistic systems into workflows where an answer can become a customer communication, a security decision, an engineering change, a clinical recommendation or a regulated record. Common controls fail when they:
- treat model confidence as evidence;
- ask another model to approve the first model;
- lose the policy, data and ruleset used for the decision;
- silently route a request into the wrong risk or domain lane;
- accept conflicting candidates without recording the disagreement;
- retain a screenshot instead of a replayable evidence record; or
- confuse a passing test with a legal, scientific, medical or operational guarantee.
EcoKure DTL provides the acceptance boundary around those systems.
| Capability | Enterprise function | Evidence produced |
|---|---|---|
| Probabilistic transition | Converts a request and candidate outputs into a canonical task, relevance state, lane and candidate set | Canonical task, input/evidence/policy/state hashes, disagreement state |
| DTL taxonomy lanes | Stores reusable, domain-specific verification knowledge with explicit trigger, patch boundary and promotion rules | Versioned lane pack, gate contract and promotion record |
| Deterministic specialist gates | Runs exact rules, calculations, schema checks, allowlists, attack corpora or domain verifiers | Structured verdict and reason codes |
| Claim control | Extracts claims from generated text and routes them to the gate that can verify them | Verified, refuted or unverifiable claim records |
| EvidencePack | Seals input, normalized state, result, code/data versions, limitations and reproduction instructions | Stable certificate and evidence-pack hashes |
| ReplayGate | Re-executes a sealed result and detects drift, missing inputs or unsafe replay | Replay match or drift verdict |
| Attestation | Signs certificate hashes and maintains a tamper-evident hash chain | Offline-verifiable signature, public key and chain position |
| Application safety guard | Applies fail-closed checks at sensitive sinks such as paths, SQL values, HTML, URLs, redirects and shell arguments | Safe, malicious or abstain decision |
| Registry and tenancy | Keeps verified lanes private by tenant unless explicitly bridged, with access control and provenance | Tenant-scoped route, ownership and bridge records |
| Enterprise metering | Exposes one partner API boundary, gate access controls, usage metering and subscription allowances | Usage records and plan/bundle state |
| Domain generalisation | Provides research and preview lanes for security, medical decision support, hardware, orbital, biological and other domains | Domain-specific test evidence with explicit maturity limits |
The central product is not a chat response. It is a governed, replayable and auditable state transition.
The public platform catalog separates established deterministic services from preview or research services. That distinction is part of the product's trust model.
| Gate or service | What it verifies | Public maturity boundary |
|---|---|---|
| SuperMath | Exact integer/rational, symbolic and calculus calculations with proof-oriented results | Deterministic math service; not a general scientific truth engine |
| UnitGate | Dimensional consistency of equations and quantities | Deterministic unit checks; domain interpretation remains external |
| ElementGate | Chemical formulae, molar mass and reaction balancing | Deterministic chemistry calculations; not laboratory validation |
| ClaimGate | Claim extraction and routing to an applicable verifier | Unverifiable claims remain unresolved |
| ClaimLint | Unsupported, over-claiming or unsafe wording detection | Wording control; not legal approval |
| EvidencePack | Hashes and seals a structured verification artifact | Preserves what was checked; does not prove reality by itself |
| ReplayGate | Replays a sealed artifact and detects divergence | Reproducibility check; external dependencies can still change |
| SecurityGate | OWASP-family vulnerability classification and guarded/fixed-mode checks | Software security evidence; not a complete security programme |
| MedGate | Rule-based clinical decision-support checks such as interactions, dosing and scores | Integration evaluation only; not a registered medical device or care decision |
| ChipGate | RTL structural safety patterns, passport and benchmark evidence | Early access/preview; not silicon readiness or physical safety |
| OrbitGate | Conjunction and collision-risk research lanes | Research preview; not flight software or certification |
| DiscoveryGate / BioGate | Protein-fold, binding-site and mutation research lanes | Research demonstration; not a biomedical conclusion |
| Research Taxonomy | Cross-domain lane packs and training interfaces | Demonstrates generalisation; each domain requires its own validation |
EcoKure DTL can sit between a company's AI applications and its operational systems, or run as a partner verification service.
flowchart LR
A[Models and AI applications] --> B[EcoKure DTL intake]
B --> C[Canonical task and relevance]
C --> D[Policy and role profile]
D --> E[DTL Registry and lane routing]
E --> F[Specialist deterministic gate]
F --> G[EvidencePack and ReplayGate]
G --> H[Attestation and audit export]
H --> I[Company workflow]
F --> J[Human review or abstention]
J --> I
Typical integration patterns are:
- API boundary: an application submits a candidate and receives a structured verdict, certificate and evidence reference.
- SDK/registry integration: an engineering team embeds lane routing, tenant access and verification into an internal platform.
- Customer-controlled verification: the customer runs the deterministic verifier and exports signed evidence to its own governance system.
- Enterprise control plane: workspaces, roles, model inventories, approvals, monitoring, connectors and signed audit exports surround the DTL runtime.
The EU AI Act is one governed EcoKure DTL application, not the definition of the whole platform. This pack maps selected obligations to technical evidence interfaces:
- operator role, intended purpose, jurisdiction and classification state;
- Article 4 AI-literacy evidence;
- Article 5 prohibited-practice screening;
- high-risk controls for risk management, data governance, documentation, traceability, transparency, human oversight and accuracy/robustness;
- Article 26 deployer controls and Article 27 assessment evidence;
- Article 50 transparency and output-marking states;
- GPAI provider documentation and systemic-risk evidence where applicable;
- Article 72 monitoring, drift and revalidation records; and
- Article 73 incident timelines and exportable evidence.
The verifier returns readiness evidence and deterministic control states. It does not decide the legal classification for the operator and does not issue a compliance verdict.
Read the company brief, the enterprise architecture and implementation guide and the official source register.
python eu-ai-act-dtl/eu_ai_act_dtl.py eu-ai-act-dtl/example.json
python -m unittest discover -s eu-ai-act-dtl -p "test_*.py"The example deliberately gives a candidate high confidence while omitting a
required transparency control. DTL blocks the candidate. Changing the
declared control to true produces an ALLOW for the corrected state.
The verifier's core contract is:
same task + same evidence + same policy/state = same accepted result
Confidence is retained as metadata. It never overrides missing evidence, failed controls, lane incompleteness or candidate disagreement.
EcoKure DTL can provide deterministic verification infrastructure, evidence preservation, replay, attestation and governance integration. It does not by itself provide:
- legal classification or an EU AI Act certificate;
- proof that a model is fair, safe, accurate or scientifically correct;
- a notified-body conformity assessment;
- a complete quality-management, privacy, cybersecurity or incident-response programme; or
- permission to deploy a domain-specific system without qualified review.
For the formal legal position, consult the Regulation (EU) 2024/1689 on EUR-Lex and qualified professionals.
Built from the JARVI3 / EcoKure DTL implementation. Public code is a reference surface; enterprise operations, private lane registries, tenancy, support, connectors and commercial terms are deployment-specific.