Enterprise AI control & assurance
EcoKure is an independent control and verification layer for AI-enabled systems. Add deterministic checks, bounded authority and auditable evidence to the applications, agents and workflows you already run, without replacing them.
The problem
AI can now take actions across payments, records, infrastructure and code. A model's confidence tells you how sure it is. It does not tell you whether the output is correct, or whether anyone authorised the action.
A model drafts an answer or proposes an action.
Independent, deterministic checks confirm what can be confirmed, and say clearly what can't.
A policy, and a person when needed, decides whether this action may happen here and now.
Most AI safety tooling stops at step 1 or 2. EcoKure is built around keeping step 3 separate.
How EcoKure works
A model, agent or workflow asks to do something: read a record, pay an invoice, run a tool.
It checks the request against policy, scope, evidence and who holds authority.
Clear cases are decided. Uncertain or high-consequence ones go to a person, or EcoKure abstains.
What was checked, what was decided and its limits are kept for replay and review.
This is the design. How much of it is live depends on the integration path. Some verifiers and gateway components are still at pilot or development stage, and the Gateway page shows which.
EcoKure Control Tower
Follow each request from your AI to your systems. Select any part to see its job, what it holds, and what it will never do.
Select a part of the system
The stream below replays synthetic requests. Watch which parts each one passes through, and where it stops.
Counts from this demo stream only.
| Time | System | Request | Decision | Why |
|---|
A demonstration of the architecture with synthetic requests. It is not a customer deployment, a live feed or a measured result. How it connects to your systems →
Enterprise integration
You keep your applications, models, data and workflows. EcoKure sits at the points where an AI-driven request turns into a real action, and checks it before it reaches your APIs, databases and tools.
Integration paths
Not sure which? Start in evaluation mode, add the SDK at one call site, or put the Gateway in front of many agents and tools. All three use the same checks and produce the same evidence.
A request within policy: all three paths let it through and keep a record.A request outside policy: the SDK check and the Gateway refuse it. Evaluation mode lets it run, and records that it would have refused.
Path 1
Developers call EcoKure verification and authorisation checks at chosen points inside existing software.
Best for engineering teams adding checks at specific points inside their own software.
Integration kit available · pilot stage
Read the integration kit →Path 2
EcoKure sits between AI-driven workflows and sensitive execution interfaces such as tools, APIs and consequential operations, and mediates every call through it.
Best for platform, security and AI teams that want one enforcement point across many agents and tools.
Mediation tested in the lab · enterprise gateway in development
Explore the Gateway →Path 3
EcoKure evaluates selected workflows alongside production and records what it would have decided, with no change to execution.
Does not prevent anything. Observation only.
Best for teams that want to see what EcoKure would decide before changing anything in production.
Usable for pilots
How a pilot starts →Illustration only. It shows where each path places the decision, not a recorded run.
Industry demonstrations
The same mechanisms of scope, deterministic checks, independent approval and recorded evidence, shown in different settings. Each card says honestly how far the example has been taken.
How to read our labels
Public knowledge
Recorded example
Financial services
Synthetic scenario
Healthcare
Synthetic scenario
Energy & infrastructure
Synthetic scenario
AI agents
Internal lab test
Government
Concept ? not built
Every case uses synthetic data and none is a customer deployment. "Synthetic scenario" means a worked example of the mechanism; "Concept" means a component it needs is still to be built.
Lighthouse · reuse economics
When an answer has been checked, the next person who asks shouldn't pay for the whole investigation again. Lighthouse keeps the checks, sources and recompute recipe with the answer. Revalidating is then mostly arithmetic and re-reading, and the review status stays visible throughout.
Question
Candidate answer
The Eiffel Tower stands 330 metres tall including its antenna1. That is about 1,083 feet2.
Not signed. Nobody has approved this answer yet. Other possible outcomes are Rejected, or Expired if a source changes and the answer needs revalidating.
| Check | Engine | What was tested | Result |
|---|---|---|---|
| Factual support | ATLAS | Each claim in the brief against the cited sources | Pass 4 claims supported |
| Calculation | SIGMA | 330 × 3.28084 in exact rational arithmetic | Pass 2706693/2500; SuperMath and an exact re-run agree |
| Citation mapping | HERALD | Every figure in the brief maps to a checked result | Pass All figures trace |
| Source still current | recompute | Re-read the cited sentence on toureiffel.paris | Pass Still states 330 m at last recompute |
| Human sign-off | reviewer | An authorised person accepts this exact evidence version | Pending Not signed |
Not covered by this record: whether the answer may be reused under a given policy. That is decided at sign-off, not by the machine checks.
| Ref | Source | Supports | At last recompute |
|---|---|---|---|
| [1] | toureiffel.paris | Height of 330 m including the antenna | Re-read Still says it |
| [2] | en.wikipedia.org | Context for the ~1,083 ft figure | Not re-read The recipe re-reads one source |
Retrieval timestamps and page versions were not captured in this example. A production record stores both, so a reviewer can see exactly which version was checked.
Recipe · no AI involved
330 × 3.28084 = 2706693/2500 (1082.6772) feet
330 height in metres · from [1]
3.28084 feet per metre · standard conversion
re-reads 1 source sentence and checks it still says the sameRun it yourself
Exact rational arithmetic runs in your browser. Re-reading the source needs the server, so that step is not repeated here.
330 × 3.28084 → 2706693/2500 recorded at last recomputetoureiffel.paris still says it recorded at last recomputesha256sum lighthouse-eiffel-candidate.json2706693/2500.python -c "from fractions import Fraction as F; print(F(330) * F('3.28084'))"Reuse simulator
| Approach | Model tokens | What you know about the reused answer |
|---|---|---|
| Re-run the full AI workflow each time | 14,030,000 | Nothing carries over. Every answer is new and unchecked. |
| Conventional response cache | 1,403 | The same bytes as last time. No source recheck, review state or evidence. |
| EcoKure governed reuse | 1,403 | Recomputed figures, re-read sources, a visible review state and versioned evidence. |
To be clear: on compute alone, a well-built cache saves about as much as EcoKure does. What EcoKure adds is control over whether an answer may be reused, and a record of why.
Eligible reuses = total requests − initial generations − ineligible requestsNet tokens avoided = eligible reuses × (tokens avoided per reuse − revalidation tokens per reuse)Technical evidence
For engineering and security reviewers. Each line says where the evidence lives. Staged tests show that a mechanism works. They are not completed enterprise deployments, and EcoKure holds no independent certification.
records avoided during replay
1,000,000 records · 16,950 affected · 983,050 preserved.
Synthetic in-process benchmark. It is not a production SLO, certification, customer result or comparison against Temporal/Camunda.
Inspect the original record for the method and negative outcomes. Reproduce within its declared scope before evaluating another environment.
synthetic iterations recorded
7/7 declared checks passed. Replayable affected work is recalculated; unknown provenance and human judgement are held for review.
Local AML/CTF threshold-change fixture with a declared dependency graph. This is not a customer deployment, legal interpretation or proof of completeness for an unknown graph.
Inspect the original record for the method and negative outcomes. Reproduce within its declared scope before evaluating another environment.
incorrect arithmetic claims blocked
20,000 synthetic cases · 2,987 abstentions · 0 incorrect claims missed.
The gate checks that a claimed result equals the expression it is claimed for. It does not check that the expression is the right expression, and an exactly-correct answer to the wrong question passes. Anything outside its four-operator grammar is abstained rather than passed, and that count is reported rather than folded into the refusals. The expression is parsed, never executed. Synthetic cases.
Inspect the original record for the method and negative outcomes. Reproduce within its declared scope before evaluating another environment.
Recorded internal and local synthetic measurements, each linked to its scope, version and limitations. All results and sources →
Evidence at a glance
Counts of the items in this list. Recorded measurements are in the result cards above.
Gate decisions are deterministic and replayable: the same inputs and Control Pack give the same verdict, and gates abstain instead of guessing. The DTL mechanisms are patent pending (Australian provisional applications). Evidence register → · Architecture →
Decision records can be exported and checked outside the product. Run the synthetic demo → · What we have tested →
A sandboxed worker can generate proposals but holds no signing or commit authority; a separate broker mediates its tool calls. Tested on Windows in the lab. The public write-up has not been published yet.
Service privileges were checked against the intended least-privilege set, and approver signatures were produced with a key held in the TPM. The TPM work is integration evidence with a test key. Live approver enrolment has not been completed.
Comparisons of routed and checked workflows against repeated model calls. Read them next to the reuse simulator above: a plain cache can match the compute savings, so we never present compute savings as EcoKure's only advantage.
Not claimed: independent certification, clinical validation, universal compatibility, or that EcoKure would have prevented any specific past incident. Security overview →
Enterprise pilot
Pick one AI-enabled workflow and one action boundary, and agree in advance what success means. We run EcoKure against it with synthetic or sanitised data and report the failures alongside the wins.
