Evidence

Follow the claim.
Open the evidence.

Inspect EcoKure’s own recorded results, signed records and implementation boundaries. Keep integrity, scientific correctness and current evidence validity distinct.

Our results & run files Browse all 23 claims Check a signed record
Record check · exampleSigned EvidencePack
Intactsignature verified
Signature
✓Matches the key it was signed with
Contents
✓Unchanged since it was signed
What it proves
Nobody has edited the record.
What it does not
!That the result inside is scientifically right.
Run the verifier yourself. It works offline and needs nothing from us.
A signature check tells you a record is intact, not that it is correct. We say both.

Evidence / Two distinct questions

Is it intact? Is it still applicable?

Follow a signed record through a source change and see why integrity and current validity need separate checks.

The record still matches its fingerprint.

Select an illustrated object to inspect its role.

01 / 03

The record still matches its fingerprint.

Integrity checks detect tested changes to a record. An intact signature links the record to the signed content.

Published local integrity checks
Conditions & scope

Authenticity does not establish scientific correctness, complete dependencies or current authority.

Inspect the integrity result ↗
Architecture illustration · No operation performed
Start your review

Three ways in.

Browse the claims23 entries, each with its supporting record and its limits.Open the register → Inspect the programmesRead every result in the environment it came from, not out of context.See the programmes → Check a signed recordDownload the standalone verifier and check a record on your own machine.Download the verifier →
Gate library

What each route actually runs.

The route finder resolves your industry to a control pack. This is what sits behind that name: the path a decision takes, what it produced when it last ran, and what it replaces. Every figure is read from its run file, so nothing here can drift from the evidence behind it.

Showing all 6 routes.

Financial services

Did an AI-assisted decision follow the rule in force?

AMLGate · open source ↗

The path transaction → aggregation rule → allow / hold / report → record
When it ran

356 of 356 structuring days caught, 57% of legitimate cash-business days held

MEASURED · 26,521 synthetic customer-days
Against what you do today No speed gain to claim

No faster than the transaction monitoring a bank already runs, which detects this today. What is new is the record binding the decision to the rule version in force when an AI assistant is in the loop.

See it run → Scope this as a pilot →
Changing evidence or data

A source changed. What does that affect?

DELA

The path source revision → dependency graph → affected / preserved → selective replay
When it ran

1,695 of 100,000 records replayed, 98,305 preserved and proven untouched

MEASURED · selective_replay_benchmark_20260901.json
Against what you do today Measured gain

Replaying everything is the alternative, and it is what most teams do: 149.7 ms across all 100,000 records against 2.6 ms for the 1,695 that changed. 57x less work, and the untouched set is proven rather than assumed.

See it run → Scope this as a pilot →
Scientific compute

Is this job fit to run before we spend the compute?

FoldInputGate · open source ↗

The path job metadata → schema and routing check → valid / abstain / block → run
When it ran

5,248 of 5,248 faulty jobs rejected, 0 of 14,752 correct ones, 983 raw sequences blocked

MEASURED · foldinputgate_preflight.json
Against what you do today Measured gain

The alternative is finding the fault after the pipeline has run, which costs a queue slot and a resubmission. Worth reading against the AMLGate row: this rule holds none of the correct work, because a schema is a contract you can check. AMLGate holds most of the cash businesses it sees, because intent is not.

See it run → Scope this as a pilot →
Physical and grid operations

Is this instruction permitted by the limits we published?

GridGate

The path instruction → limit set → authority → allow / block / abstain → record
When it ran

4,947 of 4,947 breaches blocked, 0 of 14,059 valid instructions refused, 994 stopped by an expired limit set

MEASURED · gridgate_dispatch.json
Against what you do today Measured gain

The alternative is that a breaching instruction reaches the asset and is caught by protection, after the fact. The honest cost is not in the refusals: no valid instruction was refused, because a stated limit is a contract. It is in fail-closed. Every legitimate instruction whose limit set had expired was stopped, and in a network that is an operational cost rather than a safety win.

See it run → Scope this as a pilot →
Enterprise AI governance

Is this output allowed?

MathGate · open source ↗

The path AI output → route → verify → seal → allow / block / review / abstain
When it ran

1,089 of 1,116 pass, 17 refused, 10 blocked, 0 wrong answers

MEASURED · proofbench_x_results.json
Against what you do today Measured gain

A model asked the same questions answers all of them, including the ones it cannot establish. This refuses instead, and the refusals are the point rather than a shortfall.

See it run → Scope this as a pilot →
Independent verification

Can someone else check this without trusting us?

ReplayGate

The path sealed result → offline verifier → same verdict, no network
When it ran

produce 55.7 s, re-verify 62.2 ms

REPORTED · AntifragileAI x EcoKure DTL audit report, Aug 2026
Against what you do today Reported gain

Re-running the work is the alternative. Checking the certificate costs about 895x less than producing it did, which is what makes an auditor able to afford the check.

See it run → Scope this as a pilot →

Every run above is synthetic and local. They demonstrate the behaviour of a rule; they do not establish accuracy on your documents, your transactions or your sources — that is what a bounded pilot measures, against your own data and your own current process.

Where it runs

One cloud is measured. Other environments are at different stages.

The control layer is designed to run inside your own boundary, whichever cloud or private environment that is. AWS carries a target-environment measurement; the local runtime is built while customer and disconnected acceptance work remains. Every card keeps its own status and boundary.

Showing all 4 environments.

Amazon Web Services

MEASURED

A target-representative run in ap-southeast-2, torn down afterwards.

117,290
assurance verifications completed
0
errors recorded in that run
8.9 ms
p95 under soak, 128/s
0 lost
across reboot, chain verified after 52s

Signing through AWS KMS using ECC_NIST_P256, with the public key verified independently. Records in Terraform-managed RDS PostgreSQL, provisioned and destroyed by Terraform.

Target-environment evidence, not production customer validation, certification, or a Well-Architected review.

Open the source measurement →

Microsoft Azure

NOT BUILT YET

Hoping to bring you this one. Nothing has been run here.

Azure Key Vault for signing, Azure Database for PostgreSQL for records.

No measurement exists for this environment. The architecture is the same, but that is a design intention and not evidence, and until it has been run it should not be read as one.

Google Cloud

NOT BUILT YET

Hoping to bring you this one. Nothing has been run here.

Cloud KMS for signing, Cloud SQL for PostgreSQL for records.

No measurement exists for this environment. The architecture is the same, but that is a design intention and not evidence, and until it has been run it should not be read as one.

On-premises and air-gapped

RUNTIME BUILT - VALIDATION PENDING

EcoKure.App is runnable locally; no customer or disconnected acceptance run is recorded yet.

A customer-held signing key and a local database, with no outbound call at all.

The runtime package and offline verifier exist. This is not a customer operating reference, air-gapped accreditation or production validation.

Where the evidence stops

We publish the limits as well as the results.

  • Our own measurement is not independent review.Every run here was made by us. An outside reviewer is a separate step.
  • A demo is not production evidence.The runs are synthetic. A pilot measures your data against your current process.
  • Withdrawn claims stay visible.Corrections are listed, not deleted, so you can see what changed and why.

Check it yourself.

The full evidence room has every claim, run file and correction in one place.

Open the evidence roomAsk about a pilot