Aplark™ builds mission-critical intelligence infrastructure for defence, aerospace, space and autonomous systems.
Our first product, Aplark Assure, systematically evaluates AI across operating conditions, identifies evidence gaps and creates traceable evidence for deployment decisions.
The problem
Mission-critical AI is evaluated on benchmark accuracy. Accuracy is not readiness. A held-out test score describes how a model performed on a sample somebody chose; it does not describe the operational domain, it does not account for the correlation between the samples inside it, and it says nothing at all about the conditions that actually drive risk — degraded sensing, adversarial geometry, the tail of the latency distribution.
These systems operate under uncertainty, in degraded conditions, against adversaries who are choosing the inputs. The question a certification authority asks is not how well the model scored. It is what residual failure rate you can demonstrate, at what confidence, over which declared operational domain, with what evidence.
The fact nobody sells against
Testing alone cannot establish extremely low failure probabilities for learned components when the required evidence exceeds practical test budgets.
What can be done is to state the shortfall honestly, quantify it in orders of magnitude, and move the argument to where it can actually be carried: the architectural mitigation chain, or a restricted operational domain. Aplark Assure is built to say that out loud, on day one, before the GPU time is spent.
Read the argument, with citationsAvailable now
AI assurance and evaluation
Know how your AI fails before it matters.
Aplark Assure generates, preserves, evaluates, traces and independently verifies the evidence that demonstrates a safety- or mission-critical AI system satisfies its defined requirements. It is the only Aplark product available today.
It does not certify models. It does not score models. It does not make models safer. It makes the evidence about them auditable — and it tells you, before you spend the test budget, when the thing you are trying to demonstrate cannot be demonstrated at all.
How Assure works
Nine stages, from a declared operating domain to a signed deployment decision.
Programme, item, ODD, hazards, requirements, mitigation chain and exposure weights. Produces a residual rate per hazard.
Campaign scoped to target cells, executed deterministically at a declared reproducibility class.
Artefacts hashed, environment fingerprinted, records appended to the signed ledger.
Inclusion and consistency proofs, on-demand replay, and offline verification of the exported bundle.
Sufficiency computed on effective samples. Six-dimension readiness at a named criticality level.
What is missing, what it blocks, and what it costs to close. This is the home screen.
GSN argument, explicit defeater register, evidence bound to claims.
A deterministic, provenance-linted certification package. Two generations, identical digest.
Policy evaluated against readiness and edge conformance for a specific artefact digest, offline, producing a signed decision.
The evidence
The assurance overview and the gap report, rendered exactly as they appear in the product — not illustrations.
EO/IR Detection System · Version 4.2
ODD Coverage
82%
weighted; 1,204 of 1,468 reachable cells sufficient
Partial
Requirement Satisfaction
90%
38 of 42 verified at admissible reproducibility class
Partial
Robustness Margin
+2.1
dB SNR above declared ODD boundary, worst family
Met
Edge Conformance
MET
latency, throughput, quantisation delta, hash verified
Met
Provenance Integrity
100%
all cited artefacts signed, R0/R1
Met
Residual Risk Posture
REVIEW
1 Hazardous open · 0 Catastrophic open
Review
Assessed at Hazardous. The same evidence is insufficient at Catastrophic.
You have 40,000 frames from 22 sorties. For this cell that is 22 effective samples. (ρ = 0.9 — consecutive video frames)
Hazardous demonstrability at residual rate 2.1×10⁻⁵ requires 3,000 effective samples.
Additional independent sorties: ~2,978 · est. 6.2 GPU-hours ± 1.4 · 340 GB
This gap cannot be closed by additional testing.
Actions: strengthen the architectural mitigation chain, or restrict the declared ODD.
This gap has no testing-based action. See above.
The platform
The Aplark platform is organised around a single principle: the layer that records evidence must never be the layer that produces the thing being evidenced.
Intelligence Core
Aplark Assure is available today. Every other node on this diagram is a roadmap direction — each links to a page that says so plainly.
The name
Six capabilities an intelligent system needs before anyone should trust it with a decision.
The future platform
Aplark Assure is the only product shipping today. Everything below is direction, not availability — none of it is purchasable, and none of it is described here as though it already works. Registering interest puts you on the briefing list for the product concerned.
Aplark Sim
Simulation environments
Test intelligence in worlds you can control.
IN DEVELOPMENTRegister InterestAplark Forge
Synthetic data
Build the data your AI needs.
IN DEVELOPMENTRegister InterestAplark Red
AI red teaming
Find the weaknesses before the field does.
IN DEVELOPMENTRegister InterestAplark Edge
Edge AI
Run intelligence where the mission happens.
IN DEVELOPMENTRegister InterestAplark Monitor
AI observability
Know when deployed intelligence drifts.
IN DEVELOPMENTRegister InterestAplark Twin
Digital twins
Model the system before changing the system.
IN DEVELOPMENTRegister InterestAplark Autonomy
Autonomous systems
From assisted to supervised autonomy.
IN DEVELOPMENTRegister InterestSectors
Systems that must work when the operating picture is degraded on purpose.
Where the certification argument is the deliverable, not the documentation.
No maintenance window, no rollback, and a light-delay between the fault and the operator.
Shared space with people, and a failure mode measured in physical contact.
Built for sensitive environments
Aplark Assure is designed for machines that have never had a route to the internet, because that is every machine it will ever run on.
No telemetry
Nothing about your programme is reported back to us. Ever.
No licence callback
The installation does not phone home to validate itself.
No CDN
Every asset is bundled locally. No web font, no icon set, no script.
No outbound connection
Under any configuration. There is no flag that turns one on.
This site holds itself to the same standard: it makes no third-party browser request of any kind, and a test in our build pipeline fails if one ever appears.
A working session with your programme, your hazards and your evidence — not a slide deck.
Request a Demo