What Apex commits to.
This is the public charter — the commitments Apex holds to itself and to the customers, researchers, and regulators who interact with us. It is written so we can be held to it. When the commitments are harder to keep than we claimed, we update the charter rather than the messaging.
Why we exist
The physical world runs on physics, chemistry, biology and silicon. The next decade's hardest problems — climate, drug discovery, energy, semiconductors — share that substrate. A foundation model that understands one understands the rest. We are training that model.
What kind of company
A research and product company. The research is the product. We commercialise Aether by piloting one workload at a time — not by selling the platform abstractly. Customer pilots fund the research; the research raises the ceiling on the pilots.
How we think about AGI
AGI is what you get when a model can do the work — not just narrate it. We are building it through the disciplines that have to face physical reality, because reality is the only regulariser that holds. The model that designs a rocket today is the model that runs a lab tomorrow.
- 01
We will train one model
We are training one foundation model for the physical world. Not many specialised models with shared marketing. The decision is irreversible at the architectural level and we will be explicit when we revisit it.
- 02
We will publish what we measure
Every claim ships with a reproducible benchmark. If we cannot publish the reproduction kit, we do not make the claim. Numbers and failure modes are treated symmetrically.
- 03
Safety is engineered, not promised
Refusals are encoded in the model layer. Capability gating sits between intent and action. Audit trails are immutable. We red-team quarterly and we publish the methodology.
- 04
Customer data is not training data
Base models are never trained on customer prompts, code or experiment data. Fine-tunes are scoped and never reused across tenants. Right-to-delete is enforced by the runtime, not by a post-hoc job.
- 05
Independence preserved
Labs programmes report through a separate governance line from the commercial business. Customer relationships do not influence safety findings. We will publish the operating principles of this separation when they evolve.
- 06
Mistakes get postmortems
Customer-impacting incidents are disclosed on /status within hours and a postmortem follows within ten business days. We publish the postmortems with named accountability for the corrective action.
- 07
Researchers are not adversaries
Good-faith vulnerability research is protected under our safe-harbour policy. We pay for findings. We acknowledge contributing researchers publicly. We do not legally pursue researchers who respect our published scope.
- 08
We will be present in the disciplines we serve
Forward-deployed engineers sit with customers. We do not run pilots from a sales office. The customer's hardest cases are inputs to our training-data prioritisation, not problems we tell them to work around.
The lines we will not cross.
A charter that only says yes is incomplete. These are the lines we have drawn — public, named, and enforceable by anyone who cares to check.
Wrap another lab's model
Aether is one foundation model we train ourselves. We do not call third-party chat APIs as part of the product. We will be explicit if that ever changes.
Sell intelligence as a service
We don't sell a generic AI subscription. We sell pilots scoped to specific workloads with written win conditions. Every commercial engagement is anchored to a measurable outcome.
Train on customer data without consent
Base models are never trained on customer prompts, code or experiment data. Fine-tunes are scoped, opt-in, and never reused across tenants.
Build dual-use shortcuts
We do not ship features that lower the bar for biosecurity or weapons-development misuse. The refusal corpus is versioned alongside the model. Red-team coverage on those classes is non-negotiable.
Hide the failure modes
If a published benchmark stops reproducing, we flag it. If a customer pilot misses its win condition, the postmortem is published. If a commitment in this charter is harder to keep than we claimed, we update the charter — not the messaging.
“We are training the AI that does the work, not just the chat about it. AGI through the disciplines that have to face physical reality — not through more pages of scraped text.”
Public, audited, reproducible.
Commitments without enforcement are just rhetoric. The four mechanisms below are how an outside party can hold us to what we say.
Public benchmarks
Every customer-facing claim must reproduce from a published kit. If a number on our site stops reproducing, we treat it as a P1 bug.
External red-team
We commission external red-team passes quarterly. Findings are published; awards are paid. The corpus version is pinned per model release.
Annual audit
SOC 2 Type II annually. ISO 27001 and ISO 42001 in progress. Audit reports available under NDA on the trust center.
Open challenges
Our failure modes are published alongside our wins. If we discover that a published commitment is harder to keep than we claimed, we update the charter rather than the messaging.
Who owns what.
Commitments need owners. Below is how this charter is governed in practice — who changes it, who reviews it, and how to flag a violation without going through the people who might have caused one.
Charter ownership
This charter is owned by the founder & CEO and reviewed quarterly. Material changes are flagged in the changelog. We publish a diff when wording shifts in ways that matter.
Safety advisory board
An external safety advisory board reviews biosecurity, dual-use and capability-gating practice once per quarter. Membership is published; tenure is two years; vote on refusal corpus expansions is binding.
Independent labs governance
The Labs programme reports through a separate governance line from the commercial business. Customer relationships do not influence safety findings. Labs leadership has authority to delay a release.
Whistleblowing channel
Anyone — employee, contractor, customer or external researcher — can write to trust@apexworldlabs.com. Reports are acknowledged within one business day. Retaliation against good-faith reporting is a terminable offence.
What we have not yet figured out.
A charter that is silent on its open questions is a marketing document. Below are four questions we are working on. We will publish updates as we learn.
- 01How much of the safety architecture can we open-source without compromising it? We are working on this; current answer is partial.
- 02When will sovereign deployments be possible without an Apex forward-deployed engineer on-site? The current answer is they require one; the goal is to make them stand on their own.
- 03How should the Labs programme be funded long-term? Operating revenue is fine at the current scale; we expect to need a different model at 5×.
- 04Where do we draw the line on dual-use research? We have written guidelines; the line will move as the science moves, and we will publish the changes.
Hold us to it.
If a commitment in this charter is being violated, write to trust@apexworldlabs.com. We acknowledge within one business day.