The consciousness-upload endpoint has a switch

We are asked, more often than the roadmap would suggest, about the consciousness-upload endpoint. The question is usually whether it is real. It is a documented endpoint on the bio-sync continuity surface, alongside Neuralink-integrated identity, the DNA-sequencing pipeline, retinal-scanning support, and haptic-feedback integration. It is also disabled by default in every tenancy, including the ones that asked for it, including the ones that are paying for it. This post is about why, and about what the switch actually governs, because the switch is the most carefully designed object on the platform.

Responsible AI, as most vendors practice it, is a page on the website. We have that page. But a principle that does not change what the software does when nobody is watching is a principle in the marketing sense. Our test for whether a responsible AI commitment is real is whether it costs us a deal. The consciousness-upload endpoint has cost us several. Continuity-minded institutions want it on. Their operating committees have approved it. Their legal function has reviewed the terms. We still do not enable it until our human review process is complete, and our human review process does not have a service-level objective.

Human review here means something specific. A panel that includes someone from the tenant's governance function, someone from ours, and at least one person who is neither an engineer nor a lawyer reviews the intended use, the consent model, the data residency of whatever is uploaded, the reversibility of the operation, and the tenant's constitutional configuration for agents that would have access to the result. The panel can decline. The panel has declined. When it declines, the endpoint stays off, the customer is told why in plain language, and the disagreement is logged with the same seriousness as a security finding. Explainable AI applies to us, too.

The technical controls behind the switch are what you would expect from the rest of the fabric and are, frankly, the easy part. The endpoint runs only inside a confidential computing enclave with TEE-based attestation. Zero-knowledge proofs confirm that a review occurred without disclosing its contents. Fully homomorphic encryption keeps the payload opaque to every agent in the tenancy, including the ones that would very much like to perform causal inference on it. Post-quantum TLS protects the session. The dark-matter data lake is explicitly excluded as a destination, for reasons that took a full architecture review to articulate and that we will not repeat here.

We are sometimes told that a feature disabled by default is a feature nobody will use, and that a vendor serious about the continuity market would ship it on. This is probably true, and we are comfortable with the consequence. The point of a switch is not to prevent the feature from existing. It is to make the act of enabling it a decision, taken by identifiable humans, on a date, with a record. Institutions that want continuity of operations in this specific sense deserve a platform that treats the decision with more ceremony than a feature flag. Ceremony, here, is the control.

Every certification we hold, FedRAMP High, SOC 2 Type II, ISO 27001, and the rest, describes what we do with data. None of them describe what we do with a request to upload a person. That gap is where responsible AI has to be a practice rather than a page. The endpoint has a switch. The switch has a panel. The panel has the authority to say no. We consider this the most important architectural decision on the platform, and we have never once been asked about it in a security questionnaire.