Averon Private  //  Verifiable private inference private.averon.systems Status: pre-launch  ·  waitlist open

An Averon Systems service

Frontier AI that we cannot read. And you can prove it.

Averon Private runs frontier models inside hardware-attested enclaves. Before it sends anything, your client checks which image is running and stops if that image is not the one we published. The guarantee is not a policy commitment. It is a property of a build you can reproduce and check yourself.

01 The problem

Two options today, and both are bad.

People working under privilege, clinical confidentiality, fiduciary duty, or source protection are asked to choose between capability and control. Many of them choose neither.

Option one

Hosted frontier models

Readable

Capable, and readable by the operator. Prompts arrive as plaintext on someone else's servers, pass through abuse detection and logging, sit in retention windows measured in weeks, and stay reachable by staff with production access, by legal process, or through a breach. The commitments are real. They are also revocable, and impossible to check from outside.

  • Retained for abuse review by default, in most cases
  • Reachable by subpoena, warrant, or civil discovery
  • Enforced by policy and access control, not architecture
  • No way to confirm any of it independently
Option two

Local models on your own hardware

Capability-limited

Private, and much weaker. A quantised model on a workstation sits a generation or more behind the frontier, and the gap is widest on long documents and sustained reasoning, which is most of the work this would be used for.

  • Well behind frontier capability, and staying behind
  • Context limits that break on a full case file or chart history
  • Hardware you buy, maintain, secure, and replace
  • Trust that stops at your own endpoint security

So most people take a third option and don't use it for the sensitive work. The litigation file, the patient history, the client's books, the source's identity. The material where the help would matter most is the material that never gets pasted into a prompt.

02 How this is different

Don't trust us. Check us.

Every claim on this page is meant to be falsifiable by someone outside Averon. If one stops being true, a third party should be able to show that without our cooperation.

Reproducible enclave builds

01

The enclave image is built deterministically from public source. Clone the repository, run the build, and compare your hash against the measurement your client reported during attestation. If they differ, something we did not publish is running, and the client closes the connection.

Bit-identical rebuilds · Public source

Open-source client-side attestation

02

Verification runs on your machine. The client validates the attestation report, walks the certificate chain to the silicon vendor, compares the measurement against the published reference, and binds the session key to it. Apache-2.0, so you can audit it or run a build compiled by someone with no connection to us.

TODO: repository link · Apache-2.0

Published third-party audits

03

Independent reports published in full: scope, method, and findings, including the findings still open at publication, with dates and remediation status. We publish the reports themselves, not summaries of them.

TODO: first report and auditor name

Standing security bounty

04

A permanent programme covering the enclave image, the attestation client, and the key-handling path. The highest tier is reserved for any demonstrated operator route to plaintext. Researchers may disclose 90 days after reporting, with or without our agreement.

TODO: confirm ceiling and scope document

How a session is established

  1. 01 Attest before connecting Your client requests a signed attestation report from the enclave hardware and validates it against the silicon vendor's certificate chain. Nothing is transmitted until this passes.
  2. 02 Compare against a public reference The reported measurement is checked against the published build hash, which is reproducible from source. Independent rebuild attestations are listed next to it.
  3. 03 Bind the key to the measurement The session key is sealed to the measured image. A different image cannot derive it, including a later image signed by us.
  4. 04 Terminate TLS inside the boundary Ciphertext crosses the host and is decrypted inside the enclave. No proxy or logging tier holds your text in the clear on the way in.

03 What we're honest about

What the guarantee covers, and what it doesn't.

The claim below is narrow on purpose. A broader one would not hold up.

The guarantee

Your conversation is encrypted in transit into the enclave and is not decrypted anywhere else. TLS terminates inside the attested boundary. Sealed history, if you enable it, is encrypted to a key that only exists inside that boundary. There is no operator shell into a running enclave, no hypervisor-level memory introspection, and no logging of prompts or responses. These are properties of a published image whose hash your client checked before the session started, not policy statements.

The boundary

Your conversation exists as plaintext in enclave memory while the model processes it. Inference requires that. What we can offer is that no operator, hypervisor, or host process has a path to that memory, and that you can check for the absence of such a path before you start.

Three further limits

  • This inherits the silicon vendor's threat model. A confidential computing root of trust means trusting the hardware vendor and their attestation chain. That is a narrower trust base than trusting an operator's internal controls, but it is not zero.
  • Side-channel research is active. Attacks against enclave technologies get published and mitigated, and new ones keep appearing. We track them in the open and state which mitigations each build carries.
  • Compulsion produces detection, not immunity. If we were compelled to ship an image with different properties, it would carry a different measurement and your client would refuse it. You would find out. That is worth having, but it is detection, not prevention.
Also true

Metadata still exists outside the enclave. We know that an account connected, roughly when, and how much compute it used, because billing requires it. We do not know what was said and hold no key that would tell us, but that is not the same as knowing nothing about you.

04 What we store

Three zones, with a hard line between them.

Some data has to live outside the enclave for the service to run. Here is which data, and where.

Zone 01 · Outside the enclave

Minimal identity and billing

An email address, a payment reference held by our processor, and coarse usage counters for billing: request counts and token totals, not content. This is ordinary business data under ordinary protection.

Readable by us · Kept minimal
Zone 02 · Inside the enclave

Sealed history, opt-in

Off unless you turn it on. When enabled, history is encrypted to a key sealed to the attested image and stored as ciphertext we can replicate, back up, and delete, but not open. If you lose your credential, the history is gone. Any recovery path we controlled would also be an access path.

Sealed · Unreadable by us
Zone 03 · Nowhere

Incognito sessions

Nothing is written. No history record, no generated title, no derived summary, no cache entry that outlives the session. The only persistent trace is the billing counter.

No persistence

What we don't collect

  • IP address logs. Connections terminate and addresses are not retained past the request.
  • Location history, whether stated, derived, or inferred.
  • Telemetry or analytics over prompt and response content.
  • Device or browser fingerprints.
  • Third-party trackers, ad pixels, or session replay. This page loads nothing from another domain.
  • Training data taken from your conversations, on any tier.

On deletion

Deleting sealed history removes the ciphertext, and we can show the object is gone. We cannot make claims about content we were never able to read, and neither can anyone else. If you need a stronger deletion story than that, incognito is it, because nothing was written.

Retention schedule for the billing zone: TODO, published alongside the technical brief.

05 Pricing

Tiers and what they cost.

Confidential computing hardware costs more per token than commodity inference. We would rather charge for that than take a second revenue stream out of your conversations. All tiers are pre-launch.

Free

Coming soon

$0

Open-weights, 32B class
  • Hard monthly cap, no overage, no card
  • Same enclave and attestation path
  • Incognito sessions included
  • Sealed history not included
Join the waitlist

Professional

Coming soon

$79 / month

Frontier class, standard context
  • Allowance sized for daily professional use
  • Opt-in sealed conversation history
  • Long-document work at full context
  • Priority under capacity contention
Join the waitlist

Frontier

Coming soon

$300 to $500 / month

Frontier class, extended reasoning
  • Highest reasoning and context budget
  • Sized for full case files and record sets
  • Higher sustained throughput allowance
  • Direct line to the engineering team
Join the waitlist

Enterprise

Coming soon

Contact

Dedicated attested capacity
  • Tenant-dedicated enclave capacity
  • Named measurement pinned per tenant
  • Audit support, DPAs, security review
  • Firm and practice-wide deployment
Register interest

06 Waitlist

Join the waitlist.

One email when the private beta opens, and one more if the technical brief changes in a way that affects the guarantee.

The waitlist sits outside the enclave. It is an email address in an ordinary database.