Roadmap
Where Vraelis stands.
Every product, what works today, and the source for every line.
Products
At a glance
As of 26 October 2026
| Product | Status | Where it stands |
|---|---|---|
| Vraelis Prompto | Available | Version 0.1.2, released October 21, 2026. Open source (Apache-2.0), alpha, with downloads for Windows, macOS and Linux. Testing so far is self-run; no outside user has run it yet. |
| Vraelis Contour | Private development | A private reference experiment, not a product release. Line by line below. |
| One record for every model | Next | Planned: a signed record of which model was approved, for which machine, by whom. Not available. |
| The registry for physical AI | Direction | Where the company is going. A plan, not a product. |
Vraelis Contour
Line by line, with sources
As of 17 October 2026
| What | Status | Source |
|---|---|---|
| Working in the experiment | ||
| Signed release and approval bindings | Working in the experiment | Founder brief, 10 October 2026Internal automated testing, runtime-policy revision |
| Runtime policy enforcement: ONNX Runtime 1.23.2, CPUExecutionProvider only | Working in the experiment | Founder brief, 10 October 2026Internal automated testing, runtime-policy revision |
| Real CPU inference on the approved bytes | Working in the experiment | Founder brief, 10 October 2026Internal automated testing, runtime-policy revision |
| Desired against reported loaded observations, in baseline and reconcile modes | Working in the experiment | Founder brief, 10 October 2026Internal automated testing, runtime-policy revision |
| A read-only evidence app over the recorded trace | Working in the experiment | Founder brief, 10 October 2026Internal automated review of the evidence app |
| Measured | ||
| 133 of 133 experiment tests at the runtime rejection-diagnostics revision, in the pinned network-disabled harness, reproduced by internal automated testing on a separate build on October 10, 2026 (Pacific) | Measured | Internal automated testing, October 10, 2026 (runtime rejection-diagnostics revision) |
| Earlier: 131 of 131 experiment tests at the earlier runtime-policy revision, reproduced by internal automated testing | Measured | Internal automated testing, runtime-policy revision |
| Planned next | ||
| Private collector prototype; host-checked | Live-cluster validation of the current revision pending | Private engineering record, October 17, 2026Specification v1.3 |
| Heartbeat freshness for runtime reports | Not done yet | Founder brief, 10 October 2026 |
| Live fault runs | Not done yet | Founder brief, 10 October 2026 |
| A latency study | Not done yet | Founder brief, 10 October 2026 |
| Adapters for customer runtimes | Not done yet | Founder brief, 10 October 2026 |
| Not claimed | ||
| Hardware attestation | Not claimed | Founder brief, 10 October 2026 |
| Customer deployments or adoption | Not claimed | Founder brief, 10 October 2026 |
| Certification or regulatory compliance | Not claimed | Founder brief, 10 October 2026 |
| Pricing or general availability | Not claimed | Founder brief, 10 October 2026 |
Sources
- Founder brief, 10 October 2026
- What is implemented, unfinished and never claimed.
- Internal automated testing, runtime-policy revision
- The 131 tests of the earlier runtime-policy revision, reproduced offline on a separate pinned Docker build. Runtime policy enforcement checked.
- Internal automated testing, October 10, 2026 (runtime rejection-diagnostics revision)
- 133 of 133 at source revision contour-runtime-status-rejection-v1, run with networking disabled (NetworkMode none before and after, exit 0). The retained test log was checked against its SHA-256 (441d44b4…08c5).
- Internal automated review of the evidence app
- Evidence app review: five findings, each addressed in the current version.
- Specification v1.3
- Internal proposal that describes the Kubernetes collector. Not frozen.
- Private engineering record, October 17, 2026
- Host checks of the private Kubernetes collector prototype. Live-cluster validation of the current revision has not been run. The record is not public.
Recorded toolchain
What the experiment runs on
| Component | Version | How it is bound |
|---|---|---|
| ONNX Runtime | 1.23.2 | Signed runtime policy, also pinned by the harness |
| Execution provider | CPUExecutionProvider | Signed runtime policy, this provider only |
| Node.js | 24.19.0 | Harness pin, not in the signed policy |
| Python | 3.12.14 | Recorded and pinned by the harness, not in the signed policy |
Limits
Limits of the evidence
Self-reports
The worker reports what it loaded. Host and worker are trusted.
Structure, not signatures
Opening a trace checks structure only, so an edited, self-consistent trace passes. It is labelled imported.
Synthetic models
No production or customer model has been through it.
One run, in a lab
One recorded run, in a lab that controls no equipment.
Tests
What the tests do not show
133 of 133 is the experiment's runtime rejection-diagnostics revision, run in a pinned, network-disabled harness and reproduced by internal automated testing on October 10, 2026 (Pacific). It supersedes the earlier 131 of 131 at the runtime-policy revision; it is not 133 additional tests. The evidence app's own tests are a separate suite, and no counts are added together.
- Production readiness, a formal product milestone or independent human review
- External signature verification: receipt integrity is covered by signature tests inside the suite
- Heartbeat freshness: the policy's age limit bounds how old an approval may be, not how fresh a report is
- Customer need or demand
- Behaviour with real models, other runtimes or other hardware
- Behaviour under live faults, or speed
- What a machine runs, beyond its worker's report