Decision Lab · current proof and applied research

Where Physical AI product decisions become inspectable.

Run current decision tools, inspect Auralink and Reachy evidence, and see how product, architecture, model, safety and field assumptions are turned into decision records. The planned physical Foundry remains clearly labelled further down.

Run what exists today Bring one decision

Current Decision Lab index — evidence remains classified by maturity, source and limitation.

Digital lab live · physical Foundry planned

Runnable demonstrations, measured bench work and Hyperion-owned R&D are available now. The proposed physical facility, zones and missions remain illustrative until built, tested and labelled otherwise.

Available now

Start with what exists.

These are current digital, bench and internal-R&D artefacts—not the planned physical facility. Each page states what it can and cannot prove.

Auralink

Pre-production · Hyperion-owned R&D. An inspectable reference implementation—not a client deployment or commercial-outcome claim.

Auralink is Hyperion-owned pre-production R&D awaiting hardware integration.
78% autonomous incident resolution in a controlled-environment evaluation reported in arXiv preprint 2603.08736.

Inspect Auralink

Reachy Mini + SO-101

Measured · simulated · dry-run. Real sensing and teleoperation evidence stays separate from simulated autonomy and unexecuted robot actions.

Inspect the robotics evidence

Measured on the bench

AMD Ryzen AI Max+ 395 (Strix Halo) · 128 GB LPDDR5X-8000 unified memory · medians of three runs after a discarded warm-up · no tuning

Read the complete field notes

06 · Available now

Start with what exists.

These are current digital, bench and internal-R&D artefacts—not the planned physical facility. Each page states what it can and cannot prove.

Run the demonstration

The proposition

Not a showroom of tricks. A place to make the next decision.

A model demonstration asks whether one behaviour can work under chosen conditions. The Decision Foundry asks whether the complete product should advance. Customer value, intelligence, hardware, software, safety, operations and economics are evaluated together. A convincing output is not a pass: the evidence determines whether the decision is go, redirect, sequence or stop.

One connected system

One lifecycle. One runtime. One evidence return.

The Product System governs why and when the product advances. The Physical AI system architecture governs how the system behaves in operation.

Product System

Strategy → Discovery → Productization → Launch → Scale

Five gates govern when the product may advance. Six decision lenses test whether the evidence meets the threshold for the current decision.

Physical AI runtime

Sense → Connect → Compute → Reason → Act → Orchestrate

Trust, safety and cybersecurity surround every layer.

Evidence return

Incidents, field variance, technical discovery or economics can reopen an earlier decision. Nothing advances automatically.

The visitor mission

A signal becomes a decision—or stops at the boundary.

01Sense the real decisionOutput: decision contract

Proposed lab zones

01Mission Lock
02Field Cell
03Signal Spine
04Sovereign Core
05Knowledge Forge
06Failure Theatre
07Authority Gate
08Fleet Observatory
09Evidence Chamber
10Executive Studio
11Field Notes Studio

Follow one bounded initiative from declared intent to an inspectable decision record. As you move, the architecture resolves around what must be observed, authorized and proved.

01LOCK
Sense the real decision
Name the intended use, accountable owner, system boundary, acceptance threshold and stop condition before a model is selected.
Output: decision contract
02GROUND
Ground and decide
Observe the declared baseline, connect field signals to governed knowledge and compare options with sources, uncertainty and operating context visible.
Output: bounded proposal
03GATE
Authorize or abstain
Schema, allow-list, policy and authority checks determine whether a proposal may reach deterministic control. Ambiguity or unsafe intent stops here.
Output: authorization record
04RECOVER
Act, observe and recover
Run one bounded action, inject declared variance, watch operator and system response, and prove that degradation, stop and rollback paths work.
Output: observed system behaviour
05PROVE
Prove and scale
Compare the result with acceptance and kill thresholds. Record the evidence, open assumptions and a go, redirect, sequence or stop recommendation.
Output: Product System gate record

Illustrative architecture

Stress the authority boundary

Change one condition. The model may interpret and propose; the system boundary determines what happens next.

AuthorizedInputs remain inside the declared envelope. A typed proposal passes policy and human authority, reaches deterministic control and returns evidence.

Mistral-first. System-bounded.

Intelligence may propose. Authority must be designed.

Hyperion selects Mistral first for language, multimodal understanding, speech, retrieval and model adaptation, then tests the choice against the task, deployment boundary, licence, region, latency, cost and evaluation set. The least-complex sufficient answer may still be rules, search, a smaller open-weight model, another model or no LLM.

Interpret
Available capability
Selected Ministral and Voxtral models can interpret bounded image, language, audio and operator input at the edge or through managed endpoints.
Ground
Available capability
OCR, embeddings, retrieval and Search Toolkit connect documents, incidents and technical knowledge to inspectable sources. OCR remains document understanding—not a safety-critical decision-maker.
Propose
Available capability
Structured generation compares options, exposes uncertainty and requests allow-listed tools. Function execution remains the developer’s responsibility.
Orchestrate
Public preview
MCP Connectors and durable Workflows can coordinate tools, data and human checkpoints. Preview interfaces require evaluation and change control before production use.
Adapt
Evidence-gated
Prompting and RAG come first. Fine-tuning, distillation, custom SLMs or Forge follow only when a versioned evaluation set proves the need and lifecycle burden.
Deploy
Configuration-dependent
Managed API, an explicitly configured regional endpoint or customer-controlled open-weight inference may be selected when model availability, licence, residency and infrastructure permit it.

No direct model-to-machine path

In the Foundry architecture, no Mistral response is a direct PLC, robot, charger or vehicle command. A proposal must pass schema validation, a typed allow-list, policy and authority checks, and deterministic control. An independent safety function may reject or stop the action and remains outside model authority.

Access-dependent exploration

Robostral Navigate and Mistral’s physical-world research are exploration tracks—not current Hyperion integrations. They remain experimental until access, licence, hardware, safety review and measured results are established.

Hyperion Consulting is independent from Mistral AI. References to Mistral models and services do not imply endorsement, sponsorship, certification, reseller status, partnership or special access. Provider nationality alone does not establish data residency, sovereignty or regulatory compliance.

Verify in official Mistral documentation

First mission packs

Designed around failure—not staged perfection.

Each mission begins as planned and illustrative. It becomes measured only when the reference cell exists, the protocol is published and the result is reproducible.

Grounded maintenance

Read machine state and a governed maintenance corpus, show supporting passages, state uncertainty and propose a bounded work order—without autonomously diagnosing equipment or issuing control commands.