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.
Decision Foundry · concept reference
A planned place to make Physical AI decisions inspectable.
The approved Foundry concept, proposed stations, failure missions, evidence passport and provider boundary are preserved here as a reference.
Planned · illustrative until commissioned
Back to current Lab workThe 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.
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.
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.
Mistral-only runtime. EU endpoint. Owner-authorized.
Deterministic authority first. Mistral interprets and challenges.
The deployed environment permits deterministic code or exact, pinned Mistral models through a server-only EU endpoint. The production-allowlisted JARVIS, Search and Product Council paths are active by owner decision. Open assurance items remain disclosed; there is no global-region or provider fallback.
Calculate
Live · deterministic
Economics, capacity, availability, constraints, risk propagation and hashes are computed by inspectable code with explicit assumptions, units and formulas.
Moderate
Active · owner-authorized
The pinned moderation model must approve input and output. If moderation or EU inference is unavailable, AI fails closed while deterministic simulation stays available.
Challenge
Active · owner-authorized
Small, Medium and Large serve as separate facilitator, systems-analyst and critic roles. Disagreement is preserved, and shared model-family failure modes are disclosed.
Ground
Local evidence
The VPS owns evidence identifiers, retrieval, prompt versions and state. No built-in web search, Files, Libraries, Agents or Conversations service is part of the production path.
Prove
Versioned evidence
Every substantive result records exact models, prompt and method versions, classifications, evidence IDs, deterministic hashes, limitations, review status and a human owner.
Contain
EU-only adapter
All hosted calls originate server-side at https://api.eu.mistral.ai. Other hosts, model aliases, unapproved IDs and provider fallbacks are rejected; secrets never enter browser code.
No model-to-machine or model-to-decision path
No Mistral response is a direct PLC, robot, charger or vehicle command. Deterministic code owns arithmetic and constraints; an independent safety function may reject or stop action; a named human owns investment, release, compliance, safety and customer commitments.
Explicitly outside production
OCR, embeddings, speech, Agents, Conversations, Files, Libraries, Batch, built-in web search, fine-tuning, Forge and preview services are disabled until the exact capability is verified for EU regional processing, Zero Data Retention, contract fit, security and measured need. The one exception is a public-preview model used in the Product Council: mistral-large-4-0 (Mistral Large 4), which Mistral may update.
Hyperion Consulting is independent from Mistral AI. The EU regional endpoint is a technical boundary, not proof that the complete subprocessor and control-plane chain remains in the EU. No partnership, certification, sovereignty or compliance status is claimed.
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.
Planned · illustrative until commissioned
Authority boundary
Give the system an ambiguous or unsafe instruction. Intelligence may propose or abstain; typed interfaces, policy authority and independent safety determine whether anything can reach the controller.
Planned · illustrative until commissioned
Pull the network
Disconnect external connectivity and observe local inference, cached knowledge and deterministic fallback against declared degraded-mode requirements.
Planned · illustrative until commissioned
Operating envelope
Challenge perception and learned behaviour with glare, occlusion, unfamiliar parts, drift and workstation variation to expose—not hide—the validated boundary.
Planned · illustrative until commissioned
Canary to fleet
Test a revision in simulation, release it to one controlled asset, expand only while acceptance thresholds hold and roll back when they do not.
Planned · illustrative until commissioned
Industry mission packs
One decision system. Six physical contexts.
The first three align with Hyperion’s published industry priorities. Smart infrastructure, logistics and defence are research contexts; publication does not imply sector client delivery.
Primary
Manufacturing & robotics
Inspection, manipulation, cycle time, operator intervention, degraded modes and repeatable line deployment.
Primary
Automotive & connected mobility
In-vehicle, edge and cloud boundaries; lost connectivity; staged OTA change; fleet observability and rollback.
Primary
Energy & charging infrastructure
Local autonomy, asset and fleet coordination, offline operation, incident recovery and cost to serve.
Research
Smart infrastructure & cities
Distributed sensing, multi-owner authority, privacy, cybersecurity, continuity and public-service constraints.
Research
Logistics & autonomous fleets
Exceptions, human handover, mixed fleets, site variance, utilisation and repeatable deployment.
Research · bounded
Defence & dual-use
Unclassified, non-kinetic and non-targeting missions using public or synthetic data. No weapons targeting or autonomous lethal decisions.
Evidence Passport
Every impressive moment carries its limits.
A demonstration is useful only when a buyer can tell what happened, under which conditions and what the result does not prove.
- Live
- Runnable now; not automatically production-proven.
- Measured
- Observed under named conditions; not extrapolated beyond them.
- Simulated
- Produced in a digital or synthetic environment; not field operation.
- Replay
- A fixed prior run; not a live interaction.
- Illustrative
- A proposed scenario or architecture; implementation is not implied.
- Planned
- Intended for a future build; not commissioned.
The record travels with the result
Question · intended use · data origin and rights · model, prompt, tool, firmware and dataset revisions · operating conditions · authority boundary · acceptance and kill thresholds · expected and observed result · failures · limitations · reproduction date · what the run does not prove
JARVIS · Lab OS
The guide—not the governor.
JARVIS lets a visitor observe system state, stress a bounded scenario, compare decisions and explain the result against cited evidence. It may synthesize and propose; it cannot waive an acceptance threshold, certify a system or authorize unsafe physical action.
Ask JARVIS about the Foundry- Observe
- System state, provenance and sources
- Stress
- Bounded failure injection
- Decide
- Options and lifecycle consequences
- Explain
- Evidence, limitations and decision record
How the complete system works
One Physical AI system — from sensing to dependable action
Select a stage to follow the system end to end.
01 / 08Device & Control
Sensors & environment. The physical world, sensed — cameras, depth, radar and OT signals.
Explore the engineering approach02 / 08Embedded Platform
Embedded platform. A real-time, deterministic, securely-updatable edge runtime.
Explore the engineering approach03 / 08Perception & Robotics
Perception. Mapping, detection and world-state resolved on-device.
Explore the engineering approach04 / 08Intelligence & Knowledge
Intelligence. Open-weight models and VLM/VLA policies propose an action.
Explore the engineering approach05 / 08Device & Control
Deterministic control. A deterministic controller turns a proposal into a precise command.
Explore the engineering approach06 / 08Trust, Safety & Cybersecurity
Safety & cybersecurity. An independent boundary accepts or blocks every command.
Explore the engineering approach07 / 08Perception & Robotics
Physical action. The validated command drives the machine in the real world.
Explore the engineering approach08 / 08Orchestration & Operations
Operations & fleet. Telemetry returns to operations; the system improves across a fleet.
Explore the engineering approachDecision Foundry questions
Is the Decision Foundry already open?
No. The physical facility is in concept development and is not commissioned. Current digital demonstrations and Hyperion-owned R&D are labelled separately.
Is Hyperion a Mistral AI partner?
No partnership or special access is claimed. Hyperion is independent from Mistral AI. The owner-authorized JARVIS, Search and Product Council paths use a Mistral-only runtime through the EU endpoint, with deterministic fallbacks.
Can a Mistral model directly control a machine in the proposed lab?
No. Model outputs are proposals that must cross schema, allow-list, policy and authority checks before deterministic control. Independent safety remains outside model authority.
Is the Decision Foundry a certification laboratory?
No. It does not replace legal, safety, cybersecurity, conformity-assessment or certification specialists. It creates product and architecture evidence for a bounded decision.
What should a participant leave with?
A bounded decision record: observed evidence, unresolved assumptions, limitations and a go, redirect, sequence or stop recommendation.
Which industries does the programme cover?
Manufacturing, automotive and energy are primary contexts. Smart infrastructure, logistics and defence are explicitly labelled research contexts; publication does not imply sector client delivery.
Bring one stuck initiative
Leave with the next decision.
In a 30-minute fit call, Mohammed will identify the decision, the missing evidence and whether one of Hyperion’s three mandates is appropriate—or say plainly that it is not.
30 minutes · no obligation · Mohammed leads every conversation