Hermes · physical AI assurance
Physical AI Deserves Better Evidence.
Where Hermes stands
An honest starting point.
Today
- Four questions — can it perceive, decide, fail and change safely — across twelve domains
- The job API, readiness report and failsafe are built and tested; a stand-down revokes the vehicle connector credential
- Every domain returns inconclusive, so a full scan reads “requires additional evidence”
- No customer deployments or pilots yet
Hermes does not
- Certify a vehicle or system as safe
- Claim to prevent crashes or incidents
- Take control of a real vehicle or machine
- Replace your safety team’s authority
- Treat a standards mapping as certification
From digital output to physical consequence
Sense. Decide. Act. Monitor.
Stage 1 of 4 · Sense
- Where uncertainty enters
- Sensors drift, degrade, disagree or are spoofed. The system can be confident about a world that isn’t there.
- What it could lead to
- An obstacle missed, or one that doesn’t exist acted on.
What does the system believe — and how would we know it was wrong?
Operating layers
Where a physical AI can go wrong.
Whether what the sensors report can be trusted — and what happens when it can’t.
What can go wrong
- Degraded or miscalibrated sensors
- Spoofed or injected signals
- Silent disagreement between sensors
Evidence to gather
- Calibration and health state at test time
- Behavior under degraded inputs
- Logs of sensor disagreement
Controls to test
- Plausibility and cross-checks
- Degraded-mode triggers
- Tamper detection
How the system understands the scene and anticipates what others will do.
What can go wrong
- Misclassified objects
- Missed rare cases
- Overconfident predictions
Evidence to gather
- Scenario coverage, including rare cases
- Confidence versus correctness
- Known gaps stated as unknown
Controls to test
- Uncertainty thresholds
- Conservative defaults
- Coverage reviews
How the system chooses what to do — and what it falls back to when it can’t decide safely.
What can go wrong
- Plans outside the safe envelope
- Fallbacks that never trigger
- Unsafe behavior at hand-over
Evidence to gather
- Envelope checks across scenarios
- Fallback activation records
- Hand-over timing
Controls to test
- Envelope enforcement
- Minimal-risk manoeuvre
- Human take-over paths
The networks and interfaces that let the system talk to the world — and let the world talk back.
What can go wrong
- Exposed interfaces
- Credential misuse
- Untrusted messages acted on
Evidence to gather
- Interface and access map
- Authentication of inbound commands
- Isolation between systems
Controls to test
- Least privilege
- Message authentication
- Segmentation
How software and models change after release — and what each change invalidates.
What can go wrong
- Unvalidated updates
- Rollbacks that don’t roll back
- Behavior changes nobody tested
Evidence to gather
- Version and change history
- What each update touched
- Retest results per update
Controls to test
- Signed, staged rollout
- Targeted revalidation
- Tested rollback
What the system records in service, and whether an incident can be reconstructed afterwards.
What can go wrong
- Gaps in recorded events
- Drift nobody notices
- Incidents that can’t be replayed
Evidence to gather
- Event logs with versions
- Replay in simulation
- Trend reviews across the fleet
Controls to test
- Retention and integrity of logs
- Replay before field changes
- Escalation to safety owners
Lifecycle gates
Evidence at every checkpoint.
Requirements ↔ Validation
What the system must and must not do, including its operating domain. Checked against the requirements, in the conditions that matter.
System design ↔ System verification
How sensing, planning and control fit together, with their limits. Checked end to end, in simulation and on the bench.
Component design ↔ Component verification
What each part must guarantee to the rest. Each part checked against its own guarantees.
Implementation
Software, models and hardware are built and versioned.
Change & in-service
After release, every update and incident reopens the evidence it affects.
Illustrative. Your lifecycle, gates and standards stay your own.
Component and integrated system
Testing a model is not assuring a vehicle.
| Path | Component check | Integrated system |
|---|---|---|
| Perception | The detector meets its accuracy targets on a labeled test set. | The vehicle responds safely when the detector is wrong in the scene that matters. |
| Decision | The planner stays within its envelope for given inputs. | The whole stack holds the envelope as inputs degrade and timing varies. |
| Actuation | Commands are within limits at the interface. | The physical response matches the command under real load and latency. |
| Communications | Messages are authenticated at each interface. | A compromised or delayed message cannot push the system outside its limits. |
| Fail-safe | The fallback triggers when its condition is met in isolation. | The fallback triggers in time, from the states the system actually reaches. |
| Incident analysis | Each part logs its own events. | The incident can be reconstructed across parts, versions and time. |
Where it is headed
Vehicles and fleets first.
- First focusVehicles and fleetsThe vehicle AI stack, fleet monitoring and over-the-air model updates — Hermes’s scope.See the vehicle use case
- Outside scope todayIndustrial and manufacturing AIOperational AI where equipment, processes and suppliers meet real-world effects — not part of Hermes.See manufacturing and infrastructure
- Outside scope todayRobotics and logisticsMachines that move goods and work alongside people — not part of Hermes today.Talk to us about scope
After deployment
When the software changes, so does the evidence.
Update published
perception model 4.2 · OTAObservedA new perception model is staged for part of the fleet.
Evidence marked stale
claims tied to model 4.1StaleEvery conclusion that rested on the previous model is reopened — not the whole system.
Targeted replay
simulation · benchDesignedRecorded scenarios are replayed against 4.2 offline, before any field exposure.
Decision recorded
safety ownerHuman decisionThe accountable safety owner decides whether to widen the rollout, restrict it or hold.
In-service monitoring
fleet events · versionedDesignedEvents are logged with the versions that produced them, so any incident can be traced to a change.
Safety-case support
Evidence your safety team can use.
- 01Assurance scopeSystems, versions and operating conditions covered — and what was excluded.
- 02Test conditionsSimulation, bench or recorded data; environments, seeds and configurations.
- 03Coverage and gapsWhat was assessed, what was not, and what remains unknown.
- 04FindingsObserved behavior, with severity, reproduction conditions and limits.
- 05Evidence provenanceWhere every result came from, and the versions it was true of.
- 06Human reviewWho reviewed the evidence, and the decisions that remain theirs.
Not a safety case or a compliance statement. The safety case and every safety decision stay with your team.
Mythos AI Security
Talk to us about physical AI.
Tell us what you are building or reviewing. We will be clear about what Hermes can and cannot evidence today.











