Freedom from Interference, Applied to the Message Plane of a Multi-Tenant Fleet Twin

The clause that reshapes a distributed vehicle twin most is not about the vehicle at all. It is the freedom from interference guidance in the informative annex of ISO 26262 Part 6. Read it as a requirement on the twin's own message plane. Its list of communication faults, repetition, loss, delay, insertion, masquerade, incorrect addressing, corruption and resequencing, maps almost one to one onto what goes wrong when a graph shard, an ingest broker or a tenant lens fails, and once you accept that mapping, your compliance evidence stops being a claim about uptime and becomes a claim about which fault each boundary detects.

Group the standards by what they protect

Two families matter here, and they are separated by purpose, not by year. Functional safety standards such as ISO 26262 and its SOTIF companion protect against wrong behaviour caused by faults and insufficient specification. Cybersecurity engineering standards such as ISO/SAE 21434, with the type approval regulations for cyber security management and software updates layered on top, protect against an adversary. The twin sits where the two meet, because OTA campaign targeting, remote diagnostics and fleet health triage all read from it, so the first thing a reviewer asks is whether the twin is inside or outside the safety case. If its outputs can pick which vehicles receive an update or which ones get a remote command, it influences a safety-relevant action, and you owe an interference argument at that boundary even if the twin itself is developed as QM software.

What the exchange-of-information clause forces into the code

Take the communication fault list literally and the ingest contract writes itself. Every state mutation carries a source identity, a monotonic per-source sequence counter, the vehicle timestamp and the ingest timestamp, and an integrity check computed over the payload together with its addressing fields, tenant ID and vehicle ID included. That last detail is where multi-tenant isolation actually lives: a mutation routed into the wrong tenant's lens is the same fault class as a misaddressed frame on a vehicle bus, and it is only caught if the address is part of what the checksum protects. The graph store has to stop doing last-writer-wins. It applies mutations idempotently so repetition is harmless, rejects sequence regressions instead of overwriting, and flags gaps as loss rather than silently interpolating, which makes the store slower to converge after a broker failover but means it never presents a reordered history as the current one.

Timing and memory interference get the same treatment: per-tenant executor pools and query budgets keep one tenant's runaway GraphQL traversal from starving ingest, and tenant lenses are read projections over the canonical graph, so only the ingest path can ever write to it.

The failure that looks like valid data

The failure mode worth writing up is a partitioned graph shard. Its leader loses quorum, the minority replicas keep answering reads, and a fleet SQL join that spans two shards returns a result where one half is current and the other half is frozen, consistent in shape and wrong in substance, so a campaign selects vehicles on battery state that stopped updating minutes ago. This is the delay fault class dressed up as a correct answer, and the fix is to make every result carry the age of its oldest contributing node, computed from per-shard watermarks during query planning rather than from the wall clock at response time; in our estimate the watermark bookkeeping costs a few percent of query latency. The decision rule follows directly: any query whose result can select, command or update a vehicle must be refused when that age exceeds 2 s, and read-only analytics may be served stale as long as the age is shown to the consumer. What it costs is availability. During a partition, campaigns and remote actions stall while dashboards keep working, and the next decision you will have to defend is whether a single 2 s bound is right, since state of charge tolerates far more staleness than lock status, and per-signal bounds turn one rule into a table the reviewer will want justified row by row.

Does the Annex D argument still apply if the twin platform is developed entirely as QM software?

The annex is informative and a QM component carries no ASIL obligation of its own, but when a twin output feeds a decision taken by an ASIL-rated function or process, the receiving side must show it detects the interference faults it can inherit. Producing the fault-to-mechanism mapping at the twin boundary is usually cheaper than defending its absence, and the same table doubles as threat treatment evidence for the cybersecurity case, because masquerade, insertion and incorrect addressing are also attack paths.

Back to VehicleDigitalTwins · Domains

← Back to VehicleDigitalTwins
Engineering & Architecture Brief

Freedom from Interference, Applied to the Message Plane of a Multi-Tenant Fleet Twin

The clause that reshapes a distributed vehicle twin most is not about the vehicle at all. It is the freedom from interference guidance in the informative annex of ISO 26262 Part 6. Read it as a requirement on the twin's own message plane. Its list of communication faults, repetition, loss, delay, insertion, masquerade, incorrect addressing, corruption and resequencing, maps almost one to one onto what goes wrong when a graph shard, an ingest broker or a tenant lens fails, and once you accept that mapping, your compliance evidence stops being a claim about uptime and becomes a claim about which fault each boundary detects.

Group the standards by what they protect

Two families matter here, and they are separated by purpose, not by year. Functional safety standards such as ISO 26262 and its SOTIF companion protect against wrong behaviour caused by faults and insufficient specification. Cybersecurity engineering standards such as ISO/SAE 21434, with the type approval regulations for cyber security management and software updates layered on top, protect against an adversary. The twin sits where the two meet, because OTA campaign targeting, remote diagnostics and fleet health triage all read from it, so the first thing a reviewer asks is whether the twin is inside or outside the safety case. If its outputs can pick which vehicles receive an update or which ones get a remote command, it influences a safety-relevant action, and you owe an interference argument at that boundary even if the twin itself is developed as QM software.

What the exchange-of-information clause forces into the code

Take the communication fault list literally and the ingest contract writes itself. Every state mutation carries a source identity, a monotonic per-source sequence counter, the vehicle timestamp and the ingest timestamp, and an integrity check computed over the payload together with its addressing fields, tenant ID and vehicle ID included. That last detail is where multi-tenant isolation actually lives: a mutation routed into the wrong tenant's lens is the same fault class as a misaddressed frame on a vehicle bus, and it is only caught if the address is part of what the checksum protects. The graph store has to stop doing last-writer-wins. It applies mutations idempotently so repetition is harmless, rejects sequence regressions instead of overwriting, and flags gaps as loss rather than silently interpolating, which makes the store slower to converge after a broker failover but means it never presents a reordered history as the current one.

Timing and memory interference get the same treatment: per-tenant executor pools and query budgets keep one tenant's runaway GraphQL traversal from starving ingest, and tenant lenses are read projections over the canonical graph, so only the ingest path can ever write to it.

The failure that looks like valid data

The failure mode worth writing up is a partitioned graph shard. Its leader loses quorum, the minority replicas keep answering reads, and a fleet SQL join that spans two shards returns a result where one half is current and the other half is frozen, consistent in shape and wrong in substance, so a campaign selects vehicles on battery state that stopped updating minutes ago. This is the delay fault class dressed up as a correct answer, and the fix is to make every result carry the age of its oldest contributing node, computed from per-shard watermarks during query planning rather than from the wall clock at response time; in our estimate the watermark bookkeeping costs a few percent of query latency. The decision rule follows directly: any query whose result can select, command or update a vehicle must be refused when that age exceeds 2 s, and read-only analytics may be served stale as long as the age is shown to the consumer. What it costs is availability. During a partition, campaigns and remote actions stall while dashboards keep working, and the next decision you will have to defend is whether a single 2 s bound is right, since state of charge tolerates far more staleness than lock status, and per-signal bounds turn one rule into a table the reviewer will want justified row by row.

Technical FAQ: Does the Annex D argument still apply if the twin platform is developed entirely as QM software?

The annex is informative and a QM component carries no ASIL obligation of its own, but when a twin output feeds a decision taken by an ASIL-rated function or process, the receiving side must show it detects the interference faults it can inherit. Producing the fault-to-mechanism mapping at the twin boundary is usually cheaper than defending its absence, and the same table doubles as threat treatment evidence for the cybersecurity case, because masquerade, insertion and incorrect addressing are also attack paths.

Related Architecture Specifications (AUTOMOTIVESDV): automotivev2g.com · a2humancard.com · automotivequantum.com

More Technical Articles on vehicledigitaltwins.com

Graph Databases for Automotive Digital Twins: Modeling Complex E/E Topologies →Multi-Tenant Telemetry Isolation Architectures for Connected Vehicles →