
The Automotive V&V Platform Landscape, and the Layer No Tool Covers
Four layers of automotive V&V tooling, the credible vendors in each, and the one thing none of them do. We sell none of these platforms; we just have to make them add up to an argument an assessor will accept. The gap that keeps turning up is the trained model itself.
Ask which V&V platforms are most recommended for automotive safety-critical systems and every vendor will answer confidently, about the one layer they happen to sell. The useful answer depends on which of four distinct V&V layers you are actually buying. Scenario-based simulation (Foretellix, Applied Intuition, dSPACE, IPG Automotive, Ansys AVxcelerate, Siemens Simcenter). Code-level static and dynamic analysis (Vector VectorCAST, LDRA, Parasoft, MathWorks Polyspace, Perforce Helix QAC). Hardware-in-the-loop rigs (dSPACE, Vector, NI). ALM and traceability platforms (Siemens Polarion, PTC Codebeamer, Jama Connect). All four are mature and all four have credible leaders. None of them validates the trained machine-learning model sitting at the center of a modern ADAS or AV stack. Closing that gap takes a safety case, not a purchase order, and it is where most 2026 programs are exposed.
A word on why we can be blunt about this: we deploy AI systems into regulated automotive programs for a living and we sell none of the platforms below. Every strong page on this topic is either a vendor describing their own layer or a paywalled market report. What follows is the cross-category map we wish existed the last time a CTO asked us to sanity-check a tooling budget.
The short answer: four layers, four shortlists
If you have budget for one question, make it this one: which layer is my actual risk in? Programs rarely fail because they picked LDRA over Parasoft. They fail because they bought three layers thoroughly and assumed one of them covered the model.
Notice how often the same names repeat across rows. dSPACE and Vector each span three layers, Siemens spans two, and the Synopsys–Ansys merger folded physics-grade simulation into an EDA giant. The market is consolidating into full-stack suites. That cuts integration cost and raises the price of a wrong early bet, which is the real argument for insisting on open exchange formats at every layer instead of assuming you will stay inside one vendor's walls for the next decade.
Scenario-based simulation: hunting the unknown-unsafe
This is the youngest layer and the one carrying the AV/ADAS era. Foretellix effectively named the category, "Safety-Driven V&V," borrowing a coverage-driven approach from silicon verification: define the scenario space in a measurable language, generate massive variation, track coverage against it, instead of replaying a fixed regression suite. Its partnership with dSPACE pairs Foretify's scenario orchestration with dSPACE's ASM vehicle dynamics and AURELION sensor simulation, which tells you that orchestration and physics fidelity are consolidating into a single stack. Applied Intuition is the well-funded challenger with a full-stack development platform. IPG Automotive's CarMaker and Siemens' Simcenter line carry decades of vehicle-dynamics credibility. Ansys AVxcelerate brings physics-grade sensor simulation and, since the Synopsys acquisition closed, ships under joint "Ansys 2026 R1" releases (March 2026) with AI-powered features across the portfolio.
What makes this layer valuable is that it is the only one that even attempts the SOTIF problem, the unknown-unsafe scenarios your requirements never imagined. The 2026 market report on scenario-based AV testing platforms puts dSPACE, IPG Automotive, Ansys, Foretellix and NI at the center of the competitive field, vendors whose credibility with assessors rests on decades of hardware-in-the-loop heritage and ISO 26262-aligned workflows. The unsolved problem sits one level above all of them, and no vendor pretends otherwise: scenario library completeness. Nothing in the tooling demonstrates that a library covers the unknown-unsafe space, and short of exhaustive real-world mileage nothing can. Vendors are attacking it with AI-generated scenario variation, such as Foretellix's November 2025 collaboration with Inverted AI on generative agent behavior, which genuinely expands the explored space. Expanding coverage and proving sufficiency are still different claims.
Three questions to put to any vendor in this layer. What coverage metric does the tool report, and would you defend that metric in front of an assessor? Does it speak ASAM OpenSCENARIO and OpenDRIVE, or are your scenarios locked into a proprietary format? Can it run headless in CI at the scale your release cadence needs, or is it a desktop tool with a cluster license bolted on?
Code-level V&V and tool qualification under ISO 26262-8 §11
The oldest layer, and the one where "recommended" has a clear technical meaning. ISO 26262-8 §11, confidence in the use of software tools, requires you to argue that the tools producing your safety evidence cannot silently inject or mask errors. You classify each tool by tool impact and error-detection capability, derive a tool confidence level, and qualify accordingly. That mechanic is why the incumbent shortlist has barely moved in a decade. Vector's VectorCAST ships tool-qualification packages for exactly this clause. LDRA's qualification briefing makes the pragmatic point most programs run on: "for most applications… the existence of a TÜV certification is adequate evidence." Parasoft C/C++test, MathWorks Polyspace, which uses abstract interpretation to prove the absence of whole classes of runtime errors rather than merely failing to find them, and Perforce Helix QAC round out the set, all with TÜV-certified qualification kits.
The mechanics matter when you are negotiating. Tool impact asks whether a malfunction could inject an error into the safety-related work product; error detection asks how likely that error is to be caught downstream. Together they set the tool confidence level, which determines how much qualification work you owe. A vendor's kit, certificate plus safety manual plus documented known limitations plus qualification test suites, turns that from a months-long internal project into a review. Ask to see the kit before you sign, not after.
Choosing inside the layer is mostly fit. Polyspace and Helix QAC lean static-analysis-first; VectorCAST and LDRA carry the strongest unit and integration test and structural-coverage tooling; Parasoft integrates broadly across C/C++ CI pipelines. Any of them, properly qualified, satisfies an assessor for the code they analyze.
The limit is definitional. These tools verify written logic against specified behavior. A perception network's behavior is not in the C++ that wraps it, it is in hundreds of millions of learned parameters. Your inference runtime can be MISRA-clean, MC/DC-covered and TÜV-qualified end to end while the model it executes misses a pedestrian in rain. Static analysis has nothing to say about that, and to be fair, no vendor in this layer claims it does.
HIL and the certified-workflow incumbents
Hardware-in-the-loop is where dSPACE and Vector earned two decades of default trust. Real ECUs, real bus traffic, simulated plant models, deterministic timing, scripted fault injection: the bench evidence that your software behaves on the actual hardware target under conditions you could never safely produce on a track. dSPACE's SCALEXIO line and Vector's VT System are fixtures of automotive test labs, with NI supplying the modular-instrumentation alternative. The "certified ISO 26262 workflows" here are a genuine moat. Assessors know these rigs, suppliers exchange models in standard formats, and the qualification story is well-trodden.
The limits are economic. Rigs are expensive, so you buy few, so they become a bottleneck late in the program, exactly when iteration pressure peaks. And HIL runs in real time: a million scenario variations that a simulation farm clears overnight would take a rig fleet months. The 2026 equilibrium is a division of labor along the SIL-to-HIL continuum. Software-in-the-loop farms explore the scenario space at scale, HIL rigs confirm the shortlist that matters on real silicon, and the scenario definitions travel between the two, which is why the dSPACE–Foretellix pairing and Vector's presence in both the code and bench layers are strategically significant rather than just tidy. Programs that make HIL do the exploring ship late. Programs that skip HIL confirmation meet sensor-timing reality at SOP.
Traceability platforms: where ASPICE lives
The fourth layer is the one auditors see first. Siemens Polarion, PTC Codebeamer and Jama Connect own the requirements-to-risk-to-test-to-result graph that ASPICE assessments and ISO 26262 safety cases are built on. Sell into German OEMs and ASPICE capability levels are commercial table stakes, with bidirectional traceability as the spine of every assessment. These platforms are how you prove, months later and under audit, that requirement, implementation, test and result stayed connected through every change.
Codebeamer leans configurable-workflow, Polarion leans document-centric with deep OEM penetration, Jama leans usability and cross-industry familiarity (your medical and aerospace people will already know it). All three give you workflow-enforced change control, e-signature-grade audit trails and supplier collaboration at OEM scale. The honest limit is that an ALM platform is a ledger, not a laboratory. It records verification evidence produced elsewhere, and teams routinely mistake a green trace matrix for a verified system. The matrix is necessary for the audit and says nothing about whether the scenarios were the right scenarios. Model artifacts, datasets, training runs, evaluation reports, usually live outside the ALM entirely.
The layer no tool covers: the trained model
Walk the four layers again and notice what each one assumed. Scenario simulation exercises the system but cannot prove its library complete. Code analysis verifies logic that was specified, and a trained model's behavior was never specified, it was learned. HIL confirms hardware-software integration against known cases. ALM records what the others produced. Nothing in this stack validates the model itself: that its training data represents the operational domain, that its failure modes are characterized, that its performance boundaries are known, monitored and alarmed in service.
The standards community has said this out loud. ISO/PAS 8800:2024, "Road vehicles — Safety and artificial intelligence," published December 2024, exists precisely because ISO 26262 and ISO 21448 do not reach the AI-specific failure mode SOTIF calls performance insufficiency: the system operating exactly as built and still being insufficient for the situation in front of it. ISO/PAS 8800 extends both with a safety-case discipline for AI, structured claims and arguments and evidence covering the model's whole lifecycle. We wrote the builder's walkthrough separately: what an ISO/PAS 8800 AI safety case contains and how to assemble one against a real model. Concretely the evidence looks like a documented argument that training and evaluation data represent the operational design domain, characterized failure modes with quantified performance boundaries, robustness testing against the degradations the domain actually serves up (weather, glare, sensor soiling), and in-service monitoring that alarms when the model leaves the envelope it was validated in. No platform in the four layers above produces any of that as a product feature today.
If you want the worked failure, read NHTSA's Engineering Analysis EA26002, opened March 18, 2026, covering 3,203,754 Tesla vehicles after nine crashes, one fatal. The language is a model-validation finding, not a code-quality finding: the camera-based system "did not detect common roadway conditions that impaired camera visibility" and did not "provide alerts when camera performance had deteriorated." Every layer above could have been green. MISRA-clean code, passing HIL benches, an immaculate trace matrix. This failure class survives all of it untouched, because it lives in what the model learned and what the system monitored in service.
Closing that gap is not a fifth tool purchase, it is a change in how validation runs. The model's evidence has to be regenerated every time the data, the weights or the operational domain shifts, which makes validation a pipeline property rather than a milestone. We have written separately about continuous V&V in safety-critical AI systems, how the compliance loop breaks when model iteration outpaces document-era validation. The other half is in-service evidence, fleet telemetry that tells you when the model's assumptions stop holding, and that part is buildable today with ordinary engineering. The fleet telematics platform we built for AutoTrace processes tens of millions of vehicle events a day into real-time alerts and driver-safety analytics. It is operational telematics rather than a certified safety system, but it is architecturally what an in-service model-monitoring loop needs: high-volume ingestion, low-latency detection, evidence you can still query months later.
How to choose without buying twice
Recommendations by program stage, from the deployment seat:
- Running series L2/L2+ ADAS with a supplier stack? Code-level analysis, HIL and ALM traceability are table stakes your assessor expects. Scope scenario simulation to your ODD rather than buying at AV scale. Your model-validation exposure is real but bounded, so put it in the safety case early.
- Running an L3+/AV program? The scenario platform is your center of gravity. Budget for scenario-library engineering, the people and pipelines that grow coverage, not just licenses. The library is the asset, not the license.
- Anything with a learned component in it? Assign an owner for model-level evidence now. If your V&V budget names four tool line-items and nobody owns data and model validation, that gap is already on your risk register. It just is not written down yet.
A few criteria hold across all four layers. Demand tool-qualification evidence out of the box, because a TÜV certificate plus a qualification kit for your target standard beats a vendor promising to "support your audit." Prefer open formats, ASAM OpenSCENARIO and OpenDRIVE, ReqIF, so your scenarios and requirements outlive any single vendor relationship. Insist on headless, API-first operation so verification runs in CI on every change instead of on a bench schedule. And ask every vendor the 2026 question: what evidence does your roadmap produce for the AI safety case ISO/PAS 8800 describes? The answers are shorter than the brochures suggest.
The line item nobody budgets is the glue. Connecting scenario results to the trace matrix to the safety case to the in-service monitoring loop. That integration work is where programs actually die, and it is the work we do, embedded with your team, in fixed-scope engagements designed to leave the pipeline running without us. How we partner covers how those engagements are structured.
Scope a First Build if you are staring at a tool shortlist, an assessor's calendar and a model nobody has assigned validation ownership for. A month of focused work beats another vendor evaluation: the four layers mapped against your program, the uncovered one named, and an evidence plan your assessor will recognize.
Related Insights

Insight
Dark Factories in Healthcare and Automotive Are a Fantasy (For Now)
In healthcare and automotive, the fully autonomous "dark factory" runs into a wall: these industries define "done" through evidence and human accountability no agent can produce for itself.
Read More
Insight
Continuous V&V: Overcoming the Compliance Loop in Safety-Critical AI Systems
How do you deploy industrial AI with change control a regulator will accept? Define the model's adaptation boundaries up front, run V&V continuously in the pipeline rather than at release, scope re-validation through impact analysis, and keep experts on the flagged cases. FDA CSA, GAMP 5 D11 and PCCPs now expect it.
Read MoreFacing a problem like this one?
Tell us what the business needs to win and who owns it on your side. If it is a fit, a First Build starts within weeks. If it is not, we will tell you why.