Skip to content

Perception-SLAM Artifact Compatibility Matrix

Last updated: 2026-05-24

Purpose

Perception-SLAM releases are not single artifacts. A safe runtime combination includes code, containers, models, TensorRT engines, maps, calibration, route/geofence overlays, parameters, diagnostics configuration, and cloud-side observability schemas. This matrix defines what must be compatible before an artifact can be activated on a vehicle or promoted through OTA rings.

The operating rule is simple: compatibility is proven by a signed manifest and validation evidence, not inferred from file names or "latest" labels.

MLOps Scale Fit

Compatibility evidence should grow with MLOps authority. A research checkpoint can be compared with a run note; a production fleet artifact needs a signed compatibility manifest that joins model, map, calibration, runtime, semantic taxonomy, telemetry schema, evidence, and rollback.

MLOps scaleCompatibility postureRequired compatibility artifactWhat must not happen
S0 notebook researchLocal comparison onlyRun note naming code, data, model, and map/calibration assumptionsResult reused without recording incompatible inputs
S1 repeatable prototypeBaseline reproducibilityDataset and environment manifest plus fixed validation scriptBaseline silently changes due to data, schema, or preprocessing drift
S2 single-product productionDeployable artifact setModel package manifest, dataset manifest, runtime package test, rollback targetCandidate model reaches shadow/canary with untested map, class order, calibration, or engine
S3 fleet and multi-siteSite/ODD compatibilityPer-site rollout manifest with active model/map/calibration/runtime/telemetry IDs and canary evidenceOne global artifact set activates across sites with different maps, sensor kits, or ODDs
S4 regulated safety-criticalEvidence-locked compatibilitySigned compatibility manifest linked to safety-case claims, replay packs, incident retention, and rollback drillBehavior authority changes while compatibility evidence is ticket-only, expired, or mutable
S5 platform scalePolicy-enforced compatibilityPolicy-as-code checks over registry, map, calibration, prompt/evaluator, telemetry, feature/embedding, and deployment servicesShared platform permits a tenant to bypass compatibility or reuse unsupported artifacts

This is the OTA/SUMS counterpart to ../mlops/mlops-reference-architectures-by-scale.md, ../mlops/model-registry-artifact-lifecycle-by-scale.md, ../mlops/serving-inference-operations-by-scale.md, ../mlops/mlops-scorecards-and-kpis-by-scale.md, and ../mlops/secure-artifact-attestation-profile.md: architecture defines where artifacts live, the registry lifecycle guide defines immutable artifact identity and alias authority, the serving guide defines endpoint, batch, traffic, scaling, and rollback controls, the scorecard defines what blocks release, the attestation profile defines artifact trust, and this matrix defines whether the artifact set can safely activate.

Compatibility Axes

AxisRequired metadataWhy it matters
Vehicle platformVehicle type, wheelbase, sensor kit, brake/steer interface, safety controller versionGeometry and actuation assumptions affect free-space and MRC behavior
Sensor hardwareSensor model, serial, firmware, timestamp mode, mounting position, health limitsModels and calibration are sensor-specific
Compute/runtimeGPU/accelerator, driver, CUDA, TensorRT, ROS distro, kernel, DDS profileEngines and latency behavior can change across runtime versions
Serving routeEndpoint or batch service ID, serving manifest, traffic policy, shadow/canary/champion state, autoscaling policyA valid artifact set can still be unsafe if served to the wrong cohort or under untested routing
ModelModel ID, training data ID, input/output schema, class ontology, precision, calibration fileConsumers must understand tensors, classes, uncertainty, and thresholds
MapSite, bundle ID, tile IDs, datum, layers, overlays, route graph, expiryPose, route, and geofence depend on exact map bundle
CalibrationIntrinsics, extrinsics, time offsets, sensor-to-base transform, verification stateFusion and map alignment fail silently with stale calibration
ConfigurationODD limits, monitor thresholds, planner margins, feature flags, diagnostics graphConfig can change behavior as much as code
Telemetry schemaOTel schema URL, robotics custom schema, event IDs, unitsDashboards and release gates must not misread fields
Artifact trustSubject digests, signatures, SBOM/provenance, policy result, trusted builder, alias approvalActivation must use the same trusted artifact set that was evaluated
EvidenceTest partition, benchmark manifest, shadow/canary results, safety-case claim IDsSUMS and safety case need reproducible approval evidence

Matrix

ArtifactCompatible withBlock conditionRequired evidence
Perception containerROS distro, message definitions, GPU driver, model runtime, diagnostics graphTopic/schema mismatch, unresolved dependency, untested DDS/QoS changeCI, SIL replay, interface contract check, SBOM/VEX
SLAM/localization containerMap format, TF tree, sensor drivers, calibration, timing stackFrame/datum change without migration test, timing policy mismatchReplay ATE/RPE, timing stress, map compatibility test
TensorRT engineGPU architecture, TensorRT/CUDA versions, model hash, precision calibrationEngine built on different accelerator/runtime or stale calibration cacheEngine build attestation, deserialization test, latency report
Serving manifestModel server, endpoint, batch job, traffic split, autoscaling, input/output schema, telemetry, rollback routeCandidate served outside approved ODD cell, or endpoint behavior differs from evaluated packageServing manifest, endpoint readiness, traffic policy review, rollback load test
Neural modelInput preprocessing, ontology, uncertainty calibration, runtime thresholdsClass/order/schema change not reflected in consumersModel card, dataset lineage, calibration and slice metrics
Occupancy/free-space modelGrid resolution, unknown semantics, planner contract, protected-zone policyUnknown/free encoding change or false-free-space gate failureFalse-free-space report, OOD/unknown object evaluation
Map bundleSite/route, localization algorithm, calibration, vehicle geometry, overlaysWrong active map, expired overlay, tile frame mismatchMap QA report, source traversal provenance, canary metrics
Semantic map layerSource map, tile manifest, taxonomy/class order, unknown/confidence policy, segmentation model, calibration, map-runtime consumersSemantic layer produced under a different manifest/hash, missing QA evidence, class-order mismatch, or runtime loader contract not validatedSemantic map manifest, per-tile QA, safety-class metrics, calibration/threshold report, Autoware loader smoke test
Calibration packageSensor serials, mounts, firmware, TF tree, vehicle body frameApplied to wrong vehicle/sensor kit or drift state redCalibration benchmark report, drift monitor record
Runtime configCode/model/map version set, ODD, monitor thresholds, release ringThreshold differs from validation without approvalConfig schema validation, safety impact record
Diagnostics graphNode names, diagnostic producers, operation modes, latch policyMissing critical node or changed severity semanticsDiagnostic graph test and alert routing proof
Observability schemaOn-vehicle telemetry, cloud pipeline, dashboards, alert rulesBreaking schema without dashboard migrationSchema version, migration test, sample event replay
Offboard labeler / prompt packLabel schema, prompt/model/checkpoint, retrieval corpus, projection/calibration hash, reviewer workflowLabeler output can change datasets, semantic maps, and release evidence without runtime model changesLabeler registry record, accepted/rejected statistics, reviewer QA, rollback impact
Evaluation/replay packScenario IDs, map/runtime compatibility hash, expected metrics, waiver stateA model can be approved against stale or incompatible replay evidenceReplay package manifest, deterministic replay evidence, evidence expiry check

Manifest Fields

Use the checked JSON Schema contracts as the narrow machine-readable surface for semantic-map promotion: ../../schemas/semantic-map-manifest.schema.json covers the semantic layer release manifest and ../../schemas/runtime-map-contract.schema.json covers the runtime map-loader handoff, with examples under ../../examples/map-contracts/.

FieldRequirement
manifest_idImmutable ID signed by release authority
vehicle_eligibilityVehicle classes, sensor kits, excluded serials, site IDs
artifact_setCode, model, engine, source map, semantic layer, taxonomy, calibration, config, telemetry schema, diagnostics graph, runtime map contract, and release-evidence digests
compatibility_hashHash over the full version set, not only individual artifacts
semantic_map_contractSemantic layer ID, taxonomy ID/hash, model/weights digest, calibration ID, confidence threshold file, QA report ID, source-map hash, tile-manifest hash, runtime export contract version
runtime_map_contractAutoware map contract version, map_projector_info.yaml hash, pointcloud_map_metadata.yaml hash, Lanelet2 loader evidence ID, pointcloud loader evidence ID, dynamic-load replay evidence ID if enabled
serving_contractService ID, endpoint or batch job, model server/runtime, input/output contract, traffic policy, autoscaling policy, telemetry fields, rollback route
activation_preconditionsParked/mission-complete state, battery, network, operator acknowledgement if required
rollback_setPrevious compatible artifact set and cache state
evidence_idsCI, replay, calibration, map QA, safety-case, security, and canary evidence
attestation_idsDigest-bound signatures, SBOMs, SLSA/in-toto provenance, vulnerability disposition, model/export/map-QA attestations, and policy results as defined in ../mlops/secure-artifact-attestation-profile.md
mlops_scaleS0-S5 authority level for the artifact set, because the required evidence and approvers differ by scale
labeler_artifactsOffboard labeler, prompt, evaluator, retrieval, and candidate-label bundle IDs when any of them affected training, semantic maps, or release evidence
expiryMaximum activation window and sunset date for temporary overlays/configs
signaturesUptane/TUF metadata signatures plus build provenance attestations

Release Gates

GatePass conditionBlocks
C0 inventoryFleet reports active and candidate artifact IDs with digestsUnknown active version on target vehicle
C1 cryptographic trustPackage signatures, metadata, and provenance verifyUnsigned package, expired metadata, failed SLSA provenance check
C2 compatibilityFull compatibility matrix passes for vehicle/site/cohort; semantic map manifest_id and compatibility_hash match all evidence packetsAny required axis unresolved or evidence produced under a different manifest
C3 validationRequired benchmark, replay, calibration, map, and runtime evidence attachedEvidence missing or produced under a different manifest
C4 activation safetyPreconditions and rollback cache verified on representative vehicleActivation during mission or no known-good rollback
C5 canary healthCanary metrics remain within baseline envelope for hold periodLocalization, free-space, OOD, latency, intervention, or support-ticket regression
C6 closurePost-deployment report links active manifests and residual issuesUnaccounted vehicles or unresolved safety alerts

Rollback and Quarantine

TriggerImmediate actionFollow-up
Engine deserialization failureKeep previous engine and mark candidate incompatibleRebuild engine for exact runtime and hardware
Map/calibration mismatchBlock dispatch for affected vehicle/siteReissue compatible map or recalibrate vehicle
Semantic layer mismatchKeep previous map/semantic bundle and mark candidate incompatibleRebuild semantic layer or republish manifest with matching taxonomy/model/calibration/evidence set
Runtime map-load failureKeep previous runtime map contract and block candidate activationRegenerate Autoware export cells/metadata/projection and rerun loader smoke/replay tests
Unknown schema in telemetryFreeze promotion and mark evidence invalidBackfill parser or republish telemetry schema
Canary false-free-space alertStop rollout and quarantine candidate manifestPreserve logs, replay event, update safety case
Security metadata failureAbort activation and revoke affected metadata if neededIncident review and key-rotation assessment

Governance Notes

UNECE R156 and ISO 24089 are road-vehicle software-update references, but the SUMS pattern applies directly to non-road airside fleets because behavior can change through software, maps, models, calibration, and configuration. Uptane protects OTA delivery against rollback, freeze, mix-and-match, and arbitrary software attacks, but it does not replace build provenance or safety validation. Use SLSA provenance for the build chain and the compatibility manifest for runtime activation.

  • 50-cloud-fleet/ota/software-update-management-system-ops.md
  • 50-cloud-fleet/ota/ota-fleet-management.md
  • 40-runtime-systems/software-operations/on-vehicle-supply-chain-runtime-security.md
  • 40-runtime-systems/ml-deployment/production-ml-deployment.md
  • 50-cloud-fleet/mlops/mlops-reference-architectures-by-scale.md
  • 50-cloud-fleet/mlops/model-registry-artifact-lifecycle-by-scale.md
  • 50-cloud-fleet/mlops/serving-inference-operations-by-scale.md
  • 50-cloud-fleet/mlops/mlops-scorecards-and-kpis-by-scale.md
  • 50-cloud-fleet/mlops/secure-artifact-attestation-profile.md
  • 50-cloud-fleet/mlops/model-governance-release-evidence.md
  • 50-cloud-fleet/observability/slam-timing-health-dashboard.md
  • 60-safety-validation/safety-case/safety-case-evidence-traceability.md

Sources

Public research notes collected from public sources.