Skip to content

Evaluation Platforms and Replay Gates by Scale

Last updated: 2026-05-24

An evaluation platform is the control surface that decides whether a model, map, labeler, prompt pack, runtime package, or training export has earned more authority. It is not a dashboard of aggregate accuracy. At autonomy scale, evaluation must join offline metrics, slice metrics, calibration, OOD behavior, scenario replay, runtime package smoke, shadow disagreement, canary exposure, delayed labels, safety-case claims, and rollback readiness.

Use this page when designing the evaluation layer for MLOps programs from S0 research through S5 shared platforms. Use mlops-scale-research-scope.md for the maturity ladder, experiment-tracking-reproducibility-by-scale.md for run authority, dataset-split-leakage-controls-by-scale.md for independent release data, serving-inference-operations-by-scale.md for serving manifests, endpoint readiness, package parity, traffic policy, and shadow/canary telemetry, pipeline-orchestration-release-workflows-by-scale.md for workflow placement, site-sliced-release-evidence-by-scale.md for ODD-cell release manifests, and model-governance-release-evidence.md for the release packet that consumes evaluation evidence.

The core rule is: the evaluated artifact set must be the released artifact set. A checkpoint score is not release evidence if the deployed ONNX/TensorRT engine, container, map, calibration package, semantic taxonomy, prompt/labeler, replay pack, telemetry schema, or threshold policy differs from what was evaluated.


What Evaluation Owns

Evaluation should be treated as a product with explicit owners, interfaces, and service-level goals.

Owned surfaceWhat it controlsFailure it prevents
Metric specificationPrimary metric, secondary metrics, denominator, thresholds, confidence interval, and slice definitionsTeams compare candidates with incompatible metrics
Evaluator implementationCode, container, dependency lock, hardware assumptions, random seed policy, and output schemaMetric drift caused by evaluator changes
Evaluation datasetsHoldouts, site/local splits, safety holdouts, replay packs, delayed-label samples, and benchmark snapshotsTraining, tuning, or replay data leaks into release gates
Slice taxonomySite, route, zone, weather, lighting, map state, object class, vehicle kit, sensor health, task, and release-state labelsAggregate gains hide local or safety-critical regressions
Scenario replay suiteKnown incidents, mined scenarios, hazard cases, changed-map tiles, normal-operation clips, and regression-required scenariosField failures are not converted into repeatable gates
Runtime package checkONNX/TensorRT/container loading, class order, latency, memory, determinism, and target hardware compatibilityOffline checkpoint passes but deployed artifact fails
Shadow and canary evidenceDisagreement, interventions, latency, monitor triggers, exposure denominator, and delayed labelsLive exposure is treated as anecdotal rather than evidence
Waiver and exception recordsOwner, expiry, residual risk, mitigation, and revalidation triggerKnown regressions become permanent informal exceptions
Evaluation service healthQueue time, flake rate, cost, stale scenario age, evaluator availability, and evidence completenessShared evaluation becomes too slow or too untrusted to use

Evaluation outputs should be immutable once they influence release. If a threshold, slice set, evaluator container, replay package, or dataset snapshot changes, the comparison state changes too.


Scale Ladder

MLOps scaleEvaluation postureMinimum gateAnti-pattern to block
S0 notebook researchLocal metric scripts and failure examplesRecord data pointer, split note, metric version, and representative failures when reusing a resultClaiming deployment readiness from exploratory validation
S1 repeatable prototypeFrozen split and deterministic evaluatorBaseline report with environment, metric spec, confidence interval, and failure samplesMoving the baseline while keeping the same scorecard
S2 production productCandidate evaluation laneIndependent holdout, replay smoke, calibration/OOD checks, runtime package smoke, rollback target, and release packetPromoting a checkpoint whose deployable package was not evaluated
S3 fleet and multi-siteODD-cell evaluationSite/route/weather/map-state slices, mined replay, shadow, canary, delayed-label join, and artifact compatibilityOne global aggregate approves every site or district
S4 regulated safety-criticalEvidence-locked evaluationSafety-case-linked claims, hazard replay, monitor impact, waiver expiry, rollback drill, and retention holdShipping with unresolved safety regressions hidden behind average metrics
S5 platform scaleShared evaluation serviceMulti-tenant eval registry, scenario catalog, policy templates, audit API, platform SLOs, and exception workflowEvery team creates incompatible release gates and cannot compare evidence

The transition from S2 to S3 is the major autonomy boundary. S2 can evaluate one product for one operational context. S3 must prove that the artifact is safe for this specific ODD cell, not just globally better.


Evaluation Authority States

Use explicit states so exploratory reports do not become release truth.

StateMeaningMay block release?
debug_evalLocal, ad hoc metric or visualization used to understand behaviorNo
baseline_evalFrozen comparison used to compare future candidatesBlocks only baseline mutation
candidate_evalEvaluation of a registry candidate before shadow/canaryYes, for S2+ release review
runtime_package_evalEvaluation of the deployable container, engine, map, calibration, schema, and thresholdsYes, for S2+ deployment
site_evalODD-cell evaluation for a named site, zone, route, task, weather, and artifact setYes, for S3+ expansion
safety_evidence_evalEvidence-locked evaluation tied to safety-case claims, hazards, waivers, and retentionYes, for S4+ behavior authority
incident_replay_evalRegression evaluation produced after an incident, drift event, or operator-triggered concernYes, when scenario state is regression_required
platform_benchmark_evalShared platform comparison across tenants, model families, or foundation-model toolsBlocks only if policy declares it release-relevant

Authority should flow from run manifest to evaluation manifest to release packet. A report without authority state, dataset identity, evaluator identity, and artifact-set identity is useful for debugging but not for release.


Evaluation Manifest Contract

At S2+, every release-relevant evaluation should emit a machine-readable manifest.

FieldRequired contents
eval_idImmutable evaluation ID, authority state, owner, timestamp, workflow run ID
candidate_artifactModel, map, semantic layer, prompt/labeler, adapter, runtime package, or training export under review
baseline_artifactChampion or previous approved artifact set used for comparison
artifact_set_hashHash over model, engine/container, map, calibration, telemetry schema, taxonomy, evaluator, thresholds, and replay package
dataset_scopeDataset snapshot, split IDs, local holdout IDs, safety holdout IDs, delayed-label sample, access class
leakage_reportTrain/tune/eval/replay/site-holdout overlap checks and allowed-use state
metric_specMetric names, thresholds, confidence intervals, bootstrap or statistical method, denominator, abstention/OOD policy
evaluator_identityEvaluator code commit, container digest, dependency lock, hardware class, random seed policy
slice_setSite, route, zone, weather, lighting, map state, class, sensor, task, vehicle kit, release-state labels
scenario_replayScenario suite IDs, ASAM/OpenSCENARIO-aligned package version where used, simulator/runtime config, seeds
runtime_smokePackage load, class order, latency, memory, determinism, hardware target, compatibility manifest
shadow_canaryExposure denominator, cohort, time window, disagreement taxonomy, interventions, delayed-label join
known_failuresFailure IDs, severity, waiver state, owner, expiry, mitigation, revalidation trigger
decisionPass, hold, reject, canary-only, restricted, rollback, deprecated, or requires safety review
retentionEvidence retention class, audit/export location, garbage-collection hold

The manifest should be easy to produce locally at S1, but strict enough to become a service API at S5.


Architecture Options

ArchitectureAdvantagesDisadvantagesBest use
Local script or notebook evaluatorFast, transparent, cheap, easy to modifyWeak lineage, fragile dependencies, easy to compare against moving dataS0 discovery and narrow debugging
CI/pytest-style metric smokeSimple gate, close to code, catches obvious regressionsToo shallow for release, weak slice coverage, poor large-data handlingS1 baseline checks and package-level smoke
MLflow Evaluate-style run evaluationConnects metrics, artifacts, datasets, model registry, and experiment trackingNeeds custom wrappers for robotics-specific replay, map, calibration, and runtime evidenceS1-S3 model comparison and registry-linked reports
TensorFlow Model Analysis-style sliced evaluationStrong pattern for scalable sliced metrics and model analysisTensorFlow/Beam orientation may not fit every robotics stack; replay and runtime checks remain separateS2-S4 slice-heavy evaluation and fairness/ODD analysis
Evidently-style tests and reportsStrong data-quality, drift, reference-vs-current checks, and pass/fail test framingNot a substitute for scenario replay or safety-case evaluationS2-S5 data, monitoring, delayed-label, and regression-test scorecards
Managed cloud evaluation serviceFast integration with managed registries, pipelines, endpoints, and model metadataCloud coupling, data-residency constraints, limited custom scenario semanticsS2-S5 cloud-native teams with standard model families
Custom scenario replay serviceDirectly represents incidents, maps, actors, runtime packages, and ODD cellsExpensive to build, must control simulator validity, flake, cost, and coverageS3-S4 autonomy releases and managed-site regression gates
Simulation/digital-twin evaluation platformTests rare hazards and map/site changes before field exposureFidelity limits, scenario authoring cost, possible false confidenceS3-S5 when natural exposure is too slow or unsafe
Data warehouse or BI scorecardGood for fleet-wide trends, delayed labels, and operational KPIsCan become passive dashboarding with no release authorityS2-S5 monitoring, post-release learning, and executive review

Most autonomy programs need a hybrid: lightweight model-eval tooling for metric reports, a scenario replay path for behavior regressions, a monitoring scorecard for delayed evidence, and a governance layer that controls which outputs can block release.


Evaluation Layers

LayerQuestion answeredRequired from
Unit metric checkDoes the evaluator still run and produce expected fields?S0
Offline holdoutDoes the candidate improve on independent data?S1
Slice evaluationDid any required class, site, weather, route, map state, or task regress?S2
Calibration/OOD/uncertaintyAre confidence, abstention, unknown, and threshold behavior controlled?S2
Runtime package evaluationDoes the deployable artifact load and meet edge constraints?S2
Serving manifest evaluationDoes the endpoint, batch job, traffic route, autoscaling policy, telemetry, and rollback path match the approved artifact set?S2
Replay regressionDo known incidents and required scenarios still pass?S2
Shadow disagreementDoes the candidate disagree with champion in acceptable ways on live inputs?S3
Canary/delayed-label evaluationDoes limited authority produce acceptable outcomes in the target ODD cell?S3
Safety-evidence evaluationDoes the evidence support a safety claim with retention, waiver, and rollback controls?S4
Platform benchmarkCan teams compare across shared model families, tenants, or evaluation templates?S5

The layers are cumulative. A canary does not erase a replay regression. A replay pass does not prove live distribution coverage. A strong platform benchmark does not prove local site readiness.


Replay Gate Design

Replay should be a release gate when a model can affect operations, maps, training labels, or safety evidence.

GateInputsOutput
Replay package validationRaw log ID, map bundle, semantic layer, calibration, telemetry schema, scenario file, runtime configClean-worker reproducibility report
Artifact-set compatibilityCandidate model/runtime/map/calibration/taxonomy/prompt/evaluator IDsCompatibility hash and mismatch blockers
Scenario selectionRequired suite, changed tiles, recent incidents, local ODD cells, rare hazards, normal-operation controlsSuite manifest and coverage summary
Metric executionExpected behavior, clearance, stop distance, lane/zone compliance, localization bound, perception correctnessPass/fail plus metric deltas and videos/logs
Flake controlRepeated run policy, deterministic seeds, simulator/runtime version, hardware classStable/failing/flaky classification
Waiver reviewFailure severity, residual risk, mitigation, owner, expiry, revalidation triggerHold, restricted release, or reject decision

Replay suites need both failure cases and normal cases. A suite built only from incidents can overfit release gates to known failures while ignoring routine operations, comfort, throughput, and sensor-health conditions.


Managed-Site and Urban-District Scope

Non-road managed sites require evaluation that is more local than public-road aggregate benchmarks:

  • Airports need stand, apron lane, terminal frontage, service road, jetblast, de-icing, FOD, aircraft-proximity, and ground-crew slices.
  • Ports and yards need quay, container lane, trailer, crane, worker, blind-corner, gate, and weather/dust slices.
  • Warehouses and campuses need aisle, crossing, shared pedestrian space, loading bay, temporary work-zone, and shift-change slices.
  • Urban districts that are not road-driving ODDs still need district-level map state, building frontage, sidewalk/service-lane geometry, construction, utility assets, crowds, and local operating rules.

For these settings, release evidence should be ODD-cell evidence. Public datasets and city-scale point-cloud benchmarks can pretrain models and expose failure modes, but they do not approve a local deployment. Local holdouts, replay scenarios, map-change cases, and delayed labels must carry the target site, route, task, and artifact-set IDs.


LiDAR, Image, and Semantic-Map Evaluation

Aggregated LiDAR maps and LiDAR-image pipelines need evaluation contracts that preserve modality and release-state semantics.

Evaluation topicRequired controls
LiDAR-only map segmentationEvaluate point density, occlusion, intensity, scan-angle bias, multi-session registration quality, and tile stitching seams
LiDAR plus image fusionRecord calibration/projection hash, camera availability, lighting/weather state, image timestamp skew, and fallback behavior when image evidence is missing
Image-distilled LiDAR modelTreat teacher model, prompt pack, projection QA, and distillation dataset as release-affecting artifacts if they influence labels or thresholds
Map-derived pseudo-labelsKeep semantic class separate from release-state labels such as permanent_static, dynamic_residual, static_transient, movable_static, artifact, and unknown_review
ML-related SLAM inputsEvaluate source-map geometry, loop/registration residuals, dynamic-object removal sidecars, static-transient quarantine, and changed-tile impact
Runtime semantic-map consumersVerify map layer digest, taxonomy version, schema URL, compatibility hash, and rollback map before release

For semantic maps, mIoU is necessary but insufficient. The evaluation must also show whether wrong points are release-eligible, whether false permanent structure can enter a runtime map, and whether static-but-transient objects such as stationary people, parked equipment, pallets, cones, or temporary barriers were quarantined instead of learned as permanent truth.


Training and Evaluation Architecture Coupling

Training architecture affects the evidence architecture. The evaluation platform should make the coupling explicit.

Training architectureEvaluation advantageEvaluation disadvantageBest release posture
Centralized global modelOne comparable artifact and broad data coverageHides local regressions and dominant-site biasRequire ODD-cell slices before S3 expansion
Global backbone plus site adapters/LoRAGood transfer with small local evidence packetsAdapter registry, routing, rollback, and per-site thresholds add complexityEvaluate each adapter as a release artifact
Site-specific modelsStrong local fit and simple local thresholdsFragmented evidence, high maintenance, weak cross-site learningUse only for high-value or unusual ODD cells
Federated or hybrid trainingSupports data-residency constraints and local privacyHarder client lineage, aggregation evidence, poisoning checks, and local holdout disciplineRequire client manifests, local evals, and aggregation reports
Self-supervised pretrain plus supervised fine-tuneReduces label demand and improves representation reusePretraining corpus can introduce hidden bias or privacy constraintsRecord pretrain corpus, allowed-use state, and downstream release slices
Offboard/foundation-model labeler assisted trainingSpeeds labeling and open-vocabulary discoveryPrompt/model drift can contaminate labels and eval assertionsRequire labeler registry, reviewer disposition, and eval-pack gates
Continual or triggered retrainingFaster response to drift and incidentsCan turn monitoring into uncontrolled behavior changeMine data automatically; promote only through explicit eval gates

The evaluation service should not simply score the final model. It should also evaluate whether the training route was eligible for the requested release authority.


Release Blockers

BlockerApplies fromWhy it blocks
Missing evaluation manifestS1 for baselines, S2 for releaseThe report cannot be reproduced or tied to artifact authority
Unsupported evaluator comparisonS1-S5Candidate and baseline used different metric specs, slice sets, thresholds, or evaluator versions
Evaluation data leakageS1-S5Training, tuning, pseudo-labeling, replay mining, feature building, or local holdouts contaminated the release gate
Runtime/evaluation artifact mismatchS2-S5The released engine/container/map/calibration/taxonomy is not the evaluated artifact set
Serving/evaluation mismatchS2-S5Endpoint, batch service, traffic policy, input schema, preprocessing path, autoscaling setting, or rollout scope differs from evaluated evidence
Replay suite mismatchS2-S5Required incidents, changed tiles, hazard scenarios, or ODD-cell cases are missing
Target slice regressionS3-S5Aggregate score improves while site, class, weather, map-state, or task slice regresses
Missing shadow/canary denominatorS3-S5Live exposure cannot support the requested ODD-cell decision
Stale evaluationS3-S5Map, calibration, taxonomy, dataset, evaluator, or runtime changed after the report
Flaky replay or evaluatorS3-S5Pass/fail state is not stable enough for release authority
Missing waiver owner or expiryS4-S5Residual risk has no accountability or revalidation trigger
Platform policy bypassS5Shared service cannot prove tenant isolation, evidence completeness, or release authority

The blocked state should preserve evidence. A held release is useful: it names the missing artifact, owner, and next evaluation action.


Scorecards

KPIMeaning
Evaluation manifest coveragePercentage of baseline/candidate/release evaluations with complete manifest fields
Comparable-baseline ratePercentage of candidates compared against a valid baseline with matching metric spec and slice set
Slice coverageRequired ODD/class/weather/map-state slices with sufficient denominator
Replay pass rateRequired scenarios passed by severity, ODD cell, and artifact set
Replay flake ratePercentage of scenarios whose pass/fail state changes without artifact change
Evaluation lead timeTime from candidate registration to evidence-complete decision
Evaluation costCost per candidate, replay hour, ODD-cell approval, and safety evidence packet
Stale scenario ageAge of regression-required scenarios and last passing release
Waiver agingOpen waivers by severity, owner, expiry, and mitigation
False pass / false block ratePost-release failures after pass, and rejected candidates later found acceptable
Platform SLOEval service availability, queue time, artifact resolution success, and support response

At S5, platform scorecards should measure adoption and bypass attempts, but product teams still own release decisions. A shared evaluation service can standardize evidence; it cannot absorb accountability for unsafe release scope.


  • mlops-scale-research-scope.md - MLOps maturity ladder and lifecycle controls.
  • mlops-reference-architectures-by-scale.md - architecture placement for eval services and durable interfaces.
  • mlops-migration-checklist-by-scale.md - when to add evaluation services and platform controls.
  • mlops-scorecards-and-kpis-by-scale.md - KPI families, release blockers, and operating cadence.
  • experiment-tracking-reproducibility-by-scale.md - run authority and reproducibility inputs to evaluation.
  • pipeline-orchestration-release-workflows-by-scale.md - workflow state machine for build/eval/release separation.
  • serving-inference-operations-by-scale.md - serving manifests, endpoint readiness, traffic routing, autoscaling, package parity, and rollback checks.
  • platform-sre-reliability-by-scale.md - evaluation-service SLOs, error budgets, replay capacity, incident lanes, backup/restore, and platform bypass controls.
  • dataset-split-leakage-controls-by-scale.md - split firewall and leakage controls for release evaluation.
  • model-monitoring-drift-response-by-scale.md - monitoring, delayed-label, replay, and incident response inputs.
  • site-sliced-release-evidence-by-scale.md - ODD-cell manifests and local release decisions.
  • model-governance-release-evidence.md - release packets, aliases, approvals, and rollback evidence.
  • feature-embedding-store-ops-by-scale.md - embedding/vector index evidence for scenario mining and eval data.
  • offboard-labeler-registry-by-scale.md - prompt/model/evaluator artifacts that affect labels or evals.
  • secure-artifact-attestation-profile.md - signed eval packs, replay packs, and runtime artifacts.
  • federated-privacy-preserving-training-policy-by-scale.md - evaluation gates for federated, local, and hybrid training.
  • llmops-agent-evaluation-by-scale.md - judge, prompt, RAG, VLM/VLA, and agent evaluation controls.
  • ../data-platform/replay-scenario-mining-ops.md - scenario mining and replay package promotion.
  • ../data-platform/fleet-data-pipeline.md - raw logs, data products, lineage, and retention.
  • ../../40-runtime-systems/ml-deployment/production-ml-deployment.md - runtime packaging and deployment evidence.
  • ../../60-safety-validation/verification-validation/testing-validation-methodology.md - simulation, replay, and validation methodology.
  • ../../60-safety-validation/verification-validation/shadow-mode.md - dual-stack shadow evidence.
  • ../../30-autonomy-stack/perception/overview/aggregated-map-semantic-segmentation.md - semantic-map release-state and training-export controls.
  • ../../30-autonomy-stack/localization-mapping/overview/ml-related-slam-research-scope.md - learned SLAM and map-cleaning handoff controls.

Sources

Public research notes collected from public sources.