Skip to content

Offboard Labeler Registry by Scale

Last updated: 2026-05-24

Offboard labelers are models, prompts, rules, tools, and human workflows that create labels outside the runtime vehicle stack. They include classical auto-labelers, heavy offline 3D detectors, foundation-model segmenters, VLM scene reviewers, LLM QA assistants, open-vocabulary point-cloud labelers, retrieval-augmented label search, and model-as-judge evaluators. They may never control a vehicle directly, but they can still change training data, semantic maps, replay assertions, safety reports, and release evidence.

The registry rule is simple: if a labeler output can enter training, replay, evaluation, a signed map, a taxonomy, or a safety case, the labeler is a release-affecting artifact. It needs identity, scope, evaluation, rollback, owner, allowed-use controls, the digest-bound trust-chain evidence defined in secure-artifact-attestation-profile.md, and the GenAI/agent evaluation controls in llmops-agent-evaluation-by-scale.md when prompts, RAG, judges, or tool calls are involved.


What Counts as an Offboard Labeler

Labeler typeExamplesWhy registry matters
Heavy offline detector/segmenterMulti-frame 3D detector, accumulated-cloud segmenter, Waymo 3DAL-style auto-labelerCan produce large volumes of training truth faster than humans can inspect
Promptable image/point labelerSAM/SAM2, Grounding-DINO, CLIP, DINOv2, SAM4D, SALT-style toolsPrompt/model changes can shift masks and names without code changes
Open-vocabulary 3D labelerZOPP, OpenUrban3D, VESPA, UniLiPs, LOSC-style consolidationUseful for unknown classes but risky if text names become taxonomy IDs
Map-derived label exporterSemantic map back-projection to single scansInherits source map, pose graph, calibration, release-state, and reviewer state
Weak labeling ruleIntensity threshold, map overlay, heuristic FOD candidate, ground marking extractorLooks deterministic but still needs versioning and QA
LLM/VLM reviewerScene captioner, incident summarizer, taxonomy suggestion assistantGenerated explanations can steer humans and safety evidence
Evaluator/judge modelModel-as-judge, VLM replay reviewer, prompt scorerShould route review, not replace release approval
Retrieval-augmented label assistantVector search over clips, maps, SOPs, prior labels, incidentsCorpus/index drift changes retrieved evidence

Treat the labeler as a system, not just a checkpoint. The system includes prompts, thresholds, retrieval corpora, calibration/projection code, input filters, taxonomy mappings, reviewer UI, acceptance policy, and export code.


Scale Ladder

MLOps scaleRegistry postureRequired controlMain risk
S0 notebook researchRun note and prompt textRecord model name, date, input sample, and output examplesInteresting labels cannot be reproduced
S1 repeatable prototypeVersioned labeler configFrozen eval set, deterministic decoding when possible, reviewer notesPrototype labels become hidden baseline truth
S2 single-product productionProduct labeler registryImmutable labeler version, prompt pack, thresholds, QA report, allowed-use stateTraining data shifts while deployed model appears unchanged
S3 fleet and multi-siteSite/ODD-aware labeler registryLocal prompt variants, site holdouts, reviewer correction by slice, rollback bundleOne global prompt or labeler fails local terminology or objects
S4 regulated safety-criticalEvidence-locked labeler recordsSafety-slice eval, expert review, retention hold, waiver/expiry, incident traceAuto-labeler silently creates unsafe safety evidence
S5 platform scaleMulti-tenant labeler/prompt/evaluator platformPolicy-as-code, tenant isolation, audit API, cost and quality SLOsTeams reuse unapproved labelers across unrelated products

The critical transition is S2. Once an offboard labeler output can change a model, map, replay assertion, or release packet, it must be governed even if the labeler itself is offline-only.


Registry Record

FieldMinimum contents
labeler_idStable name, version, owner, intended task
labeler_typeDetector, segmenter, prompt pack, rule, map exporter, LLM/VLM, evaluator, human workflow
model_artifactsCheckpoint, provider/API model ID, container, dependency lock, hardware class
attestation_refsSubject digests, signatures, provenance, SBOM where relevant, policy result, trusted-builder or provider identity
prompt_packSystem instructions, user templates, variables, examples, local terminology, prompt owner
retrieval_contextCorpus snapshot, embedding model, index build, filters, citation/trace policy
input_contractSensor streams, map tiles, calibration, source-map state, image coverage, data-quality gates
taxonomy_contractControlled taxonomy version, candidate class aliases, unknown policy, class order
thresholdsConfidence gates, abstention policy, class-specific acceptance/review/discard thresholds
projection_contractCamera-LiDAR projection, pose source, time sync, occlusion policy, map back-projection code
evaluation_setGold set, local holdouts, rare-class slices, map-state slices, reviewer disagreement sample
output_statesCandidate, pre-label, review label, QA passed, training approved, safety evidence approved
allowed_useResearch only, pre-annotation, training, replay, release eval, map publication, safety evidence
rollbackPrevious labeler/prompt/threshold bundle and affected downstream datasets
expiryEvidence expiry, revalidation triggers, waiver owner

For API-hosted foundation models, a version string is not enough. The record should include provider, model ID, hosted region, data-retention setting, prompt pack, safety settings, decoding parameters, and evaluation date.


Labeler State Machine

StateMeaningAllowed output
experimentalResearch-only labeler or promptLocal candidate examples
candidate_registryReproducible config exists but eval is incompleteCandidate pre-labels only
qa_approvedEval set and reviewer QA passed for a bounded scopePre-labels and review acceleration
training_approvedOutput may enter a named training snapshot after QAapproved_for_training labels
release_eval_approvedOutput may support replay/eval assertions under frozen evidenceRelease-eval labels, not safety approval
safety_evidence_approvedExpert-reviewed and evidence-locked for safety claimsSafety evidence labels within scope
restrictedUsable only with named exclusions or mitigationsBounded use with waiver/expiry
quarantinedRegression, incident, prompt/model drift, or data restriction foundNo new downstream consumption
deprecatedReplaced or retiredHistorical reconstruction only

Labeler approval is scoped. A prompt pack approved for "baggage tractor" candidates at one airport is not automatically approved for aircraft proximity, FOD, personnel, terminal frontage, or another country's site terminology.


Output State Contract

Output stateMeaningDownstream use
candidate_labelMachine suggested a region, class, caption, or assertionReview queue, active learning, search
pre_labeledCandidate was imported into an annotation toolReviewer acceleration only
review_labelHuman reviewer accepted, corrected, or rejected the candidateQA sampling and adjudication
qa_passedQA policy passed for the label batchTraining candidate if allowed-use permits
approved_for_trainingData steward approved for a training snapshotModel training and ablation
approved_for_release_evalFrozen for evaluation or replayRelease metrics only
approved_for_safety_evidenceEvidence owner approved and retention lockedSafety-case support within scope
rejectedCandidate was wrong or unsupportedError analysis and labeler improvement
unknown_reviewCandidate cannot be mapped safelyUnknown-region evidence or taxonomy backlog

The registry should make illegal state transitions impossible. candidate_label -> approved_for_safety_evidence should be blocked unless the workflow records reviewer, QA, allowed-use, evidence owner, and retention state.


Evaluation Gates

GateWhat to measureScale where it becomes blocking
ReproducibilitySame inputs produce same candidate set under recorded configS1
Gold-set qualityPrecision/recall/IoU, reviewer correction, class confusionS2
Slice qualitySite, zone, weather, map state, rare classes, personnel, FODS3
Prompt stabilitySmall prompt/model/corpus changes do not flip critical outputs unexpectedlyS2-S3
Projection qualityCamera-LiDAR/time/pose errors do not put labels on wrong pointsS2
Taxonomy compatibilityCandidate names map to controlled classes or explicit unknownS2
Reviewer workloadCandidate improves throughput without raising defect rateS2
Safety slice false acceptanceHazard labels are not falsely promotedS4
Retrieval trace qualitySources and top-k evidence can be reconstructedS3-S5
Downstream impactModels trained on the labels improve target slices without leakageS2-S5

Do not let "human in the loop" hide weak labeler evidence. The reviewer workflow is part of the labeler system and needs its own defect taxonomy, sampling plan, disagreement rate, and correction analytics.


Architecture Comparison

Labeling architectureAdvantagesDisadvantagesBest use
Manual labels onlyHighest trust for small safety-critical batchesSlow, expensive, weak coverageS0-S2 baselines, S4 hazard evidence
Closed-set offline auto-labelerHigh precision on known classes, easy metricsMisses novel objects and local terminologyCommon vehicles, aircraft, people, GSE, routine labels
Promptable 2D labeler lifted to 3DFast rare-object discovery, image foundation-model leverageCalibration/projection-sensitive, lighting-sensitiveCandidate masks and reviewer acceleration
Open-vocabulary 3D labelerFinds unknown concepts in point cloudsNaming instability, taxonomy risk, uneven calibrationTaxonomy discovery and local benchmark bootstrapping
Map-derived pseudo-label exporterMassive label multiplication from one reviewed mapSource-map and release-state errors contaminate trainingAggregated-map flywheel after map acceptance
LLM/VLM scene reviewerGood for triage, captions, and taxonomy suggestionsCan hallucinate, over-explain, or miss geometryReview routing and incident summarization
Model-as-judge evaluatorScales qualitative checksBias and drift; not approval authorityRouting review, comparing candidate outputs

The production pattern is layered: machine candidates reduce work, human review decides taxonomy and correctness, QA decides batch quality, data stewardship decides allowed use, and release governance decides whether a downstream model or map can consume the batch.


Semantic Map and Non-Road Rules

Aggregated-map and non-road managed-site labelers need stricter controls than generic image annotation:

  • Preserve semantic class and release state separately. A correctly named parked GSE unit, cone, worker, pallet, hose, or FOD candidate may still be wrong as permanent map truth.
  • Store source-map acceptance ID, map tile IDs, pose/back-projection code, calibration package, and hygiene layer digest for every map-derived batch.
  • Keep local terminology in prompt packs: stand, apron, tug, belt loader, ULD, terminal frontage, quay edge, warehouse aisle, mine bench, utility cabinet, campus crossing.
  • Require site/ODD slices for reviewer correction rates. A prompt that works on apron scenes can fail on terminal-frontage crowds or utility corridors.
  • Route new names through taxonomy promotion, not direct class creation.
  • Treat stationary people and staged movable assets as high-risk false-permanent examples.
  • Link labeler updates to map-derived-pseudo-label-invalidation-protocol.md when source-map, prompt, model, reviewer, taxonomy, projection, or release-state artifacts change.

For safety-critical labels, a useful auto-labeler is one that reduces reviewer burden while preserving auditability. It is not a replacement for release authority.


Monitoring and Scorecards

MetricS0-S1S2-S3S4-S5
Candidate yieldUseful examples per runAccepted/rejected/reworked by class and siteEvidence-bearing accepted yield by hazard slice
Reviewer correctionManual noteCorrection rate by labeler version, prompt pack, site, ODDExpert disagreement, false acceptance, adjudication aging
Prompt/model driftManual rerunRegression suite by prompt/model/corpus versionRelease blocker and audit trace
Taxonomy actionsNotesNew alias, parent, split/merge, unknown backlogSafety-impacting taxonomy review
Projection defectsVisual spot checkCalibration/projection error samplesSafety-slice projection audit
Allowed-use violationsNot trackedBlocked by workflow statesPolicy-as-code and incident report
CostPer batchCost per accepted label and reviewer hour savedCost per evidence pack and platform tenant

Labeler scorecards should be reviewed before training data freezes, map publication, release evaluation, and any safety-case evidence update.


Failure Modes

Failure modeConsequenceControl
Prompt text becomes class IDTaxonomy drifts and backward compatibility breaksTaxonomy promotion record and controlled class mapping
Provider model updates silentlyLabels change under the same promptProvider/model ID, eval date, regression pack, rollback bundle
Threshold changed without reviewTraining distribution shifts invisiblyThreshold version, acceptance report, downstream impact query
Reviewer accepts machine biasSystematic false labels pass QARisk-weighted QA and independent adjudication
Open-vocabulary candidates treated as truthUnknown objects become unsafe map classesCandidate-only state and reviewer taxonomy decision
Projection error lifts image labels to wrong points3D training labels are corruptedCalibration/projection QA and rejected-projection logging
Labeler improves average yield but harms hazard slicesMore labels, worse safety evidenceHazard-slice blockers and expert review
Retrieval corpus drift changes evidenceRAG-assisted label decisions cannot be reconstructedCorpus/index snapshot and retrieval trace
Labeler rollback ignoredBad labels remain in datasets after rollbackImpact graph across datasets, maps, replay, and models

  • mlops-scale-research-scope.md - MLOps maturity ladder and prompt/foundation-model operations scope.
  • model-governance-release-evidence.md - release packet evidence for offboard labelers and prompt packs.
  • map-derived-pseudo-label-invalidation-protocol.md - invalidation for map-derived training exports.
  • feature-embedding-store-ops-by-scale.md - retrieval corpus and vector-index controls.
  • secure-artifact-attestation-profile.md - signatures, SBOM/provenance, and policy verification for release-affecting labeler artifacts.
  • llmops-agent-evaluation-by-scale.md - prompt, RAG, judge, tool-agent, trace, and GenAI eval controls.
  • data-flywheel-airside.md - auto-labeling pipeline, quality gate, and closed-loop training.
  • ../data-platform/active-labeling-budget-ops.md - label-budget states and promotion boundaries.
  • ../data-platform/3d-annotation-tools.md - annotation tooling, pre-labels, and reviewer workflows.
  • ../data-platform/data-catalog-lineage-quality-ops.md - data product promotion and lineage.
  • ../../30-autonomy-stack/perception/overview/aggregated-map-semantic-segmentation.md - offboard semantic-map candidate lane.
  • ../../30-autonomy-stack/perception/methods/losc.md - LiDAR open-vocabulary segmentation consolidation.

Sources

Public research notes collected from public sources.