Skip to content

Map-Derived Pseudo-Label Invalidation Protocol

Last updated: 2026-05-24

Map-derived pseudo-labels are high-leverage and high-risk. A reviewed semantic map can back-project millions of labels into single-scan training data, but every exported label inherits the source map, pose graph, calibration, taxonomy, release-state decision, cleaner policy, reviewer state, and projection code that produced it. If any upstream artifact is corrected, the derived labels may become stale or unsafe.

This protocol defines when a map-derived pseudo-label batch must be quarantined, rebuilt, or permanently retired. It is the P0 control behind the mlops-scale-research-scope.md backlog item "Map-derived pseudo-label invalidation protocol." Use dataset-split-leakage-controls-by-scale.md for the companion split-firewall policy that prevents map-derived labels from contaminating release evaluation, replay, or local holdouts.


Scope

The protocol applies to any label derived from an aggregated map and then consumed by training, evaluation, replay, active learning, benchmarking, or release evidence:

Derived artifactExampleWhy invalidation matters
Back-projected scan labelsSemantic map labels projected into contributing LiDAR sweepsPose or calibration fixes can move labels onto the wrong points
Map-derived training tableSingle-scan segmentation rows created from a semantic mapTraining can learn stale map mistakes as ground truth
Release-state maskspermanent_static, movable_static, fod_candidate, artifact, unknown_review masksA release-state change can convert a positive label into an ignore or review label
Pseudo-label confidence fieldsPer-point or per-cluster confidence from a map segmenter or labelerThreshold or calibration changes alter allowed use
Replay/eval labelsMap-derived expected outputs for scenario replayA candidate can pass against stale expected labels
Foundation-model candidate labelsSALT/SAM/OpenUrban3D/VLM prompt outputs over map tilesPrompt/model/corpus changes can alter candidate evidence

The safest rule is simple: a pseudo-label batch is valid only for the exact source-map, semantic-layer, taxonomy, release-state, calibration, projection, and reviewer state recorded in its manifest.


Invalidation Triggers

TriggerExamplesDefault action
Source-map correctionPose graph update, loop-closure fix, tile re-registration, map merge rerunQuarantine derived labels for affected tiles and sessions; rebuild back-projection
Calibration updateLiDAR extrinsics, camera intrinsics, time offset, sensor serial swap, thermal drift correctionReproject or invalidate labels that used the old calibration
Taxonomy changeClass split/merge, new unknown policy, class order change, facade or utility sublabel additionRebuild label mappings; block training until class compatibility is proven
Release-state changemovable_static becomes permanent_static, FOD review overturns, artifact confirmed, temporary overlay expiresRecompute training masks and remove invalid positives
Cleaner or permanence policy changeDynamic removal threshold, static-transient quarantine, FOD retention rule, map hygiene policy updateRebuild hygiene layers and all downstream training exports
Projection/back-projection code bugFrame transform, timestamp interpolation, occlusion, nearest-neighbor assignment, tile stitch bugQuarantine affected batches and models trained on them
Reviewer decision overturnedAudit finds label error, safety owner rejects promotion, waiver expiresMark derived labels suspect and create rebuild/review task
Split or leakage violationTraining export overlaps validation/test/replay package or held-out siteRetire affected split; retrain/evaluate with clean split
Prompt/labeler/evaluator changePrompt pack, model checkpoint, retrieval corpus, judge threshold, open-vocabulary mapping changesRe-evaluate candidates; keep old batch frozen or deprecated
Privacy/deletion requestCustomer/site offboarding, retention expiry, data-use restrictionPropagate deletion or restriction to derived labels, features, models, and evidence
Incident linkageField incident suggests source labels contributed to unsafe behaviorFreeze evidence, quarantine related labels, and run root-cause review

Not every trigger requires retraining. It does require an impact query that can answer which labels, datasets, model versions, evaluation packs, semantic maps, and release packets consumed the affected source.


State Machine

StateMeaningAllowed use
activeBatch is current for its manifest and allowed usesTraining/eval/replay as declared
suspectUpstream trigger may affect the batch; impact not yet knownRead-only inspection; no new release training
quarantinedBatch is affected or cannot be proven unaffectedNo training or release evidence; active-learning review allowed
rebuiltBatch was regenerated from corrected sourcesCandidate use only until QA passes
reapprovedRebuilt batch passed QA and split checksTraining/eval/replay as declared
deprecatedHistorical batch retained for audit, not future useIncident/audit reference only
purged_or_restrictedData must not be used due to privacy, contract, or retention ruleRemove from derived stores; retain only legally permitted audit record

Promotion from suspect to active is allowed only with a negative impact analysis: the trigger did not touch the batch's source map tiles, calibration scope, taxonomy fields, release-state labels, split membership, or allowed use.


Impact Graph

Every pseudo-label batch should be reachable through this graph:

Upstream nodeDownstream nodes to query
Raw survey/sessionsource map, semantic map, training export, replay/eval packages
Pose graph / source mapsemantic layer, map-hygiene layers, back-projected labels, localization replay
Calibration packagecolorization, projection, back-projection, fusion labels, runtime map contract
Dynamic-removal / hygiene sidecarrelease-state masks, training positives/negatives, FOD/hazard labels
Semantic taxonomytraining labels, model heads, class-order consumers, metrics, runtime loaders
Reviewer decisionlabel state, QA report, release packet, safety-case evidence
Prompt/labeler/evaluatorcandidate labels, reviewer tasks, label QA, model cards
Training exportdataset snapshot, model training run, experiment, registry version
Model registry versionevaluation pack, release packet, deployment manifest, incident evidence

The data catalog should store this as lineage, not as prose. At minimum, each edge needs source ID, output ID, code version, parameter hash, timestamp, owner, and allowed-use state.


Batch Manifest Fields

FieldRequirement
pseudo_label_batch_idImmutable ID for the exported label batch
source_map_manifest_hashExact source map and pose graph used
semantic_layer_idSemantic map layer ID and digest
map_hygiene_layer_digestsDynamic, transient, FOD, artifact, and review layers consumed
taxonomy_id / class_id_mapVersioned semantic taxonomy and ordered class mapping
release_state_schemaAllowed release-state values and training permissions
calibration_idSensor/camera/extrinsic/timing calibration used for projection
projection_code_hashBack-projection/materialization code and parameters
split_idTrain/val/test/replay/local-holdout split membership, grouping keys, allowed use, and leakage checks
review_stateCandidate, reviewed, QA-passed, waived, rejected, or expired
allowed_useTraining positive, auxiliary task, evaluation, replay, review only, or blocked
invalidation_statusActive, suspect, quarantined, rebuilt, reapproved, deprecated, or restricted
upstream_artifact_idsSource maps, prompts, labelers, evaluators, cleaner configs, and QA reports
downstream_consumersDataset snapshots, training runs, registry versions, release packets, and replay packs

These fields can begin as JSON sidecars. At S3-S5 they should be data-catalog fields with queryable lineage and policy checks.


Scale Requirements

MLOps scaleMinimum invalidation control
S0 notebook researchNote the source map and projection script; do not reuse pseudo-labels outside the experiment without promotion
S1 repeatable prototypeBatch manifest, fixed split, and manual invalidation note when the source map or taxonomy changes
S2 single-product productionCataloged batch state, lineage to model registry version, release-ticket hold when a batch becomes suspect
S3 fleet and multi-siteSite/ODD impact query, active-learning rebuild queue, and local holdout protection for affected sites
S4 regulated safety-criticalEvidence freeze, safety-case impact review, retention hold, rollback assessment, and approver record
S5 platform scalePolicy-as-code invalidation propagation across catalog, registry, feature/embedding store, evaluator, and deployment services

At S2+, invalidation should fail closed: if impact cannot be proven, treat the batch as quarantined until rebuilt or waived.


Airside Rules

  • permanent_static is the only default positive state for base semantic class supervision.
  • movable_static, static_transient, fod_candidate, artifact, and unknown_review may train auxiliary heads, hard negatives, or review queues, but not permanent-static positives without explicit declaration.
  • FOD and personnel corrections are safety-critical invalidation triggers even when the corrected region is small.
  • Temporary overlays must expire or renew; labels exported from expired overlays become suspect by default.
  • Airport/site-specific taxonomy aliases must not propagate into global labels without taxonomy approval.
  • Back-projected labels from a corrected source map must not share split IDs with the old labels unless leakage checks are rerun.

  • mlops-scale-research-scope.md - scale ladder and backlog.
  • dataset-split-leakage-controls-by-scale.md - split manifests and leakage controls for map-derived labels, replay, local holdouts, and release evidence.
  • data-flywheel-airside.md - map-derived semantic label branch.
  • model-governance-release-evidence.md - release evidence and registry controls.
  • ../data-platform/data-catalog-lineage-quality-ops.md - catalog states, lineage, and data quality.
  • ../../30-autonomy-stack/perception/overview/aggregated-map-semantic-segmentation.md - semantic-map training export and release-state masks.
  • ../../30-autonomy-stack/localization-mapping/maps/airside-map-hygiene-ground-truth-protocol.md - release-state labels and QA fields.
  • ../ota/perception-slam-artifact-compatibility-matrix.md - artifact-set compatibility and activation gates.

Sources

Public research notes collected from public sources.