Skip to content

LLMOps and Agent Evaluation by Scale

Last updated: 2026-05-24

This page defines MLOps controls for LLM, VLM, VLA, RAG, model-as-judge, and tool-using agent systems in autonomous vehicle and managed-site operations. These systems may be advisory, offline, or back-office, but they can still affect labels, semantic maps, incident reports, operator guidance, evaluation decisions, safety-case prose, and release packets.

The operating rule is conservative: a generative model may propose, summarize, retrieve, rank, or route review, but it does not become release truth unless the downstream artifact passes the normal data, model, map, runtime, and safety gates.


Artifact Scope

ArtifactWhy it mattersRequired identity
Prompt packPrompts change behavior like codePrompt digest, owner, task, variables, examples
Model endpoint/checkpointProvider/model updates change outputsProvider, model ID, checkpoint/API version, region, retention mode
Decoding and safety configTemperature, top-p, filters, and abstention alter stabilityConfig digest and allowed-use scope
Retrieval corpusRAG answers depend on indexed contentCorpus snapshot, embedding model, index build, access tier
Tool policyAgents can read, write, or mutate systemsTool allowlist, permission scope, approval gate
Agent graphPlanner/router/tool sequence changes outcomesGraph version, state schema, memory policy
Evaluator/judge modelJudges can approve or reject candidatesJudge prompt/model/calibration set
Evaluation packMetrics define what "good" meansGolden set, rubrics, thresholds, slices
Trace schemaDebugging and audit require full pathInput/output digest, retrieved docs, tool calls, reviewer disposition
Guardrail policyBlocks unsafe, private, or unsupported actionsPolicy version, test set, exception owner

Treat any of these as release-affecting once output can influence training data, map publication, replay assertions, safety evidence, operations decisions, or deployment approval.


Scale Policy

MLOps scaleLLMOps postureRequired evidenceWhat to block
S0 notebook researchLocal prompt experiments and manual reviewPrompt text, model name, sample inputs/outputs, limitationsReusing generated outputs as labels or evidence
S1 repeatable prototypeVersioned prompt/eval notebookFrozen eval set, prompt digest, model/config record, failure examplesDemo prompt becoming hidden production workflow
S2 single-product productionPrompt/model/retrieval registry and release packetEval pack, trace capture, reviewer workflow, rollback prompt/modelPrompt/model update changing labels, reports, or tools without review
S3 fleet and multi-siteSite/ODD-aware prompt, retrieval, and evaluation slicesLocal terminology tests, site holdouts, corpus snapshots, canary tracesOne global prompt or corpus applied to every site
S4 regulated safety-criticalEvidence-locked generative outputs and human approvalSafety-case links, red-team/prompt-injection tests, abstention rules, immutable tracesGenerated prose or judge score replacing accountable approval
S5 platform scaleMulti-tenant GenAIOps platformRegistry, eval service, policy-as-code, telemetry, cost/SLOs, tenant isolationTeams using ungoverned agents, prompts, tools, or retrieval corpora

The most important transition is S2. Once a prompt, RAG system, VLM labeler, judge, or agent can affect a product artifact, it needs the same candidate/shadow/champion/rollback discipline as other MLOps artifacts.


Autonomy Use Cases

Use caseAllowed roleRelease boundary
Open-vocabulary auto-labelingCandidate labels, taxonomy suggestions, reviewer accelerationLabels need reviewer, taxonomy, QA, allowed-use, and provenance
Semantic-map tile reviewCandidate class names, anomaly flags, uncertainty triagePublished map needs source-map acceptance and map QA
Incident summarizationDraft timeline, evidence links, suspected factorsIncident record needs human owner and source-linked evidence
SOP/NOTAM/RAG assistantRetrieve and summarize operational rulesActive operations need cited, current, approved documents
Model-as-judge evaluationRoute review and compare candidatesJudge score cannot replace release approval in S4 contexts
Agentic data miningQuery fleet data, create candidate batches, open ticketsWrites need scoped tools, dry-run mode, and human approval
VLM scene co-pilotAdvisory scene description and risk cueVehicle control remains behind deterministic runtime assurance
VLA action proposalCandidate trajectory or task plan for simulation/reviewRuntime action requires planner, monitor, and safety gate

For managed-site autonomy, generative systems are most useful as offboard accelerators and reviewer aids. Direct control authority should remain outside the generative layer unless a separate safety architecture proves bounded behavior.


Architecture Comparison

PatternAdvantagesDisadvantagesBest fit
Prompt registry onlySimple, low overhead, enough for prototypesWeak eval and trace controlsS1 prototypes
Prompt plus eval packEnables regression testing and prompt comparisonDoes not control retrieval/tool driftS2 production prompts
RAG with frozen corpus snapshotsGrounded answers and reproducible retrievalCorpus/index invalidation burdenSOP, NOTAM, incident, map-doc assistants
VLM labeler with human reviewSpeeds rare-class and open-vocab discoveryProjection, prompt, and taxonomy riskCandidate labels and map-tile triage
Model-as-judgeScales qualitative comparisonBias, drift, and circular evaluationReview routing, not safety approval
Tool-using agentAutomates multi-step workflowsWrite actions, prompt injection, hidden state, cost runawayTicket creation, batch mining, report drafts
Agent platform serviceShared governance and telemetryPlatform bypass risk if too slowS5 multi-team GenAIOps

Prefer the smallest pattern that preserves evidence. A RAG assistant does not need an agent planner if retrieval plus a cited answer is enough. A tool-using agent should start in dry-run mode until traces prove it behaves within scope.


Evaluation Layers

LayerWhat to testAutonomy example
Task correctnessOutput matches task-specific ground truthCorrectly classify pushback tug, FOD, stand closure, worker, or construction barrier
Grounding and citationClaims are supported by source evidenceNOTAM answer cites the active closure and route segment
Spatial consistencyText agrees with metric geometry and tracks"Loader is clear of aircraft" matches 3D clearance
Retrieval qualityRight documents/clips/tiles are retrievedSimilar incident search returns the relevant site and map state
Tool correctnessTool calls use correct inputs and permissionsAgent opens a candidate-label ticket without mutating release labels
Prompt-injection resistanceMalicious or irrelevant context does not override policyRetrieved document cannot make the assistant approve a release
Abstention and uncertaintyModel refuses or routes review when evidence is weakLow-quality night image becomes unknown_review
RobustnessSensor, text, document, and context corruptionsPrompt typo, stale SOP, occluded worker, rain/fog image
RegressionPrompt/model/corpus update does not break slicesNew prompt improves apron scenes but not terminal frontage
Human review loadReviewer correction and disagreementAuto-labeler saves time without raising false acceptance
Cost and latencyToken, tool, retrieval, and runtime costAgent mining job does not starve release replay

Public benchmark scores are not sufficient. Every production GenAI artifact needs a task-specific eval pack and site/ODD slices that reflect the actual operating context.


Agent Evaluation Contract

Agent evaluation must inspect the trajectory, not only the final answer.

FieldRequired evidence
Task IDScenario, ticket, label batch, map tile, incident, or operator request
Input digestUser prompt, image/clip/map/document IDs, context window
Planning traceSteps proposed, tools selected, branch/loop decisions
Tool callsTool name, arguments, return value digest, permission result, error handling
Retrieval traceQuery, filters, top-k IDs, score, corpus/index version
Memory stateWhat short/long-term memory was read or written
OutputFinal text, structured JSON, labels, tickets, or proposed action
Reviewer dispositionAccepted, corrected, rejected, escalated, waived
Policy resultGuardrail decisions, blocked actions, exceptions, expiry
Cost/latencyTokens, tool time, retries, timeout, rate-limit behavior

For safety-relevant or release-affecting tasks, the agent should not have direct write permission to labels, semantic maps, registry aliases, deployment manifests, safety-case records, or incident closure. It can create candidate artifacts that a governed workflow reviews.


Release Gates

GatePass conditionBlocks
G0 inventoryPrompt/model/retrieval/tool/eval artifacts have IDs and ownersUnknown prompt or model endpoint
G1 reproducibilitySame inputs, model/config, corpus, and tools can reconstruct traceMissing prompt/corpus/tool version
G2 task evalEval pack passes aggregate and slice thresholdsPoor grounding, wrong labels, unsupported summaries
G3 security evalPrompt injection, tool misuse, data exfiltration, and unsafe output tests passOWASP LLM risk unresolved
G4 human reviewReviewer correction rate and disagreement are within boundsModel overloads reviewers or creates false confidence
G5 compatibilityDownstream taxonomy, map, data, runtime, or release packet accepts artifact stateCandidate output used as release truth
G6 rolloutShadow/canary traces show expected behavior under real workflowSite-specific prompt/corpus regression
G7 closureActive prompt/model/corpus/tool IDs are monitored and rollback existsNo rollback prompt/model/corpus bundle

At S4, the release gate must explicitly say which generated outputs are evidence, which are supporting context, and which are merely drafts.


Observability and Telemetry

GenAI telemetry should be structured enough for debugging, cost control, and audit:

SignalExamples
Trace spansModel call, retrieval call, tool call, planner step, evaluator step
MetricsLatency, tokens, cost, tool error rate, refusal rate, unsupported-claim rate
Logs/eventsPrompt/model IDs, corpus/index version, policy result, reviewer correction
Safety countersPrompt-injection hits, blocked tool calls, private-data suppression, hallucination reports
Release linkageArtifact IDs, release packet, ODD/site scope, rollback bundle

OpenTelemetry GenAI semantic conventions are still evolving, so record the convention version emitted by instrumentation. Do not depend on free-text chat logs as the system of record for release evidence.


Airside and Non-Road Managed-Site Rules

  • Site terminology belongs in versioned prompt packs and retrieval corpora: stand, apron, tug, ULD, belt loader, quay, aisle, bay, mine bench, utility cabinet, terminal frontage.
  • VLM labels over images or maps should enter candidate_label or unknown_review, not release_label.
  • RAG assistants must cite active documents and show corpus freshness; stale SOP, NOTAM, work-order, or map-overlay data can be worse than no answer.
  • Agent tools that open tickets, label tasks, or map hygiene reviews should write only candidate records until human approval.
  • Generated incident reports must preserve evidence IDs, raw clips, timestamps, active artifact IDs, and reviewer decisions.
  • A VLA action proposal is a simulation or advisory artifact unless deterministic planning, runtime monitoring, and safety-case evidence accept it.

Failure Modes

Failure modeConsequenceControl
Prompt update without evalLabels, reports, or advice drift silentlyPrompt registry and eval pack
RAG corpus changes invisiblyAnswers cannot be reproducedCorpus/index snapshot and retrieval trace
Judge model approves its own familyInflated eval scoresHuman calibration set and judge/model separation
Agent writes to production systemsBad labels, tickets, maps, or releases propagateTool scopes, dry-run, human approval
Prompt injection through retrieved docsTool misuse or false approvalRetrieval sanitization, policy checks, injection tests
Missing abstentionModel fabricates on weak evidenceUnknown/review states and abstention thresholds
Trace gapsIncident cannot be reconstructedStructured trace schema and retention
Site terminology mismatchWrong object or rule interpretationSite-specific prompt and eval slices
Cost runawayAgent loops or starves GPU/eval capacityToken/tool budgets, timeouts, queue policy
Generated prose becomes evidenceSafety case rests on unsupported claimsSource-linked evidence and human sign-off

  • mlops-scale-research-scope.md - MLOps scale ladder and foundation-model operations.
  • offboard-labeler-registry-by-scale.md - labeler, prompt, evaluator, retrieval, and reviewer workflow registry.
  • feature-embedding-store-ops-by-scale.md - retrieval corpus and vector-index controls.
  • secure-artifact-attestation-profile.md - prompt/eval/tool artifact signing and provenance.
  • model-governance-release-evidence.md - release packets and rollback evidence.
  • data-flywheel-airside.md - auto-labeling, active learning, and scenario mining.
  • ../../30-autonomy-stack/vla-vlm/vlm-scene-understanding.md - VLM scene-understanding use cases and reliability notes.
  • ../../30-autonomy-stack/vla-vlm/vlm-vla-reliability-benchmarks.md - VLM/VLA benchmark design.

Sources

Public research notes collected from public sources.