CLI Evaluation Input Compilation

Status: implemented production input boundary under the Evaluation Core vNext RFC. Parse, resolve, compile, Runtime assembly, host workflow, Core execution, and artifact persistence now form the authoritative omk eval path.

1. Purpose and boundary

The CLI host must translate flags, eval.yaml, files, and local resources into the same host-neutral measurement contract. It must not build a second plan beside Core's sealed RunPlan. Input handling therefore has five distinct responsibilities:

CLI flags / eval.yaml
        │ parse precedence and syntax
        ▼
CliEvaluationRequest
        │ resolve files, artifacts, samples and host resources
        ▼
ResolvedCliEvaluationInput + ResolvedHostResources
        │ pure deterministic compilation
        ▼
EvaluationDefinition + MeasurementPolicy
RuntimeBindingRequest + EvaluationOrchestrationOptions
EvaluationPresentationOptions + static RunOptions metadata
        │
        ├── Core prepare: schema, references, capabilities, runtime identity, sealing
        └── adapter preflight: doctor, credentials, connectivity, physical environment

Phase	Owns	Must not do
Parse	parseCliEvaluationRequest() applies CLI > eval.yaml > explicit host default, validates syntax, and records source attribution	Read files/environment/network/clock, accept parser-injected defaults as explicit flags, create a Runtime, calculate measurement digests
Resolve	Materialization, git pinning, content/tree digests, mock and workspace descriptors, locator binding	Call Core prepare, trust self-reported capabilities, emit a sealed plan
Compile	Pure mapping to canonical, schema-valid, statically coherent Core contracts and host-side requests	Read filesystem/environment/network/clock, create AbortSignal, writer, store, or run ID
Core prepare	Revalidate the contract, qualify verified Runtime capability/identity, and seal the only RunPlan	Perform connectivity or doctor checks
Adapter preflight	Credentials, connectivity, path permissions, doctor, physical health	Replace Core capability validation or rewrite sealed measurement design

2. Output ownership

• EvaluationDefinition owns data projections, Target behavior, evaluator instruments, metrics, experiment design, analysis, comparisons, and decision policy.
• MeasurementPolicy owns execution/evaluation concurrency, timeout, retry, cache, evidence, failure, event delivery, and the shared Run budget ledger.
• RuntimeBindingRequest v4 contains only implementation and resource-lease requirements derived from Definition／resolved host resources. Executor qualification reuses the exact canonical TargetDefinition.executionRequirements; it does not maintain a second approximation. A registry may resolve bindings, but cannot override execution requirements, model, effort, prompt variant, protocol, evaluator identity, or behavior config. Its complete assembly contract is specified in Evaluation Runtime Adapter.
• ResolvedHostResources binds a stable resource ID and digest to an effect locator. It is not a Core schema and never enters canonical measurement JSON.
• EvaluationOrchestrationOptions owns dry-run, resume locator, batch, independent Series repeats, preflight switches, diagnostic post-processing, gold post-hoc workflows, and managed-evidence append behavior.
• Sample-bundle requires is normalized as host-only dependencyRequirements, together with its baseDirectoryLocator, for the later doctor/preflight workflow. Relative files and preflight commands therefore keep the sample bundle's resolution semantics. This host context does not enter Core measurement digests and is never silently discarded.
• EvaluationPresentationOptions owns output locator, index scope, language, server, verbosity, layered view, and CLI exit presentation. None changes DecisionResult.
• Static RunOptions metadata may contain serializable annotations and summaries. The orchestrator creates run ID, cancellation, event writer, and buffers only when a Run actually starts.

HostExecutionPlan is intentionally absent. The only object named Plan is produced and sealed by Core prepare.

3. Identity, lineage, and resource safety

Behavior identity and source lineage are separate axes:

• Target config contains Target-wide behavior facts such as artifact, mock, sandbox, model, and effort. Canonical executionControls separately holds Target defaults and sample overrides for workspace and tool authority; its workspace descriptors contain no locator.
• Target executionRequirements is the aggregate capability request derived from resolved behavior and every effective sample control: explicit system-instruction use, copy-on-write workspace, native MCP config, pre-tool-call mock interception, tool allow-list, skill-discovery policy, and sandbox ID. It enters Definition and ExecutionPlan identity; only Core prepare may compare it with Runtime features. It never grants the aggregate workspace or tool authority to one Trial.
• Host resources contain locator, resolved commit, repository origin, and materialization evidence. Moving the same content between absolute/relative paths or machines does not invalidate execution identity.
• A behavior change changes the Definition digest. A lineage-only change is assessed later by explicit comparability and provenance policy, not smuggled into Target config.

Mock rules and payloads are separate secret, digest-bound descriptors. Raw tool／match values never enter Core or static Target JSON; the Runtime adapter reads them only through the run-scoped verified lease before any business process starts. sampleIds, strict mode, the rule descriptor, and ordered payload descriptors remain canonical Target behavior, so changing rule bytes or return order changes measurement identity. sampleId is supplied by Core to the adapter and is never appended to the model prompt. Compile also requires every reference role to match its host resource kind: artifact, workspace, MCP config, mock rule, mock payload, evaluator content, and gold dataset cannot be substituted for one another even when a descriptor happens to match. Runtime adapters must verify the digest immediately before use. Missing interception, allowed-tool, skill-discovery, MCP, cancellation, seed, or sandbox capability fails closed during Core prepare; adapters must never drop mocks or fall back to real external calls. Heterogeneous sample cwd and allowedTools values compile into canonical sample overrides; they are never unioned, and adapters receive only the exact effective Trial control.

ResolvedHostResources v3 makes descriptor.size mandatory and represents pinned Git verification as {verificationKind, verifiedDigest, commitId}. The commit ID is a normalized 40–64 character lowercase hexadecimal object identity; a branch or tag name is not a pin. File-only MCP config, mock rule, mock payload, and evaluator-content resources require content-digest; mock rules additionally require application/json. MCP config and both mock-control resource kinds must be classified secret. Workspaces require tree-digest or pinned-git; pinned Git is limited to artifacts and workspaces. gold classification and gold-dataset kind are mutually required. Earlier shapes are rejected without a compatibility reader.

Dataset projections preserve the Gold boundary: Executors see only input + executionContext; evaluators may receive expected + evaluationContext; analysis receives only explicit membership and analysis context. Gold locators remain host resources. Post-hoc gold comparison is exploratory and cannot masquerade as a preregistered decision.

4. Measurement mapping

• Control/treatment is represented only by Comparison; artifact contents are Target behavior.
• Assertions, rubric judges, dimensions, composites, and RAG metrics remain distinct Evaluator/Metric/AnalysisGraph concepts. An Evaluator template owns its algorithm implementationId, instrument, and runtime prompt variant; a judge member owns only provider executor, model, effort, and ensemble identity.
• Judge members carry explicit instrument, ensemble member, replicate group, and replicate index. No analysis may parse evaluator IDs to infer hierarchy.
• Holdout/cohort membership is analysis-only and fixed before execution.
• Bootstrap, correction, thresholds, and trivial-difference gates are Definition facts in AnalysisGraph or DecisionPolicy.
• Every treatment has its own explicit paired control comparison; multi-treatment requests never collapse into one ambiguous treatment identifier.
• Deterministic assertion evaluators emit explicit structural-missing observations when a criterion does not apply. LLM assertions and rubric dimensions instead seal canonical applicableSampleIds; Core omits non-applicable Evaluator coordinates from plan identity, coverage, execution, and analysis rather than scoring an unintended sample.
• The production design uses paired blocks with seedCoupling: uncontrolled: current provider adapters expose no exact sampling-seed control. The host may deterministically randomize coordinate order, but it must not claim coupled model randomness.
• Length debias changes rubric evaluator config; presentation/tone neutralization remains always on.
• Legacy total USD maps to a shared Run provider-cost limit. Per-sample USD and milliseconds map to per-coordinate provider-cost and active-duration limits. They are not implemented by a host AbortController or report rewrite.

The repetition hierarchy is explicit:

Concept	Contract
Trial	Repeated Target measurement inside one sealed Run
Retry attempt	Infrastructure recovery for one coordinate
Judge replicate	Repeated observation by the same instrument
Ensemble member	A distinct judge instrument/Runtime
--repeat	Independent Runs in an Evaluation Series
Batch child	A different artifact workflow, not a Series replicate

For --repeat > 1, the orchestrator allocates a unique seriesInstanceId before resolution. Compilation binds that host-owned instance identity to the complete Series design—measurement design, repeat count, comparison scope, and minimum status—to derive Core's final seriesId. The compiler never invents identity from a clock or random source, and reusing an instance ID with a different design cannot alias the same Series.

4.1 Cache and replay

Fresh measurement is the default. The normalized policy always carries two independent fields:

cache:
  executionMode: disabled # disabled | replay-only | transparent-deterministic
  evaluationMode: disabled # disabled | reuse

executionMode and evaluationMode enter their respective Core contract digests. A non-disabled mode also requires a stage-specific source locator in host-owned orchestration.cacheSources; the locator never enters Core canonical JSON. The Runtime adapter must assemble exactly that source into the corresponding cache port and must not substitute an environment-selected or global default.

replay-only is a fail-closed read path. A missing, unavailable, corrupt, or identity-mismatched coordinate terminates the run and never falls back to a live Target call. Replayed records retain the original trial identity, Runtime identity, usage, cost, and provenance; they add neither a native invocation nor an independent replicate. Until Series has an explicit effective-independent-sample model, compilation rejects every non-disabled cache mode together with an independent Series repeat. It also rejects mixing cache reuse with resume in one request. transparent-deterministic is available only when Core prepare verifies deterministic execution and a verified Runtime identity. Evaluation reuse is independent and remains bound to the complete evaluation contract, including evaluator/model/prompt variant, replicate identity, Gold-facing inputs, metrics, and evidence policy.

The current product CLI disables both Execution and Evaluation cache. The existing resume contract is unchanged. Core and Runtime retain their independent cache contracts; product cache reuse is not exposed through CLI flags or eval config.

5. Determinism and validation

parseCliEvaluationRequest() is the pure normalization boundary for raw CLI/config inputs. The host passes only flags explicitly supplied by the user, an already syntax-validated EvalConfig, and any environment-selected defaults as explicit values with provenance. CLI and config candidates pass through the same canonical field validators before precedence is applied. Provenance is emitted only for a value that exists at this stage; later-derived values acquire provenance only when they are actually derived. Oclif-injected defaults must not be passed as explicit CLI values. Judge disablement is resolved before judge-model parsing, so an unused malformed judge source cannot fail a no-judge request.

compileCliEvaluationInput() accepts only resolved, serializable IR. It performs no I/O and returns deeply frozen output. It uses Core's public normalizeEvaluationDefinition() and validateDefinitionSemantics() contracts before emitting output: schema validation, canonical set-like ordering, reference integrity, and static semantic checks therefore have one owner. The Core boundary revalidates those contracts before runtime qualification. Object property order, nested membership/cohort-filter order, host locator spelling, CLI/YAML source, and lineage do not affect Definition or Policy digests; actual behavior bytes do.

Parse and compilation errors are host CliEvaluationInputError values with stable codes and optional source/field paths. Core static-validation failures retain their Core code/details inside the host error without leaking Core Run errors across the pre-Run boundary. They are not Core EvaluationError records, because a Run has not started. Runtime capability claims are not accepted by compile; only Core prepare can qualify them.

6. Migration boundary

This layer is the production boundary. omk eval consumes its contracts through Runtime assembly and the Core host workflow, then persists the Core Plan, Bundles, and Report. The deleted legacy pipeline is neither double-run nor shadow-run, and no later layer reparses CLI input.

The current boundary emits omk.cli-evaluation-request/v3, omk.resolved-cli-evaluation-input/v6, omk.resolved-host-resources/v3, and omk.runtime-binding-request/v5. Request v3, resolved input v6, and binding request v5 carry the optional host-declared judge deployment revision from eval.yaml into evaluator Runtime qualification. Earlier request shapes are not accepted.

--no-cache and the noCache config field have been removed because they no longer changed product behavior. Remove them from commands and eval.yaml; no replacement option is needed. Existing Core cache contracts and current measurement policy digests remain unchanged. Old cache files are not read.

7. Exhaustive input registry

The declarative registry classifies every live omk eval flag and every machine-enumerable EvalConfig path. CI compares both source-key sets strictly, so an unclassified new field fails. The table is generated from the registry by yarn build:docs.

Source	Field	Normalized field	Priority	Normalized default (source)	Owner	Digest stage	Runtime qualification	Error / migration
CLI	--batch	orchestration.batch	300	false (documented)	Orchestration	none	—	CLI_INPUT_INVALID retain
CLI	--bootstrap	definition.analysisGraph.bootstrap	300	true (documented)	Definition	analysis	—	CLI_INPUT_INVALID retain
CLI	--bootstrap-samples	definition.analysisGraph.bootstrap.resamples	300	1000 (documented)	Definition	analysis	—	CLI_INPUT_INVALID retain
CLI	--budget-per-sample-ms	policy.budget.perCoordinateActiveDurationMs	300	—	MeasurementPolicy	run	—	CLI_INPUT_INVALID retain
CLI	--budget-per-sample-usd	policy.budget.perCoordinateProviderCostUSD	300	—	MeasurementPolicy	run	—	CLI_INPUT_INVALID retain
CLI	--budget-usd	policy.budget.totalProviderCostUSD	300	—	MeasurementPolicy	run	—	CLI_INPUT_INVALID retain
CLI	--concurrency	policy.executionConcurrency	300	1 (documented)	MeasurementPolicy	execution	—	CLI_INPUT_INVALID retain
CLI	--config	orchestration.configLocator	300	—	Orchestration	none	—	CLI_INPUT_INVALID retain
CLI	--control	definition.targets.control	300	—	Definition	execution	—	CLI_INPUT_INVALID retain
CLI	--control-cwd	resources.controlWorkspaceLocator	300	—	Orchestration	none	—	CLI_INPUT_INVALID retain
CLI	--dry-run	orchestration.dryRun	300	false (documented)	Orchestration	none	—	CLI_INPUT_INVALID retain
CLI	--effort	definition.targetRuntime.effort	300	"low" (documented)	Definition	execution	model-effort	CLI_INPUT_INVALID retain
CLI	--executor	definition.targetRuntime.implementationId	300	— (environment-selection)	Definition	execution	executor-protocol model-effort	CLI_INPUT_INVALID retain
CLI	--global	presentation.indexScope	300	"project" (documented)	Presentation	none	—	CLI_INPUT_INVALID retain
CLI	--gold-dir	orchestration.gold.resourceLocator	300	—	Orchestration	none	—	CLI_INPUT_INVALID retain
CLI	--holdout-ratio	definition.dataset.analysisCohorts	300	—	Definition	analysis	—	CLI_INPUT_INVALID retain
CLI	--judge-models	definition.judges.members	300	— (environment-selection)	Definition	evaluation	evaluator-instrument model-effort	CLI_INPUT_INVALID retain
CLI	--judge-repeat	definition.judges.replicateCount	300	1 (documented)	Definition	evaluation	—	CLI_INPUT_INVALID retain
CLI	--lang	presentation.language	300	"zh" (environment-selection)	Presentation	none	—	CLI_INPUT_INVALID retain
CLI	--layered-stats	presentation.layeredView	300	false (documented)	Presentation	none	—	CLI_INPUT_INVALID retain
CLI	--mcp-config	resources.mcpConfigLocator	300	—	Orchestration	none	tool-mock-sandbox	CLI_INPUT_INVALID retain
CLI	--model	definition.targetRuntime.model	300	— (environment-selection)	Definition	execution	model-effort	CLI_INPUT_INVALID retain
CLI	--no-debias-length	definition.judges.lengthDebias	300	true (documented)	Definition	evaluation	—	CLI_INPUT_INVALID retain
CLI	--no-diagnostic	orchestration.diagnostic	300	"enabled-outside-core" (documented)	Orchestration	none	—	CLI_INPUT_INVALID retain
CLI	--no-evidence	orchestration.managedEvidence	300	"append" (documented)	Orchestration	none	—	CLI_INPUT_INVALID retain
CLI	--no-gate	presentation.exitMode	300	"gate" (documented)	Presentation	none	—	CLI_INPUT_INVALID rename → --report-only
CLI	--no-judge	definition.judges.enabled	300	true (documented)	Definition	evaluation	—	CLI_INPUT_INVALID retain
CLI	--no-serve	presentation.serve	300	true (documented)	Presentation	none	—	CLI_INPUT_INVALID retain
CLI	--no-strict-baseline	definition.baselineIsolation	300	true (documented)	Definition	execution	—	CLI_INPUT_INVALID CLI_INPUT_BASELINE_ISOLATION_CONFLICT retain
CLI	--output-dir	presentation.outputDirectoryLocator	300	—	Presentation	none	—	CLI_INPUT_INVALID retain
CLI	--repeat	orchestration.independentSeries.repeatCount	300	1 (documented)	Orchestration	run	—	CLI_INPUT_INVALID retain
CLI	--report-only	presentation.exitMode	300	"gate" (documented)	Presentation	none	—	CLI_INPUT_INVALID retain
CLI	--resume	orchestration.resumeSourceLocator	300	—	Orchestration	none	—	CLI_INPUT_INVALID retain
CLI	--retry	policy.retryCount	300	0 (documented)	MeasurementPolicy	execution	—	CLI_INPUT_INVALID retain
CLI	--samples	orchestration.samplesLocator	300	—	Orchestration	none	—	CLI_INPUT_INVALID retain
CLI	--skill-dir	orchestration.skillDirectoryLocator	300	"skills" (documented)	Orchestration	none	—	CLI_INPUT_INVALID retain
CLI	--skip-connectivity	orchestration.preflight.connectivity	300	"required" (documented)	Orchestration	none	—	CLI_INPUT_INVALID retain
CLI	--skip-doctor	orchestration.preflight.doctor	300	"required" (documented)	Orchestration	none	—	CLI_INPUT_INVALID retain
CLI	--strict-baseline	definition.baselineIsolation	300	true (documented)	Definition	execution	—	CLI_INPUT_INVALID CLI_INPUT_BASELINE_ISOLATION_CONFLICT retain
CLI	--threshold	definition.decisionPolicy.threshold	300	— (derived)	Definition	decision	—	CLI_INPUT_INVALID retain
CLI	--timeout	policy.executionTimeoutMs	300	120000 (documented)	MeasurementPolicy	execution	—	CLI_INPUT_INVALID retain
CLI	--treatment	definition.targets.treatments	300	—	Definition	execution	—	CLI_INPUT_INVALID retain
CLI	--treatment-cwd	resources.treatmentWorkspaceLocators	300	—	Orchestration	none	—	CLI_INPUT_INVALID retain
CLI	--trivial-diff	definition.decisionPolicy.trivialDifference	300	— (derived)	Definition	decision	—	CLI_INPUT_INVALID retain
CLI	--verbose	presentation.verbose	300	false (documented)	Presentation	none	—	CLI_INPUT_INVALID retain
eval.yaml	bootstrap	definition.analysisGraph.bootstrap	200	true (documented)	Definition	analysis	—	CLI_INPUT_INVALID retain
eval.yaml	bootstrapSamples	definition.analysisGraph.bootstrap.resamples	200	1000 (documented)	Definition	analysis	—	CLI_INPUT_INVALID retain
eval.yaml	budget	policy.budget	200	—	MeasurementPolicy	run	—	CLI_INPUT_INVALID retain
eval.yaml	budget.perSampleMs	policy.budget.perCoordinateActiveDurationMs	200	—	MeasurementPolicy	run	—	CLI_INPUT_INVALID retain
eval.yaml	budget.perSampleUSD	policy.budget.perCoordinateProviderCostUSD	200	—	MeasurementPolicy	run	—	CLI_INPUT_INVALID retain
eval.yaml	budget.totalUSD	policy.budget.totalProviderCostUSD	200	—	MeasurementPolicy	run	—	CLI_INPUT_INVALID retain
eval.yaml	concurrency	policy.executionConcurrency	200	1 (documented)	MeasurementPolicy	execution	—	CLI_INPUT_INVALID retain
eval.yaml	decision	definition.decisionPolicy	200	—	Definition	decision	—	CLI_INPUT_INVALID retain
eval.yaml	decision.minimumComparisonUnits	definition.decisionPolicy.sampleSize.minimumComparisonUnits	200	20 (documented)	Definition	decision	—	CLI_INPUT_INVALID retain
eval.yaml	decision.power	definition.decisionPolicy.sampleSize	200	—	Definition	decision	—	CLI_INPUT_INVALID retain
eval.yaml	decision.power.assumptionSource	definition.decisionPolicy.sampleSize.assumptionSource	200	—	Definition	decision	—	CLI_INPUT_INVALID retain
eval.yaml	decision.power.expectedDifferenceStandardDeviation	definition.decisionPolicy.sampleSize.expectedDifferenceStandardDeviation	200	—	Definition	decision	—	CLI_INPUT_INVALID retain
eval.yaml	decision.power.minimumDetectableDifference	definition.decisionPolicy.sampleSize.minimumDetectableDifference	200	—	Definition	decision	—	CLI_INPUT_INVALID retain
eval.yaml	decision.power.targetPower	definition.decisionPolicy.sampleSize.targetPower	200	0.8 (documented)	Definition	decision	—	CLI_INPUT_INVALID retain
eval.yaml	decision.threshold	definition.decisionPolicy.threshold	200	— (derived)	Definition	decision	—	CLI_INPUT_INVALID retain
eval.yaml	decision.trivialDifference	definition.decisionPolicy.trivialDifference	200	— (derived)	Definition	decision	—	CLI_INPUT_INVALID retain
eval.yaml	effort	definition.targetRuntime.effort	200	"low" (documented)	Definition	execution	model-effort	CLI_INPUT_INVALID retain
eval.yaml	executor	definition.targetRuntime.implementationId	200	— (environment-selection)	Definition	execution	executor-protocol model-effort	CLI_INPUT_INVALID retain
eval.yaml	goldDir	orchestration.gold.resourceLocator	200	—	Orchestration	none	—	CLI_INPUT_INVALID retain
eval.yaml	holdoutRatio	definition.dataset.analysisCohorts	200	—	Definition	analysis	—	CLI_INPUT_INVALID retain
eval.yaml	judgeModels	definition.judges.members	200	— (environment-selection)	Definition	evaluation	evaluator-instrument	CLI_INPUT_INVALID retain
eval.yaml	judgeModels[].deploymentRevision	definition.judges.members[].deploymentRevision	200	—	Definition	evaluation	model-effort	CLI_INPUT_INVALID retain
eval.yaml	judgeModels[].executor	definition.judges.members[].executorId	200	—	Definition	evaluation	evaluator-instrument	CLI_INPUT_INVALID retain
eval.yaml	judgeModels[].model	definition.judges.members[].model	200	—	Definition	evaluation	model-effort	CLI_INPUT_INVALID retain
eval.yaml	judgeRepeat	definition.judges.replicateCount	200	1 (documented)	Definition	evaluation	—	CLI_INPUT_INVALID retain
eval.yaml	lengthDebias	definition.judges.lengthDebias	200	true (documented)	Definition	evaluation	—	CLI_INPUT_INVALID retain
eval.yaml	mcpConfig	resources.mcpConfigLocator	200	—	Orchestration	none	tool-mock-sandbox	CLI_INPUT_INVALID retain
eval.yaml	model	definition.targetRuntime.model	200	— (environment-selection)	Definition	execution	model-effort	CLI_INPUT_INVALID retain
eval.yaml	noDiagnostic	orchestration.diagnostic	200	"enabled-outside-core" (documented)	Orchestration	none	—	CLI_INPUT_INVALID retain
eval.yaml	noJudge	definition.judges.enabled	200	true (documented)	Definition	evaluation	—	CLI_INPUT_INVALID retain
eval.yaml	repeat	orchestration.independentSeries.repeatCount	200	1 (documented)	Orchestration	run	—	CLI_INPUT_INVALID retain
eval.yaml	samples	orchestration.samplesLocator	200	—	Orchestration	none	—	CLI_INPUT_INVALID retain
eval.yaml	skipDoctor	orchestration.preflight.doctor	200	"required" (documented)	Orchestration	none	—	CLI_INPUT_INVALID retain
eval.yaml	strictBaseline	definition.baselineIsolation	200	true (documented)	Definition	execution	—	CLI_INPUT_INVALID retain
eval.yaml	timeoutMs	policy.executionTimeoutMs	200	120000 (documented)	MeasurementPolicy	execution	—	CLI_INPUT_INVALID retain
eval.yaml	variants	definition.targets	200	—	Definition	execution	—	CLI_INPUT_INVALID retain
eval.yaml	variants[].allowedSkills	definition.targets[].behavior.allowedSkills	200	—	Definition	execution	—	CLI_INPUT_INVALID retain
eval.yaml	variants[].artifact	resources.targets[].artifactLocator	200	—	Orchestration	none	—	CLI_INPUT_INVALID retain
eval.yaml	variants[].cwd	resources.targets[].workspaceLocator	200	—	Orchestration	none	—	CLI_INPUT_INVALID retain
eval.yaml	variants[].git	resources.targets[].gitSource	200	—	Orchestration	none	—	CLI_INPUT_INVALID retain
eval.yaml	variants[].git.ref	resources.targets[].gitSource.ref	200	—	Orchestration	none	—	CLI_INPUT_INVALID retain
eval.yaml	variants[].git.spec	resources.targets[].gitSource.spec	200	—	Orchestration	none	—	CLI_INPUT_INVALID retain
eval.yaml	variants[].git.url	resources.targets[].gitSource.url	200	—	Orchestration	none	—	CLI_INPUT_INVALID retain
eval.yaml	variants[].name	definition.targets[].targetId	200	—	Definition	execution	—	CLI_INPUT_INVALID retain
eval.yaml	variants[].role	definition.targets[].experimentRole	200	—	Definition	execution	—	CLI_INPUT_INVALID retain