@zhuoyuezs/ml-platform 0.1.8 → 0.1.10

package/skills/feature-management/references/operator-authoring.md

@@ -1,175 +0,0 @@
-# Operator Authoring
-
-Read this file whenever a workflow creates or changes executable Operator code.
-
-## Contents
-
-1. Package contract
-2. Feature entrypoint contract
-3. Causal calculation rules
-4. Required tests
-5. Build and publication
-6. Security boundary
-
-## Package Contract
-
-Build a standalone pure-Python wheel. Do not import report-local, experiment-local, or unregistered business packages. Declare runtime dependencies explicitly, but keep the package small and compatible with `py3-none-any`.
-
-Minimal `pyproject.toml`:
-
-```toml
-[build-system]
-requires = ["hatchling"]
-build-backend = "hatchling.build"
-
-[project]
-name = "pressure-features"
-version = "1.0.0"
-requires-python = ">=3.10"
-dependencies = ["pandas>=2.3.3"]
-
-[tool.hatch.build.targets.wheel]
-packages = ["src/pressure_features"]
-```
-
-Keep package version and Operator Registry version independently explicit. Change both when code behavior changes unless a documented release policy maps them differently.
-
-## Feature Entrypoint Contract
-
-Implement:
-
-```python
-def compute_features(context: Any) -> pandas.DataFrame:
-    ...
-```
-
-Available context fields:
-
-```text
-manifest
-features
-inputs
-config
-requested_output_columns
-operator
-parameters
-metric_frames
-target_times
-prediction_horizon
-cutoff_times
-furnace_id
-computation_hash
-```
-
-Each `metric_frames[parameter_key]` is a standard long frame with source timestamps and values. Resolve frames from `context.inputs`; do not reach into a database, API, filesystem secret, or environment credential.
-
-Use `context.cutoff_times` for causal feature windows. It is derived once from
-`context.target_times - context.prediction_horizon` using the DatasetManifest
-prediction contract; do not parse a duplicated Feature-level horizon.
-
-An Operator requires an explicit forecast contract whenever its formula or code
-uses `context.cutoff_times`, or any `input_schema.history_requirements` entry uses
-`anchor: cutoff`. In that case always declare `input_schema.prediction` with
-`required`, `minimum_horizon`, and `maximum_horizon`. This includes a strict
-nowcast: declare `required=true` and both bounds as `0min`; do not omit the
-section and silently accept the runtime's `default_nowcast` fallback. Declare
-source history as `input_schema.history_requirements`, using `anchor: cutoff`
-for causal windows and `anchor: target_time` only for intentional target-aligned
-formulas. Scope a requirement with `output_columns` when only some outputs need
-it. Include alignment slack such as hourly floor boundaries in the declared
-lookback.
-
-For realtime reads, `history_requirements` is also the per-Parameter fetch
-contract. Do not rely on a package-local lookback map or a single fixed window
-shared by all inputs. The realtime runner widens a half-open adapter read by one
-grid step, then applies `timestamp <= cutoff`; an Operator must never consume a
-post-cutoff row.
-
-Return requirements:
-
-- Return a pandas DataFrame.
-- Return exactly one row per `context.target_times` entry.
-- Return `event_time` in the same order and with the same timestamps.
-- Optionally return `furnace_id`.
-- Return every requested physical output column.
-- Avoid calculating unrequested expensive columns when practical.
-- Reject unknown requested output columns.
-- Preserve numeric missing values rather than silently filling them without a declared rule.
-
-One Operator may return multiple physical columns. Each public Feature still binds one `output_column` and exact inputs.
-
-## Causal Calculation Rules
-
-State every time rule in tests and descriptions:
-
-- forecast horizon and cutoff derivation;
-- whether an event exactly at cutoff is included (`<=`) or excluded (`<`);
-- rolling window left/right closure;
-- timezone and daylight-saving assumptions;
-- duplicate timestamp resolution;
-- sparse-event fallback behavior;
-- rounding stage and decimal count;
-- behavior before and after dated business-rule changes.
-
-Never infer a rule from a Feature name alone. Require a business decision when a boundary is unspecified.
-
-## Required Tests
-
-Use deterministic fixtures and cover each applicable item with a separately
-named test. A broad formula or happy-path test does not substitute for the
-cutoff-before, cutoff-at, cutoff-after, duplicate/missing, empty-history,
-requested-output, event-time-order, or dtype/rounding cases below:
-
-1. expected formula values;
-2. the event immediately before cutoff;
-3. an event exactly at cutoff;
-4. an event immediately after cutoff;
-5. missing and duplicate inputs;
-6. empty history or insufficient lookback;
-7. requested output subsets;
-8. exact `event_time` row count and order;
-9. output dtype and rounding;
-10. every dated business-rule branch.
-
-Run package tests before building the wheel. Then inspect the wheel tag and ensure it is `py3-none-any`.
-
-## Build And Publication
-
-Build into the catalog path referenced by `catalog.json`:
-
-```bash
-uv build <catalog>/operator_package \
-  --wheel \
-  --out-dir <catalog>/operator_package/dist
-```
-
-Keep the draft Operator fields null:
-
-```json
-{
-  "code_hash": null,
-  "package_uri": null,
-  "code_artifact": null
-}
-```
-
-Catalog publication uploads the wheel, computes SHA-256 and size, records its immutable object URI, and registers the resolved OperatorSpec.
-
-Do not reuse an Operator version with a different wheel. If the same key already exists with different content or digest, create a new version.
-
-## Security Boundary
-
-The current runtime blocks socket networking, removes storage/source credentials from the child environment, verifies the wheel hash, enforces a timeout, and extracts only pure-Python wheels without native libraries or unsafe paths.
-
-This is not a complete OS sandbox. Therefore:
-
-- publish only reviewed internal code;
-- do not read arbitrary host paths;
-- do not spawn subprocesses;
-- do not access secrets or environment configuration;
-- do not use dynamic code evaluation;
-- do not add native libraries;
-- do not perform network calls;
-- keep resource limits conservative.
-
-Request an engineering/security review when the formula requires capabilities outside this boundary.

package/skills/feature-management/references/platform-capability-guide.md

@@ -1,75 +0,0 @@
-# Platform Capability Guide
-
-Use this guide to explain what the platform can execute. It is not business
-evidence. A platform default never answers an unresolved business question.
-
-## Platform In One Paragraph
-
-The platform is a versioned data-contract, registry, execution, and delivery
-layer for algorithm projects. It turns confirmed source Parameters and approved
-deterministic Operators into ordered Features and reproducible DatasetManifests,
-then either materializes a batch DatasetArtifact or serves one causal realtime
-read. It owns source adapters, quality/missingness/freshness evidence, lineage,
-replay metadata, and data Job execution. This Skill covers that data subsystem;
-use the separate model lifecycle Skill for training, evaluation and packaging.
-Model serving, source-table creation and Kubernetes administration are outside
-this Skill's scope. Do not present a Skill boundary as a platform capability gap.
-
-The following distinctions are part of the platform contract:
-
-- A `Parameter` is a public source-data contract, not an arbitrary raw table or
-  a model feature.
-- An `Operator` is versioned executable code; a `Feature` is exactly one output
-  column; a `FeatureSet` is the consumer-visible column order.
-- A `DatasetManifest` is a versioned request for time range, read policy,
-  rowsets, missing/endpoint rules, prediction, and one FeatureSet.
-- A `DatasetArtifact` is an output of an authorized build, not evidence that the
-  source contract or business formula was semantically approved.
-- Realtime inference is a causal cutoff read. It must not run a batch build,
-  write Parquet, or publish a DatasetArtifact.
-
-Dry-run and registry resolution are structural evidence. They may not detect a
-missing source relation; source-backed success must be established by the build
-or realtime fetch response. Artifact metadata is separate from downloaded-file
-evidence: schema, column order, file hashes, and numeric parity require the
-complete files.
-
-## Public Assets
-
-| Asset | Stores | Use when | Do not use for |
-|---|---|---|---|
-| Project | Namespace and isolation boundary | Every independent business release | A dataset version or a source table |
-| Parameter | One readable source contract and its quality/time semantics | Raw or independently readable source values/events | Rolling means, ratios, trends, model columns |
-| Operator | Versioned deterministic executable code | A formula needs computation over declared inputs | A business definition without confirmed formula |
-| Feature | One immutable output-column contract | One model/input column produced by an Operator | A group of columns or an unnamed formula |
-| FeatureSet | Ordered Feature references | Consumer column order is part of the contract | Copying formulas or creating multiple datasets |
-| DatasetManifest | Dataset mode, time grid, rowsets, policy and FeatureSet reference | A reproducible dataset contract is ready | Filling unknown source or business semantics |
-| DatasetArtifact | Output of an authorized build | Build was explicitly approved and completed | Proving a Catalog is semantically correct |
-
-## Decision Rules
-
-- Parameter answers **what source value is exposed**; Operator/Feature answers
-  **how a confirmed business formula is computed**.
-- FeatureSet order is consumer-facing and must be confirmed; it is not inferred
-  from filesystem order or JSON discovery order.
-- Dataset `time_range`, `prediction`, `rowset_splits`, `abnormal_windows` and
-  `endpoint_policy` are separate contracts. A platform default does not choose
-  a business policy.
-- `apply --dry-run` checks structure, references, package and immutability. It
-  does not prove source correctness, formula correctness or artifact parity.
-- `publish` creates immutable Registry resources. `build` creates an execution
-  Job and must have a separate explicit approval.
-
-## Required Semantic Asset Review
-
-Before writing Catalog JSON, produce a local review table with one row per
-Parameter, Operator, Feature and Dataset field:
-
-```text
-asset_key | business_meaning | source/evidence | confirmed_by | unresolved | proposed_value
-```
-
-Stop before Catalog generation if any required `source/evidence`, formula,
-time boundary, unit, null policy, output dtype, FeatureSet order, read policy,
-or approval field is unresolved. The review table is a proposal for the user;
-it is not a Registry asset and must not contain guessed values.

package/skills/feature-management/references/supervised-datasets.md

@@ -1,101 +0,0 @@
-# Supervised dataset preflight
-
-Use before the first immutable publication for training/evaluation. Examples
-describe the current contract; deployed capabilities still require validation.
-
-## Labels and clocks
-
-Keep labels out of FeatureSet order. Training rejects overlap with `data.labels`
-even when an input adapter would omit that column. Generic supervised labels use
-`label_materializations`, mutually exclusive with legacy `target`.
-
-Merge this fragment into a complete manifest, replacing source identities:
-
-```json
-{
-  "mode": "training",
-  "prediction": {"horizon": "5min"},
-  "label_materializations": [{
-    "name": "future_pressure",
-    "source": {"kind": "parameter", "ref": {
-      "name": "pressure", "version": "v1", "project": "default"
-    }},
-    "output_column": "future_pressure",
-    "event_time": {"source_field": "timestamp", "offset": "0min"},
-    "alignment": {"method": "exact"},
-    "missing_policy": "report",
-    "dtype": "float64"
-  }]
-}
-```
-
-`source_field` names a column in the normalized Parameter frame, commonly
-`timestamp`, not automatically the output Feature frame's `event_time`. Other
-names need source-frame evidence. The schema lists `dataset_column` and
-`operator_output`, but the current materializer supports Parameter sources;
-resolve before building rather than treating schema acceptance as runtime support.
-Choose `report`, `reject` or `drop` according to the intended label population.
-
-- `prediction_time = event_time` is the causal cutoff.
-- `label_time = prediction_time + prediction.horizon` is the forecast target.
-- `event_time.offset` shifts the source observation clock before alignment. At
-  exact alignment, source time `s` matches label time `t` when `s + offset = t`.
-
-For cutoff 10:00 and horizon 5min, expect label time 10:05 and, with zero offset,
-the source value at 10:05. A positive offset is not a shortcut to future values
-and does not change the forecast horizon. `offset_minutes` belongs to legacy
-`target`, not generic label materialization. Free-form dictionaries can accept
-unused keys: inspect resolved semantics and sampled built rows too.
-
-Check every Operator's `input_schema.prediction` before publication. A maximum
-horizon of zero is incompatible with a 5min manifest. Review a new immutable
-Operator version and causal tests, or select a compatible Operator; do not
-compensate with label offsets or silently reduce the requested horizon. Zero
-horizon does not itself disable named splits.
-
-## Splits and schema must precede training
-
-`rowset_splits` is a mapping, not an array, and is mutually exclusive with
-`rowset`. Values are RowsetStrategy objects: `strategy`, optional `time_range`,
-and `grid` for `fixed_grid`. See [contracts.md](contracts.md) for the full split
-example. Fixed-grid split grids must match the manifest grid; ranges are
-half-open and must not overlap.
-
-Building splits adds `rowset_split` and can change the complete DataSchema hash.
-Settle split/metadata columns, labels, time columns, dtypes, roles and FeatureSet
-order before training. Matching feature values alone does not ensure evaluation
-compatibility after adding columns.
-
-Dataset split declaration does not populate the TrainingRun binding. Hand this
-fragment to the lifecycle consumer to merge under `data`:
-
-```json
-{
-  "splits": {
-    "assignment_column": "rowset_split",
-    "train": ["training"],
-    "validation": ["validation"],
-    "test": ["test"]
-  }
-}
-```
-
-Keys are consumer roles; array values are actual artifact assignment values.
-For a manifest split named `train`, use `["train"]`. Hand off the exact artifact
-identity, DataSchema hash, FeatureSet order, label/temporal contracts, observed
-split counts and binding together.
-
-## Evidence checkpoints
-
-Before publication, review schema and Operator limits and run catalog dry-run.
-After publication, resolve labels, source fields, horizon and dependencies before
-building. After build, inspect DataSchema and sampled Parquet rows for the clock
-equation, source-label alignment, nulls and split counts. A small authorized
-diagnostic build can test source semantics but cannot replace final artifact
-identity or full-data validation.
-
-Before training, validate the TrainingRun with explicit splits and confirm
-trainer shape/device and evaluation runtime support. Do not create successive
-immutable dataset versions or launch training to discover field names. Stop at
-the first unexplained error, preserve the request and evidence outside the
-repository, and correct the draft once the cause is understood.

package/skills/model-lifecycle-management/SKILL.md

@@ -1,38 +0,0 @@
----
-name: model-lifecycle-management
-description: Manage ITSMP model training, evaluation, and immutable model packaging through the deployed platform API. Use when a user asks to list or inspect registered TrainerDefinitions, register a trainer, validate or submit a training run, inspect/retry/cancel training jobs, configure or run governed evaluation, inspect evaluation evidence, validate a ModelArtifact, or create and inspect a ModelPackage. Do not use for feature catalog authoring, dataset builds, model deployment, Kubernetes administration, or image release engineering.
----
-
-# Model Lifecycle Management
-
-Use the `ml-platform` executable installed with the same npm release as this Skill. Verify `ml-platform version`, `ml-platform show-config`, and server health before API operations. Never use a platform source checkout or a second client runtime.
-
-```text
-DatasetArtifact -> TrainingRun -> TrainingJob -> ModelArtifact
-  -> EvaluationRun/Job -> EvaluationResult/Summary -> ModelPackage
-```
-
-Read [references/training.md](references/training.md) for training work, [references/evaluation.md](references/evaluation.md) for governed evaluation, and [references/packaging.md](references/packaging.md) for ModelArtifact validation and packaging. Read only the references required by the request.
-
-Read [references/discovery.md](references/discovery.md) before locating existing
-resources or selecting runtime bindings. It also defines Job/package preflight
-and version-skew handling.
-
-## Shared Gates
-
-Before training a model that will be evaluated, also read
-[references/evaluation.md](references/evaluation.md). Check the final dataset
-schema, training split binding and evaluation runtime before submission.
-OpenAPI free-form dictionaries are not complete nested contracts: use the
-examples in these references and same-release validation, not successive 422s
-to discover fields. Preserve release/error evidence when deployment differs.
-
-- Establish exact immutable inputs, TrainerDefinition `v1`, runtime image digest, model inputs/targets, resources, evaluation policy, and package destination from authoritative contracts. Do not infer them from model names.
-- Validate, register, submit, retry, cancel, and package creation are separate actions. Read-only discovery and validation do not authorize mutation or workload submission.
-- Before submission, show the exact JSON path, immutable identities, expected workload, and target API. Obtain explicit authorization. Retry and cancel require separate authorization for the exact Job.
-- Never use `latest` image tags. During development TrainerDefinition stays at `v1`; rebinding it follows the reviewed release procedure and requires confirming no active Job references the old definition.
-- A succeeded Job is insufficient evidence. Training requires a valid ModelArtifact and signature; evaluation requires immutable Result/Summary evidence and coverage; packaging requires `READY`, an immutable image digest, manifest, and passing test report.
-- Do not register caller-computed evaluation results. Workers produce predictions, metrics, decisions, and summaries from frozen DatasetArtifact rowsets.
-- Keep credentials, registry secrets, Kubernetes details, and model binaries out of request JSON and reports.
-
-Report immutable input identities, returned run/job/package IDs, status, validation or gate failures, and the next authorized action. Never claim deployment; it is outside this Skill.

package/skills/model-lifecycle-management/agents/openai.yaml

@@ -1,4 +0,0 @@
-interface:
-  display_name: "模型训练、评估与打包"
-  short_description: "管理 ITSMP 模型训练、评估与不可变模型打包流程"
-  default_prompt: "使用 $model-lifecycle-management 管理模型训练、评估或打包流程。"

package/skills/model-lifecycle-management/references/discovery.md

@@ -1,89 +0,0 @@
-# Discover exact contracts before authoring requests
-
-Use the same-release CLI/API. Start with `version`, `show-config`, `health` and
-`get-model-runtime-release`. If a documented command is absent from `--help`,
-or a new endpoint returns 404, record client/server release evidence and upgrade
-to the matching release; do not guess another resource name or submit work as a
-capability probe.
-
-## Effective runtime policy
-
-```bash
-ml-platform --profile server get-model-runtime-release
-ml-platform --profile server list-trainer-definitions
-ml-platform --profile server get-trainer-definition NAME v1 --project PROJECT
-```
-
-Discovery returns `configured`, `release_id`, `approval_policy`, and `bindings`.
-Each binding contains the exact `project/name:version:device` trainer key,
-immutable training and Serving base images, input modes and architectures.
-Compare the registered TrainerDefinition image against the binding before Job
-preflight. A binding is release approval, not proof that a matching definition
-is registered. `configured=false` means capability-based validation without a
-release allowlist; it does not mean every trainer/device is approved by a release.
-`evaluation.runtime_identity` and `default_executor` are effective server values.
-`worker_dependencies_verified=false` explicitly means discovery has not tested
-imports, image pulls or model prediction in the Worker.
-
-## Resource lookup and pagination
-
-```bash
-ml-platform --profile server list-training-runs -q pressure --limit 50 --offset 0
-ml-platform --profile server get-training-run RUN_ID
-ml-platform --profile server list-training-jobs -q RUN_ID
-ml-platform --profile server list-model-artifacts -q JOB_ID
-ml-platform --profile server get-model-artifact ARTIFACT_ID
-ml-platform --profile server list-evaluation-configs
-ml-platform --profile server list-evaluation-runs -q ARTIFACT_ID
-ml-platform --profile server list-evaluation-jobs -q RUN_ID
-ml-platform --profile server list-metric-definitions
-ml-platform --profile server list-executable-packages
-ml-platform --profile server get-executable-package NAME VERSION
-ml-platform --profile server list-model-packages -q ARTIFACT_ID
-```
-
-All these list commands return `items`, `total`, `limit`, `offset`; default limit
-is 50, maximum 500. Continue with offset plus returned item count until total is
-reached. `-q` is a case-insensitive substring of serialized metadata, not a query
-language or project authorization filter. Always verify exact identities after
-search. Collections are global metadata views; no project filter is claimed.
-Jobs in lists are immutable specs; use the exact get-job command for refreshed
-status. Lists do not reconcile or schedule workloads. `get-training-run` returns
-`training_run` plus `spec_hash`; artifact/package detail returns the object.
-
-## Read-only preflight
-
-```bash
-ml-platform --profile server validate-training-job training-job-request.json
-ml-platform --profile server validate-model-package model-package-request.json
-```
-
-Training preflight requires an already registered Run with its exact hash and
-resolves the same Job spec used by submission, including device/runtime approval,
-resources and distributed capability. Packaging preflight resolves the same
-artifact eligibility, selected runtime, input modes and immutable-version conflict
-checks used by creation. Both return `status=valid`, the proposed `job_spec` or
-`package`, `submitted=false` and `unchecked` execution checks. They do not register
-Jobs/packages or invoke schedulers. Artifact reads can populate local caches.
-
-A valid preview does not prove image availability, cluster capacity, credentials,
-framework imports, training success or packaging test success. Read `unchecked`
-and inspect an existing package's status: an immutable failed version is not
-made runnable by validation. Submission rechecks current state; preflight is not
-a reservation. Preserve the preview alongside the approved submission request.
-
-## Structured request errors
-
-Training Run `task`, `data`, `data.artifact`, `features.feature_set`, `splits`,
-`temporal` and `trainer` are explicit objects in OpenAPI. Evaluation artifact,
-config references, members and execution, plus ModelPackage requests, are also
-structured. Unknown keys return 422 with `detail[].loc` identifying the field.
-Keep algorithm-specific `trainer.parameters` and documented plugin configuration
-maps extensible; validate them against the registered plugin schema. Do not treat
-all remaining `additionalProperties` as a platform schema defect.
-
-These HTTP checks preserve accepted values and do not rewrite historical stored
-contracts or hashes. Legacy records remain readable. A previously ignored typo
-is now rejected on a new submission; correct the input rather than removing
-validation. Omit evaluation runtime identity for the server default; if execution
-is supplied it must name `executor` explicitly.

package/skills/model-lifecycle-management/references/evaluation.md

@@ -1,172 +0,0 @@
-# Evaluation Commands
-
-## Request bodies and prerequisites
-
-Evaluation reuses the model's original TrainingRun, replaces its artifact
-reference and compares the complete evaluation DataSchema hash with the trained
-Job's hash. Matching feature names is insufficient: added split/metadata columns
-or changed roles/dtypes can fail. Settle the final schema before training and
-validate compatibility before concluding that retraining is required.
-
-Example `evaluation-config.json` (exploratory, not a release gate):
-
-```json
-{
-  "name": "pressure_metrics",
-  "version": "v1",
-  "task_type": "regression",
-  "mode": "exploratory",
-  "metrics": [{
-    "name": "regression.mae", "version": "v1",
-    "targets": ["future_pressure"], "calculation_space": "business"
-  }]
-}
-```
-
-Inspect `get-metric-definition regression.mae v1` before relying on this metric.
-Release mode additionally requires at least one required validation rule with a
-business-approved threshold; do not invent thresholds to obtain PASS.
-
-Example `evaluation-request.json` shape; replace illustrative identities before
-validation or submission:
-
-```json
-{
-  "model_artifact_id": "replace_model_artifact_id",
-  "dataset": {
-    "project": "default",
-    "dataset_id": "replace_dataset_id",
-    "manifest_hash": "replace_with_exact_artifact_manifest_hash"
-  },
-  "rowset": "test",
-  "targets": ["future_pressure"],
-  "config": {"name": "pressure_metrics", "version": "v1"},
-  "execution": {
-    "executor": "kubernetes",
-    "resources": {"cpu": "2", "memory": "3Gi"},
-    "deadline": "90min",
-    "retry_limit": 0
-  }
-}
-```
-
-`config` is a name/version reference, not an inline policy. `targets` are
-model-bound label column names. The HTTP validation and submission endpoints
-fill omitted `runtime_identity` with the server's supported evaluation identity;
-omit it rather than guessing. Explicit pinning requires that exact identity
-(`sha256:` plus 64 hex digits), not an image digest or trainer hash. Inspect the
-returned frozen Job's `policy.runtime_identity` and retain it as evidence.
-Resources above illustrate shape, not Chronos-2 sizing guidance.
-
-If `execution` is omitted entirely, these HTTP endpoints use the configured
-server execution backend. If supplied, include `executor` explicitly: an empty
-execution dictionary falls back to the domain's `local` default. Do not assume
-the domain default is the deployed HTTP default. Kubernetes selection alone does
-not prove Worker dependency availability. Validate before submitting and confirm
-the approved Worker supports the model plugin. Successful training in another
-image does not establish evaluation support.
-
-## Diagnose before another build or training run
-
-| Symptom | Actual check and next action |
-| --- | --- |
-| `requires a named split` | Inherited TrainingRun `data.splits` is not a dictionary and requested rowset is not `all`. Inspect the original training request, not only manifest/resolved/Parquet splits. This check has no horizon-zero branch. |
-| `rowset is not declared` | Inspect `data.splits` role keys (`train`, `validation`, `test`), nonempty value arrays, assignment column and observed values. Manifest `training` can map to consumer role `train`. |
-| DataSchema mismatch | Compare the full schema with the trained Job's hash, including split columns, roles and dtypes. Do not remove columns or falsify hashes to pass. |
-| Label leakage | Materialize label columns separately from FeatureSet order; input selection alone cannot repair the contract. |
-| Chronos-2 runtime unavailable | Record executor, error stage/Job, runtime identity and Worker image evidence. The prediction plugin cannot load a dependency; request runtime support from the platform owner. New dataset/model versions do not install dependencies. |
-
-Do not change to `rowset=all` to bypass a requested holdout. Adding `data.splits`
-to EvaluationRunRequest is unsupported and cannot repair an immutable original
-TrainingRun. Explain the missing binding and validate a corrected training draft
-before any authorized retraining.
-
-`ExecutablePackage` is a custom metric Python wheel with verification and
-publication lifecycle, not a model Serving container. Built-in metrics need no
-user-created wheel. `ModelPackage` is a separate model image packaging object;
-creating one does not configure evaluation Workers or automatically satisfy the
-evaluation runtime identity. Empty package registration cannot repair a
-Chronos-2 import failure.
-
-## Discovery and evidence retention
-
-Submission returns `run.run_id`, singular `job.job_id` / `job_status` for the
-first member, and `jobs[].spec` / `jobs[].status` for all members. Track all
-members when present. `get-evaluation-job` returns `spec`, `status`, `attempts`;
-inspect `status.phase` and, on success, `status.result_id`. Top-level
-`status=accepted` is not completion. Read `spec.policy.runtime_identity` for the
-resolved runtime. There is no evaluation wait CLI; poll the get command with a
-bounded interval/deadline and preserve the ID on timeout. Do not use dataset
-`wait-job` for a training or evaluation Job.
-
-Evaluation request reuse is normally deduplicated. `force_rerun=true` explicitly
-requests another execution; do not set it to work around an unexplained timeout
-or missing result. Retry keeps the immutable contract; corrections to inputs
-require a corrected request, not a retry with hidden changed semantics.
-
-Default coverage requires at least one row and both prediction and label
-coverage of 1.0. Null labels and sequence context loss can prevent a gate pass
-despite successful computation. Do not lower coverage to obtain PASS. If the
-business policy intentionally permits incomplete coverage, declare the approved
-values under `coverage.minimum_rows`, `minimum_prediction_coverage` and
-`minimum_label_coverage`, and report the excluded population.
-
-A release-rule shape is shown below; the threshold is illustrative and must be
-replaced by a business-approved value. Merge under the policy and set
-`mode=release`; `metric` and version must also appear in `metrics`.
-
-```json
-{
-  "validation": [{
-    "metric": "regression.mae", "version": "v1",
-    "target": "future_pressure", "operator": "<=", "threshold": 1.0,
-    "required": true, "level": "job", "source": "metric", "slice": "overall"
-  }]
-}
-```
-
-`slices` names metadata/entity-key columns, not filter expressions or arbitrary
-Feature columns. `SUCCEEDED` describes execution; `PASS`, `FAIL`, `INCONCLUSIVE`
-describe decisions. Inspect required decisions and coverage before reporting
-release readiness. A Result and a Summary are distinct evidence objects;
-record the actual returned IDs rather than deriving one ID from another.
-
-Use `list-evaluation-configs`, `list-evaluation-runs`, `list-evaluation-jobs`
-and their get commands to locate prior evidence; see [discovery.md](discovery.md)
-for paging and exact identity recovery. `list-evaluation-attempts JOB_ID` lists
-attempt history, not all runs. Preserve request/config versions and returned
-run/job/result/summary IDs; do not search metric names instead of provenance.
-
-## Commands
-
-```bash
-ml-platform --profile server register-evaluation-config evaluation-config.json
-ml-platform --profile server register-metric-definition metric-definition.json
-ml-platform --profile server register-executable-package executable-package.json
-ml-platform --profile server verify-executable-package <name> <version>
-ml-platform --profile server publish-executable-package <name> <version>
-```
-
-Custom metric packages must be deterministic, digest-bound, network-restricted, verified, and separately authorized for publication.
-
-```bash
-ml-platform --profile server validate-evaluation-run evaluation-request.json
-ml-platform --profile server submit-evaluation evaluation-request.json
-ml-platform --profile server get-evaluation-run <run-id>
-ml-platform --profile server get-evaluation-job <job-id>
-ml-platform --profile server list-evaluation-attempts <job-id>
-ml-platform --profile server retry-evaluation-job <job-id>
-ml-platform --profile server cancel-evaluation-job <job-id>
-```
-
-An EvaluationRunRequest binds ModelArtifact(s), a frozen DatasetArtifact rowset, targets, EvaluationConfig, runtime identity, and optional baseline. Retry/cancel require separate authorization.
-
-```bash
-ml-platform --profile server get-evaluation-result <result-id>
-ml-platform --profile server get-evaluation-result-metrics <result-id>
-ml-platform --profile server get-evaluation-result-artifacts <result-id>
-ml-platform --profile server summarize-evaluation-run <run-id>
-ml-platform --profile server get-evaluation-summary <summary-id>
-```
-
-Report candidate, eligible, labeled and evaluable coverage; calculation space; slices; baseline identity; required decisions; runtime/package digests; and Result/Summary hashes. Exploratory or incomplete evidence is never a release gate pass.