openpond-sdk 0.7.0 → 0.8.1

package/MODEL_STARTERS.md

@@ -147,3 +147,28 @@ retain a null metric value. The client verifies workspace, manifest,
 population, grading role, metric references and result hashes on readback.
 Hosts remain responsible for owner-authenticated evidence and durable
 dispatch/cancellation; a content hash is not producer authentication.
+
+### Dataset evaluations without a model configuration
+
+SDK 0.8 allows an evaluation request to set `modelProjectId: null` and select an
+exact published dataset release directly. The hosted executor still resolves the
+complete package and checks its supported runtime before admitting the run.
+Model-configuration-backed requests continue to require their attached release.
+
+`policy.messages` optionally supplies system, user and assistant context before
+each example. Those messages form part of the immutable execution policy hash;
+changing them produces a different evaluation configuration. Dataset reference
+answers are never included in these messages by the host.
+
+The optional `project` field contains `{ id, revision, contentHash, targetId }`.
+Its hash identifies the selected Project **content**, and `targetId` may be null
+for an ad hoc evaluation inside that Project. Hosts authorize the Project and
+resolve its retained revision before execution. Request, summary and manifest
+must agree on this provenance. Project attribution is not inferred from shared
+datasets or model configurations.
+
+`runs.list({ projectId })` reads that Project's attributed evaluations;
+`runs.list({})` reads authorized workspace history. Existing
+`runs.list({ modelProjectId })` filtering remains available. Callers must handle
+nullable `summary.modelProjectId` for independent evaluations. No trained model
+or serving deployment is created by running an evaluation.

package/TASKSETS.md

@@ -304,3 +304,11 @@ Parquet collections through the dataset reader. Search never examines private
 answers. Draft edits invalidate the index, and a published release remains
 separate from its draft. No task listing, editing, or labeling operation starts
 a training job.
+
+Independent dataset writes may include `originProjectId` at creation. This is
+immutable creation provenance, not exclusive ownership or a sharing grant. The
+host verifies that the Project belongs to the same workspace and is active.
+Receipts and list entries retain the origin so unpublished drafts remain visible
+in their originating Project. A release can still be used across Projects through
+explicit associations; changing or removing that association does not rewrite
+where the draft was created.

package/TRAINING_PROTOCOL.md

@@ -78,6 +78,34 @@ checkpoint inventory or qualification metrics, when the portable client needs
 that evidence to validate the artifact. This metadata is immutable output
 evidence, not private worker-control, lease, or provider state.
 
+## Manual run preparation
+
+Hosts supporting reviewed-data training expose these authenticated, team-scoped
+routes through `createTrainingClient`:
+
+```text
+POST /v1/training/preparations
+GET  /v1/training/preparations/{preparationId}?teamId=
+POST /v1/training/preparations/{preparationId}/start
+POST /v1/training/preparations/{preparationId}/cancel
+```
+
+`prepareRun` accepts a stable operation ID, the expected configuration revision
+and ETag, an exact sealed batch reference, an optional organizing Project, a run
+name and a maximum spend. Preparation freezes the executable inputs and stages
+artifacts without starting compute or creating a recurring learning policy.
+The returned plan binds its creator, request hash, source and submission hash.
+The SDK verifies both plan and request identity before returning the receipt.
+
+`startPreparation` and `cancelPreparation` require the retained plan hash and
+observed receipt revision. Starting uses the stored submission; mutable
+configuration changes do not silently recompile a prepared run. Providers must
+serialize cancellation with job and dispatch creation, retain the submitted job
+ID, and make identical retries return that same job. A cancelled preparation
+cannot start. Once submitted, use the ordinary job cancellation controls.
+Reading a preparation does not authorize starting it; the host rechecks scope,
+creator permissions, resource access, runtime availability and spending limits.
+
 ## Deterministic grading identity
 
 For portable training bundles, derive `job.rewardSource` with