Evaluation Core artifact persistence

Current host persistence, reuse, and downstream projection contracts. Production omk eval, Studio, Gold, resume, batch, and evolve consume these Core artifacts. Transported JSON never becomes a trusted Core capability.

1. Boundary

Evaluation Core remains pure and in-memory. The host persists one immutable run artifact set:

1. the versioned RunPlan wire document;
2. ExecutionBundle;
3. EvaluationBundle;
4. AnalysisBundle;
5. EvaluationReport.

The Plan is required because Bundle digests identify a sealed measurement contract but cannot reconstruct it. Persisting only results would force resume to trust summaries or recompile against ambient state. The host manifest is a locator and integrity projection, not a sixth measurement fact.

2. Authority and verification

On publication and full read, the store validates:

• every wire schema and native artifact digest;
• the Plan's recomputed dataset, stage, randomization, and run-contract digests;
• Execution → Evaluation → Analysis → Report parent digests;
• Report Bundle references, status, budget summary, Decision binding, and provenance parents;
• manifest document digests and maximum captured-content classification;
• every referenced content descriptor through an injected resolver.

This proves transport integrity and exact lineage, not provenance authenticity. A persisted RunPlan is returned as a wire document and is never rebranded as SealedRunPlan. Resume must prepare the current Definition and MeasurementPolicy, obtain a fresh sealed Plan capability, and run the Core plan-aware Bundle verifiers before reuse. Digest equality alone cannot mint source trust, cache receipts, or Runtime attestation.

3. Atomic publication

Each run uses an opaque directory derived from the SHA-256 of runId; user identifiers never become path segments. The host writes all five documents and the manifest into a private staging directory, writes the manifest last, then publishes the directory with one same-filesystem rename. Readers enumerate only published run-<digest> directories. Interrupted hidden staging directories are invisible and can be cleaned independently.

A runId is immutable. Re-publishing the exact document set is idempotent; a different set under the same ID fails. In-process writes are serialized per run, and the directory publication step also resolves cross-process races to one complete winner. Corrupt published runs fail explicitly instead of disappearing from list results.

Directories use owner-only permissions and files use owner read/write permissions because a full Plan can contain expected or evaluation-only inputs. Paths are host effects and never enter Core documents or measurement digests.

4. Manifest and index

omk.core-run-artifact-manifest/v2 uses qualified discriminants (manifestKind, documentKind). Its five references have fixed relative filenames, schema identity, native identity digest, and a canonical full-document digest. It also records run/report identity, run-contract digest, orthogonal Report status, Execution/Evaluation replayability, creation time, and maximum captured-content classification. Layout v2 names the authoritative report report.json; older manifest bundles are not read.

An index card is a pure projection of a validated manifest. Listing may rebuild cards without resolving large content because it does not claim evidence availability. Loading a full run requires descriptor closure; a missing resolver or corrupt content fails instead of projecting empty evidence.

5. Host content store

The Node content store implements the existing Execution and Evaluation content ports. It verifies the caller-provided value digest before writing and stores a private, content-addressed envelope. Identity includes value digest, media type, and classification so the same bytes classified as public and gold cannot alias or downgrade each other. Descriptors use an opaque omk-content: URI and never expose a filesystem path.

Resolution revalidates envelope digest, descriptor, canonical value digest, media type, byte size, and classification. Failures use stable, redacted codes; raw filesystem paths and content do not become Core errors.

6. Resume disposition

Resume is complete-fact reuse, not record copying or cross-process checkpoint continuation. The host first prepares the current Definition and MeasurementPolicy to obtain a fresh SealedRunPlan. The admission adapter rejects a transported Plan that merely has the same JSON shape, then loads the located run and invokes the Core Execution, Evaluation, Analysis, Decision, and Report verifiers in order.

Only a completed run with complete evidence can be reused. A missing, corrupt, partial, contract-mismatched, under-trusted, cache-indeterminate, or budget-indeterminate source produces a stable reason code. The caller must explicitly choose fail-closed or start-fresh, a minimum source trust, and whether cache receipts and budget accounting must be verified. Rejection never copies successful rows, creates a new trial, or changes the old lineage.

Verification contexts are independent host evidence. A caller must never construct provenance attestations or cache receipts from Bundle claims. The admission digest records only artifact identities, normalized Core verification facts, and the explicit policy; it excludes raw Dataset, Gold, outputs, traces, and receipt material.

7. Project and global overlay

An overlay writes only to its primary store and reads primary before fallback layers. Its index card includes a rebuildable artifact-set digest over the five document references. The same runId may appear in multiple layers only when this digest is identical; otherwise get, list, and exists fail explicitly. Writes never shadow an existing fallback ID, including an identical one, so project and global stores cannot silently select different facts.

8. Batch child runs

omk.core-batch-manifest/v1 is a host index, not a Core Report. Each ordered item contains a qualified batchItemKind, stable itemId, independent child runId locator, artifact-set and Report identities, status, and maximum captured-content classification. The child keeps its own Plan, Bundles, Report, provenance, and statistical unit.

Publication resolves and fully verifies every child before atomically publishing the private batch manifest. A missing, corrupt, duplicate, or identity-changed child fails. Full batch reads revalidate child locators; batch listing validates only the manifest and therefore does not claim child evidence availability. Interrupted staging directories remain invisible.

9. Core-native downstream projections

Downstream consumers read immutable Core facts through pure, side-effect-free projections. These APIs do not import the legacy Report or result-row types. Production consumers now use these projections after the atomic CLI cutover.

The artifact graph projection consumes one validated stored run. It emits qualified Core node kinds for Run, Target, Evaluator, Metric, Execution result, Evaluation result, Analysis result, and Decision. Node status comes only from explicit stage state. Numeric observations may appear as view metrics, but no score band or display threshold becomes a verdict. Evidence references bind exact Core document hashes and JSON pointers; raw input, expected value, output, trace, evidence, and error text are omitted.

Gold comparison requires the caller to select one sealed Target, Evaluator, and numeric Metric. Optional trial and measurement coordinates refine that identity, but must match the Plan. Gold and Metric scales must match exactly. A sample resolving to multiple observations fails instead of averaging across trials, ensemble members, or replicates; a scale transformation belongs in a preregistered Analysis. The v2 projection binds a canonical Gold dataset digest and uses null, not non-JSON NaN, for undefined statistics. It records complete bootstrap-draw coverage and uses the fixed-expected-disagreement reliability bootstrap; perfect agreement has an explicit non-applicable interval. An optional, caller-selected minimum alpha produces a post-hoc assessment against the interval lower bound; there is no universal default and the assessment is not a release DecisionResult. The Krippendorff alpha, weighted kappa, and Pearson point formulas remain unchanged; the v2 interval is a BREAKING-COMPARABILITY change from v1's conditional-finite-draw bootstrap.

Evolution evidence consumes one exact Evaluation Series Plan → SeriesAnalysisBundle → EvaluationSeriesReport chain. Series fixes experimentalUnit = run; the projection preserves member status, trust, comparability, analysis standards, assumptions, coverage, and registered Decision. It never ranks member runs or infers a winner from display scores. Only a decided Series Decision makes the view decision-ready; completed analysis without a Decision remains analysis-only.

The CLI dry-run projection accepts only the in-process SealedRunPlan capability returned by prepare. It exposes stage digests, Dataset／Target／Evaluator／Metric counts, experimental-unit identity, Analysis outputs, Decision policy identity, and redacted preflight records. It contains no input, expected value, evaluation context, artifact config, locator, or provider secret. A transported Plan document cannot be projected as a successful prepare result.

The CLI run projection accepts one fully validated stored artifact chain. Operational completion, evidence completeness, conclusion status, and registered Decision remain separate inputs to the gate. A verdict is not release authority by itself: PROGRESS requires the stable release-gates-passed reason and SOLO requires solo-layer-gate-passed. report-only may skip the release gate only for a completed run; cancellation, budget exhaustion, and failure remain non-zero operational outcomes. Batch projection verifies every manifest child against its exact artifact-set identity and projects independent child outcomes without pooling scores, Reports, or statistical units.

10. Non-goals

• legacy EvaluationReport readers, migration, or dual writes;
• partial-record resume or a cross-process checkpoint engine;
• Studio projections;
• artifact signatures or provenance attestation.

The production switch was delivered as one atomic BREAKING-SCHEMA change. No current command reads both legacy and Core facts.