Skip to content

Inspect and preserve run evidence

A run bundle is a self-describing directory. During execution, new results and journal generations are committed atomically. Terminal bundles should be preserved unchanged. Copy the complete directory when transferring evidence; loose summaries omit the integrity and provenance needed for interpretation.

Discover through the manifest

Start with manifest.json, then follow its referenced input, schema, result, artifact and execution indexes. Paths inside those records are contained relative paths; filenames outside the indexes have no contract meaning.

flowchart LR
    accTitle: Read a run through its manifest and indexes
    accDescr: The manifest references authored and resolved inputs, the result index, artifact index and execution index. Results provide assessments and lineage; artifacts provide registered files; execution records provide attempts and committed checkpoints.
    MANIFEST["Run manifest"] --> INPUTS["Authored and resolved inputs"]
    MANIFEST --> RESULTS["Result index and envelopes"]
    MANIFEST --> ARTIFACTS["Artifact index and registered files"]
    MANIFEST --> EXECUTION["Execution index and committed state"]

Follow the referenced paths and verify their digests. These are related views of one run, not independent files to combine across runs.

Record Why it matters
Manifest Run/study IDs, lifecycle, application revision, seed and input snapshot digest
Authored/resolved inputs What was requested and exactly what resources/defaults resolved
Result index and envelopes Domain ownership, execution/assessment, payload, issues, metrics and lineage
Artifact index Registered reports, plots, logs and domain artifacts with media type, byte size and digest
Format-2 execution index Committed graph/state, attempts, events, telemetry, fingerprint and checkpoints

Use sindri status my-run --json for the supported status projection. It verifies the bundle before reporting status. Machine integrations must honor the embedded schemas and indexed references; do not scan for a familiar filename and silently accept a mismatched payload.

Distinct states

Layer Values Interpretation
Run lifecycle running, completed, failed, incomplete, cancelled Overall execution state
Result availability succeeded, partial, failed, skipped, absent Whether/how a planned result is available
Result execution succeeded, partial, failed, skipped What happened to an emitted result operation
Engineering assessment pass, warn, fail, not_assessed Conclusion under represented criteria

Absence is not an invented empty result. Failed/skipped execution has no engineering assessment; partial execution cannot claim a pass. A system-level completion or convergence label cannot override individual assessment failures. See system closure and evidence limits.

Integrity and provenance

Readers verify referenced paths, sizes and SHA-256 digests; result/index metadata, snapshot binding, producer/artifact relationships and domain consistency must also agree. A mismatch rejects the bundle. Do not “repair” a hash to make edited output look original.

A digest proves content identity/integrity within this record chain, not the scientific truth of a result. Keep requirement IDs, metric units, issue codes, model/resource identity and applicability in exported summaries. Reports should display the recorded assessments, not recalculate a new score from issue severity.

Compatibility

Spacecraft evaluations and advisory reviews are separate immutable artifact directories outside the original run. Preserve their complete manifests, policy snapshots and revision histories together with the referenced source bundle. Evaluation and review never backfill a source run.

Current writing/recovery uses bundle format 2. Format 1 remains supported for integrity checking and inspection only; it cannot be modified, resumed or used as a replay source. This format-1 support does not cover arbitrary pre-1.0 folders or loose historical summary files.

Historical study-v1 snapshots and compact sweep results remain readable within supported bundles. They are not executable or resumable through the current study-v2 path. Preserve their original files and frame meanings; prepare a new design study from separate current authored inputs.

Contract-suite version, bundle format and domain payload version are separate identities. Compare the actual schema/payload kind, not just a filename or the application's headline version. Historical authority metadata may be unavailable when its recorded registry digest differs from the local resource revision.

Exact interfaces: manifest, result index, result envelope, artifact index and typed issues.