Understanding an assessment
Last updated 6 September 2026
PR comment and dashboard
After frontier upload, GitHub Actions upserts one aggregate-only PR comment. The same fields appear on the run-detail page. Entity IDs and warehouse row values are never included in the comment.
Values below that describe missed entities, mismatched rows, failed validations, or FULL_REBUILD_REQUIRED are correctness problems. Rows avoided, candidate no-ops, and FULL_REBUILD_RECOMMENDED are efficiency warnings unless paired with a failed check.
Field guide
| Field | Meaning | Signal |
|---|---|---|
| Assessment status | PASSED or FAILED (comment headline). | Correctness if failed. |
| Model | Target mart from the pinned semantic manifest. | Context. |
| Assessment type | SQL-change, CDC, or combined. CDC runs also show provider and batch checkpoint. | Context. |
| SQL operator | First classified change kind, such as FILTER_CHANGED. | Context. |
| Changed source rows | Source rows whose predicate or event membership changed. | Context. "Not measured" when a full rebuild skipped discovery. |
| Event-derived candidates | Entities routed from source events or CDC logical events. | Context. |
| SQL-change candidates | Entities selected by compiled SQL-change predicates. | Context. |
| Candidate union | Deduplicated candidate keys (event ∪ SQL-change). | Context. Conservative set. |
| Confirmed changes | Candidates whose mart output actually changed. | Context. Smaller than the candidate set when no-ops exist. |
| Candidate no-ops | Candidates whose mart row did not change. | Efficiency. Not a miss. |
| Missed entities | Full-reference changes absent from the candidate set. | Correctness problem. |
| Full and targeted row counts | Full entity count versus frontier/candidate recomputes. SQL comments may show before → after row count. | Context. "Not measured" when rebuild skipped execution. |
| Rows avoided | Percent of the full population not recomputed. | Efficiency. |
| Validation differences | Sum of failed check difference counts. | Correctness if greater than zero. |
| Evidence level | none, aggregates, equivalence, or empirically validated. | Empirically validated is evidence for this dataset, not a formal proof. |
| Impact compilation status | COMPILED or FULL_REBUILD_REQUIRED per modified model. | FULL_REBUILD_REQUIRED is fail-closed correctness. |
| Impact execution status | EXECUTED, NOT_RUN, NOT_EVALUATED, or FAILED. | FAILED is a correctness/runtime problem. NOT_RUN is expected when rebuild is required. |
| Manifest version / fingerprint | Pinned snapshot used for this run, not the later active mapping. | Context. Mismatch with dbt is a correctness setup issue. |
| Base and PR artifact fingerprints | Hashes of the compared compiled SQL artifacts. | Context. Traceability. |
| Targeted repair safety | Safe, not safe, or skipped. Safe requires zero mismatched final rows, zero missed entities, and no required/recommended rebuild. | Not safe is a correctness problem. Skipped follows rebuild policy. |
| Full-backfill decision | Required, recommended, or not required. | Required is fail-closed. Recommended is an efficiency warning. |