> For clean Markdown of any page, append .md to the page URL. > For a complete documentation index, see https://docs.nvidia.com/docsfab/llms.txt. > For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/docsfab/_mcp/server. # DocsFab Review JSON Result Reference > Look up the machine-readable review result, report artifacts, and retained evidence boundaries. Look up the machine-readable review result, report artifacts, and retained evidence boundaries. ## Consume the JSON Result Outer agents and continuous integration wrappers must run the workflow with `--json` and consume the single JSON object on stdout. Do not scrape the human footer or skill Markdown. ```bash docsfab run review-documentation \ --repo /absolute/path/to/repository \ --revision \ --json ``` When a run record exists, the outer envelope remains `schemaVersion: 1` and `kind: "docsfab-run-result"`. This remains true when a projected string exceeds 512 characters and is truncated. Inspect the quality limitations to detect that truncation. Use `product.findings` for workflow outcome codes such as `review-required`. Use the root `reviewFindings` array for actionable `ReviewFindingV1` objects. Each object contains `id`, `severity`, `code`, `summary`, and a relative `path` or `null`. It can also contain `line`, `origin`, `evidence`, and `recommendation`. Findings lifted from admitted sealed review skills use `origin: "agent"` and include evidence and a recommendation. Current `review-documentation` results always include `reviewFindings` and `reviewFindingsProjection`. When no findings can be safely projected, `reviewFindings` is `[]`. Check projection availability and limitations before interpreting that empty array as a zero-findings result. Older envelopes can omit these additive fields. The final validated quality evaluation keeps its source finding IDs. Rich source locators remain in `evidence`; the projection adds a safe first path and line only when the locator supports them. Artifact admission proves schema and run bindings. It does not independently prove that an agent-authored finding is factually correct. Findings from a completed technical-verification step can remain in `reviewFindings` when a later evaluator fails. They are retained evidence and do not certify the overall run: execution status, quality validation, and product disposition still report their failure or unavailability. ```json { "product": { "findings": ["review-required"] }, "reviewFindings": [ { "id": "finding:review-documentation:EDIT-001:1", "severity": "warning", "code": "EDIT-001", "summary": "The explanation is unclear.", "path": "docs/guide.md", "line": 18, "origin": "agent", "evidence": "source-1: docs/guide.md:18", "recommendation": "Clarify the explanation." } ], "reviewFindingsProjection": { "projector": "docsfab-review-findings-v2", "availability": "partial", "artifactLocator": "record:sha256:aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa", "limitations": ["review-findings-unparsed:review-documentation"] } } ``` The excerpt shows the two finding surfaces. Read the complete `docsfab-run-result` envelope for execution, quality, report, next-step, and result-digest fields. If quality report projection fails, the envelope still exists with `quality: null`, a `quality-projection-unavailable` product limitation, and an actionable `nextSteps` entry. `reviewFindingsProjection.availability` is separate from finding count and takes one of the following values: - `complete` means no projection limitations remain. An empty `reviewFindings` array is a complete zero-findings projection. - `partial` means admitted evidence remains with projection or coverage limitations. The findings array can be empty. Inspect the retained report for the named source. - `unavailable` means required evidence, a binding, an evaluator, an artifact, or a schema is unavailable. - `invalid` means finding admission or integrity validation failed. Invalid findings are excluded. When independent valid findings survive another source's failure, availability is `partial`. `docsfab runs show --json` reloads the same content-addressed artifact and produces the same findings projection and recovery action as initial delivery. Recovery depends on the recorded state, as the following list describes: - When a bound artifact locator exists but the object is absent, restore the matching known-good record and artifact set, then replay locally. - When the installed evaluation schema is unavailable, restore that runtime schema, preserve the retained bytes, and replay locally. - When integrity or binding validation fails, recover a matching known-good set or fix the producer. Never edit sealed evidence to make it pass. - When no admitted evaluation artifact was recorded, there is no artifact to restore or replay. The evaluation producer owner must supply validated evidence before findings can be projected. A valid finding is a review result, not a transient provider failure. Inspect the cited evidence and repair it only under a separate write grant; do not rerun the unchanged review as the recovery step. The human `QUALITY` section and JSON `product.limitations` also report what the retained evidence establishes. `review-method-static-read-only` means the recorded review steps used only the admitted read tools. `package-command-checks-not-executed` means that record contains no package command execution; passing artifact validators does not establish package checks. When a native host-operation summary does not carry typed command execution evidence, both the review method and package-check disposition remain `not-established`; DocsFab preserves the summary but does not scrape its JSON or prose into evidence authority. A typed native-check projection remains future work. Human or agent audience outcomes emitted by the model remain `model-claimed-only` even when their envelope passed host validation. An empty audience outcome is `unavailable`, not a successful audience check. `review-findings-human-review:` preserves an admitted source verdict that requires human review. A `review-findings-coverage-unverified::` limitation identifies each safe target path whose technical coverage was explicitly unverified. Either limitation keeps an empty findings array partial rather than turning missing evidence into a zero-findings result. When the evaluator explicitly closes a complete findings set without any admitted findings, DocsFab records `evaluation.findings` as `[]` and does not create placeholder findings. Current retained technical inventories must remain canonical, schema-valid, and covered by the frozen review scope or the succeeded authoring plan. An invalid or missing current authoring plan binding cannot widen admission. Historical records with no current inventory remain recoverable, but carry an explicit inventory-unavailable limitation and therefore cannot establish complete findings availability.