Finding follow-ups
Compare later source evidence with an original finding and review correctness separately.
Watch the workflow: Check follow-up source prerequisites · Authored dummy data
This recording uses an earlier Settings layout. It shows the missing review-case source state and a historical Connections panel. In the current app, follow Open Settings, then choose Integrations → Braintrust to inspect saved connection metadata. The recorded connector-checkpoint panel is not present in the current form; Connected only confirms saved configuration. A connected ingestion project does not register a review-case source or establish stored-arrival or review coverage. No follow-up comparison, original correctness judgment, source registration, permission change, provider probe or new connection is performed. This opening covers prerequisites; the source-dependent review workflow remains unavailable.
Read the walkthrough transcript
- 00:00 — Open Review cases in Tensile. No review-case source is connected here.
- 00:06 — Inspect Connections in Settings. A connector checkpoint does not establish review coverage.
Finding follow-ups show what happened after an original review: an observed document or field change, a retained warning, a change that was not originally flagged, a new later defect, or missing evidence. They are available in Review cases for workspaces with a registered review-case source whose Braintrust case and trace projects are connected. Every brokerage account in that source is included; there is no separate account list to maintain. If the projects are disconnected, Review cases says so and saved revisions return once they are reconnected.
Inspect a follow-up
Open a saved export revision, choose a run, and inspect Finding follow-ups. Each observation shows its source evidence cutoff, original and later versions, observed state, deduced action, explicit source response, proposed correctness, and your saved draft. The authorized original reviewer assesses independently: other reviewers’ answers stay hidden until the reviewer locks their own assessment. Customer viewers can read the team’s locked assessments. Each finding keeps Response, Later change, Proposed correctness (candidate), and Review separate. Saved Unknown reviews are counted separately from definitive judgments. A saved original-defect judgment remains visible when coverage is incomplete, while its missed-finding classification stays unknown; pending and ambiguous later evidence have separate counts. The visible hold or Unknown reason explains what is missing. Expand Inspect evidence to check the exact subject, field, version relationship and rule references. In the selected run’s review view, later comparisons, source provenance and their counts appear after the reviewer locks eligible original assessments. This presentation does not restrict existing approved source pages, sibling runs, decision journals or exports; operational original review is not a statistically blind benchmark. Missing-rule observations remain available after those locks and to customer viewers.
A document hash change shows that a document changed. It does not establish which field changed or prove that a missing signature was corrected. A possible correction needs matching field evidence, a source-owned field/version relationship and the original applicable rule. Scalar comparisons without that relationship are ambiguous and excluded from eligible follow-ups. Same-document fingerprint comparisons support only the narrower document-change observation. Source responses and suggested labels remain separate from human review.
Review original correctness
When the required original rule and readable evidence are available, choose Review original correctness. Assess whether an applicable defect was observable at the original review. Save a draft or lock your answer in the existing review queue. The initially authorized verified FD reviewer owns this review; other reviewers need explicit delegation. Customer acceptance remains separate. Corrections retain the earlier answer history. After the original lock, changed definitive answers are refused because later or peer evidence may be visible; correctness unknown remains available. A temporary source-read error prevents saving; retry first. If you deliberately lock correctness unknown while original pages are unavailable, that lock cannot become a definitive answer when pages return. Reassessing the same original evidence with another reviewer is not supported in this release. Only a genuinely changed original rule or original evidence creates a new case; a later correction does not.
A later signature, negotiated amendment, changed rule or new defect does not by itself establish that the original finding was correct. A proposed miss also needs complete original check coverage. Choose Correctness unknown when the original evidence is insufficient or disputed.
Understand missing evidence and counts
Missing rules hold correctness review while a narrower observed change can still be shown. Missing follow-ups remain pending. Incomplete history and evidence outside the cutoff remain unavailable; wrong subjects, unreadable comparisons and unverified field relationships are ambiguous. These are excluded from eligible follow-up counts. Original findings, follow-up observations, candidates and your locked reviews count different units.
Frozen passing check coverage can expose an unflagged check for independent review. These original-check prompts are counted separately and do not count as observed changes or eligible follow-ups. Standing failed or missing checks do not imply an omitted finding. A human must establish the original applicable defect before it becomes a confirmed miss.
For a retained-only source, the FD team reviews an explicitly selected saved revision; live import is unavailable. New retained revisions need a separately checked scope. A connected live source can import a new bounded export revision to observe newly available evidence; a brokerage account that first appears in that export is included without any configuration change. The account filter only narrows the rows of a saved revision. Saved revisions remain fixed. Later drift or a late source outcome does not retire the same original assessment. When its exact earlier revision is reviewable, Review exact original revision returns to that evidence and later comparisons remain deferred until its original lock. A temporary lookup failure keeps comparisons deferred for retry. If no reviewable original revision exists, correctness stays unknown; no original evidence is manufactured. A changed original rule or original evidence needs a new review. Later evidence and a newer cutoff retain the original assessment when the original-time facts match. Reserved review populations remain held under the existing isolation rules.
Team review does not establish customer acceptance or approved evaluator accuracy. Follow-ups do not close source findings, change documents, contact customers or execute remediation.
Original document judgments require a generation-frozen value inventory, complete page indices and the original render. Each review view checks that the original PNG bytes remain readable and match the source-owned page checksum before revealing later evidence or saved correctness labels. Definitive judgments check those bytes again when saving. If an attachment becomes unavailable or changes, the review view holds later evidence until the original bytes are restored; the saved answer history stays intact. WebP-only images, incomplete page sets, missing original images and sets larger than twenty pages remain held for correctness; narrower document-change observations can still proceed. The source owner supplies the approved original evidence and access grant; the authorized original reviewer owns the next original judgment. No source permission is inferred from a review answer.