A new NotebookLM card differs from an actual table or video ready for its intended use. Its title and content should be inspectable, and downloaded bytes require another check. Northstar records separate stages from opening a feature panel through rereading its existing result.
Opening a feature panel does not report generated content
The open path returns panel-opened with contentGenerated=False after displaying configuration. Visible menus and newly generated content are different events. Basic-template work also differs from a customized instruction flow: an explicit report instruction cannot be silently ignored when no input action is present.
The registered feature count describes catalog entries and execution entrances, not every tested style, language and export combination. I distinguish available panels from actually verified generation and file outcomes. A large capability list should not become a promise of unestablished completeness to a customer.
Do not count hover-only more buttons as results
Early detection could vary with more buttons visible only on hover. NotebookStudioCards instead reads recognized card buttons containing source-count/time metadata and extracts their titles. Studio-created checks that enough cards were added and generation/failure indicators are absent.
That stage does not semantically review all card contents. It is screen evidence of a prepared result. Card existence cannot establish a table's calculation accuracy or a video's quality. It identifies which result to inspect next while preserving the difference between existence and usefulness.
Read the original title and actual content after uncertainty
The bounded status reconciliation path finds an existing result matching proofArtifact. It combines the title with a table containing data rows or audio/video with positive duration. A matching card title alone cannot clear uncertainty. Substituting another result would lose the original request's state.
This is not a semantic check of every cell or playback of an entire video. The relevant evidence is which current content was linked to the existing result. The response says reconciled-existing-artifact, readOnly and contentGenerated=False. ReadOnly here means no fresh generation, not absence of UI clicks needed to open the result.
A download receipt is another proof after artifact observation
The historical record checks the video's title/media and then compares a download receipt with saved size and digest. I reread the retained local file during this writing task and confirmed its bytes still match that proof. No fresh remote download or playback test was performed.
A historical parser error was fixed by reading the nested download object rather than top-level metadata. Since the file already existed, only that particular record was reconciled through its evidence. The result did not release unrelated uncertain generation, publication or upload records in bulk. Recovery needs a target and bounded evidence too.
Retain both verified features and remaining scope
Historical records include text sources, questions, reopened notes, table/video results and file receipt. Mind-map generation cards and title were observed, while internal branch operation and saving remained unverified because the actual frame's provenance was not established. A passing independent frame test was not relabeled as successful real mind-map operation.
Sharing, deletion, permission changes and every advanced style, language and export combination were not all executed. This task reads those records, functions and fixtures. Tests of card recognition and exact-title reconciliation do not establish complete service coverage. The remaining scope stays visible to the next worker.
Explain how far the requested outcome was checked
A customer needs to know which result came from their material and how far it was verified, more than the menu's feature count. Generation start, prepared card, title/content and download each identify an unfinished stage when evidence is missing. One completed label should not attempt to satisfy every expectation.
The shared observer compares synthetic new-response and media readiness data; it does not implement all NotebookLM card, table or frame handling. I use the common principle without claiming identical adapter contracts. Extensibility needs both shared concepts and readable evidence specific to each capability.