Two refined how document evidence is presented. It implemented human-readable file and location descriptions in answers, with PDF source links opening the selected page in the workspace. Updated code and self-restart records support that the guidance was applied. Two continued this document work and the restarts without a new observer request. Saved browser checks of link behavior used isolated examples; the observer did not run a new real-user conversation. The comparison reused the earlier questions about synthetic documents. The agent recorded every new answer as passing its own review, yet also acknowledged a minor location inaccuracy in an answer that counted a heading among paragraphs while still awarding a pass. The observer checked that the saved answers matched the grading records, not that every answer was independently certified perfect. The earlier comparison material also combined answers from the initial run with answers from later repairs of selected cases. To address the remaining ambiguity, Two added Word locations described as paragraphs or tables under the nearest heading. Headings themselves are excluded from body-paragraph counts, and comments anchored at those locations use the same descriptions. The observer compared saved results from the synthetic document and confirmed that the original file and existing content were retained. A further self-managed restart applied this guidance. For Excel, Two added sheet and cell locations with explanations of hidden rows and columns, and the anchor and interior cells of merged regions. These distinguish a cell absent from the file, a stored blank, numeric zero, and a formula without a saved calculated value. Formulas are not recalculated; stored values and states remain the evidence. The observer compared saved before-and-after results and confirmed that values, states, sources, and the selected range were unchanged apart from the new location labels and explanatory wording. This was not a fresh execution of the reader. Updated code and another self-managed restart support application of the Excel guidance. Actual model use of the new descriptions has not yet been measured. Word locations derive from document structure and headings, not exact rendered page or line coordinates. Repeated headings or complex tables can still need context. Self-grading on reused questions and application of code do not establish answer quality on new user questions. A fresh real-user conversation using this guidance, from material selection through reading, citation, and answering, remains to be verified.
Edited for public reading from the observation journal. If later observations change an interpretation, it is recorded in a new note.