LabTrace: Experimental Change Memory

Inspiration: Researchers and lab teams accumulate protocols, notes, manifests, and reports on their NAS across runs. Reconstructing what changed between runs means reopening records and trusting memory, so differences are missed, conflicting values go unnoticed, and explanations outrun evidence. What it does: LabTrace compares a selected experiment run with an earlier run from the same study in a researcher-authorized NAS folder. It retrieves both runs' protocols, notes, manifests, and reports, reports source-linked differences, flags missing or conflicting metadata, and assesses progress or regression only against an explicitly defined criterion. Every field is status-labelled MATCH, DIFFERENT, MISSING, CONFLICT, or UNREADABLE, and explanations are separated into recorded explanations, candidate explanations labelled as unverified hypotheses, or no supported explanation. Product principle: Show what the records support. Keep observation, interpretation, and causal explanation separate. How I built it: LabTrace is a read-only parent Skill, experiment-run-diff, written in Markdown with five linked references covering workflow, comparison rules, report format, field schema, a worked example, and three synthetic test fixtures with expected outputs. It delegates each step to one registered UGOS expert agent, carrying an evidence record forward so no claim enters the report without a source: file-agent, ug-filesystem (find, stat), Steps 3, 4, and 9; document-agent, ug-document (extract, qa, reconcile), Steps 5, 6, and 8; knowledge-agent, ug-knowledge (save-text), Step 9 only after the researcher confirms the exact destination. Originals are never altered. The Skill selects no model and calls no external service; if the user selects a cloud model in Uliya, document-agent text may go to that service, disclosed in the Skill and proposal. Tools disclosed: uliya-skill-kit 0.7.3 and an AI assistant were used to prepare the package. No external runtime services, datasets, or code are included in the Skill ZIP. Challenges: The supported capability catalog does not support derived-metric calculation, statistical analysis, significance testing, or unit conversion, so those were removed rather than simulated. Reconciliation output detail is not fully documented, so comparison logic is defined in the parent Skill from extracted values. Accomplishments: A complete source-linked workflow that withholds classifications the records do not support, labels candidate explanations as unverified hypotheses instead of causes, flags conflicts instead of resolving them silently, and never selects an ambiguous comparison run without asking. What I learned: Designing for evidence first, with an evidence record for every extracted field, produces a report a researcher can audit, unlike conventional AI that merges documents into a fluent conclusion. What's next: Finalist validation on the assigned UGREEN AI NAS if selected, using two core cases (explicit metric with complete reports; differences with no metric defined, classification withheld) and one edge case (run identifier conflict while multiple conditions change).