Network-analysis functions and intended uses
A browser-based network-analysis tool whose numerical calculations have been checked independently, while their substantive interpretation remains a separate problem
- Reads exports people already hold (Slack, email, calendars), surveys and network files, in the browser
- Builds ties by visible rules, then measures the network and compares it with random networks
- Generates organizations with planted structure, so an analysis can be checked against a known answer
Intended uses: network-analysis instruction, organizational consulting, and analysis of personal data exports.
About 21 million comparisons with networkx, exact calculations and textbook formulas support the calculations for the tested inputs. They do not show that the records identify the right people, that the tie rules are appropriate, or that a ranking supports a claim about someone's importance.
First-round findings
Round 1: 154 issue reports, many about how the network relates to the organization being analyzed
Consultant
A departed Slack account ranked second among brokers without a departure warning in the reported result.
Journalist
Four personal accounts were never joined, so "my network" came out as four islands.
Instructor
The recovery check never appeared, and planted and detected groups were both called "Community".
Accessibility and design
Strong basics, 7 major accessibility issues, and drift from the site's type and buttons.
Departure-status indicators and distinct-contact counts
The same data in two application versions: departure status is visible, and the revised table reports distinct contacts
Known-answer evaluation of the revised interface
[test heading]
[gap]
Design: five synthetic participant profiles, four tasks, and two application versions using fictional organizations with planted answers. Tasks, scoring rules, and predictions were specified before testing. Agents are not people: this shows which paths to a wrong conclusion the interface leaves open, not how much students learn.
Regression check for departure warnings
A deterministic check found three contexts where the departed employee appeared without a status warning
| Where the departed broker is named | 2 Oct | 6 Oct |
|---|---|---|
| People table, her row | not marked | marked |
| Her profile | not marked | marked |
| "Who stands out" (three rankings) | not shown | not marked |
| Brokerage summary sentence | not shown | not marked |
| Note beside the contacts measure | not shown | not marked |
No AI. The fictional organization's data are imported through the application's own interface into two saved, read-only versions. The check examines selected references to the departed employee for an accompanying status marker.
Two omissions beyond the one the agent sessions exposed. The check returns the same answer every time, so it can run after every change. Agents answer a different question: whether a reader reaches the right conclusion.