Who sees what
The rule that cannot be broken: students never read mentor feedback, scores or the agreed requirements. It must hold even against a logged-in student calling the tool's API directly, not only in the screens.
Roles
| Role | Can do |
|---|---|
| Customer | Submit and revise their own proposal, reply to review comments, see their team's reports and milestones, sign off reports, give feedback |
| Student | See all project briefs, rank preferences, submit their team's weekly reports, milestones and requirements, submit peer assessments |
| Mentor | See teams' work and every team's briefing, fill in weekly check-ins, score every team at presentations, approve the feedback sent to their own team |
| Coordinator | Everything, including setup |
| Course manager | Read everything; decide grades |
| AI service account | Read what a playbook needs; write only AI outputs and proposal comments; cannot write scores or grades |
Students and customers use simple portal pages; they cannot open the staff screens at all.
Records
| Record | Written by | Read by | Hidden from students |
|---|---|---|---|
| Project proposal, with comment thread | Customer; staff and AI comment | That customer, staff | Yes |
| Project brief | Staff | Everyone | No |
| Sealed specification | Staff with the customer | That customer, staff | Yes |
| Project preference and avoid list | Each student | That student, coordinator | From other students |
| Team, weekly report, sign-offs, milestones, requirements, dependency log | Team (sign-offs: mentor and customer) | Team, staff, its customer | No |
| Mentor check-in | Mentor | That mentor, coordinator, course manager | Yes |
| Presentation score | Each mentor | That mentor only, coordinator, course manager | Yes |
| Peer assessment | Each student | That student, the team's mentor, coordinator | From other students |
| Customer feedback | Customer | Staff | Yes |
| AI output (digest, briefing, coverage report, decision sheet) | AI service account | Staff | Yes |
| Merged feedback, grade records | AI draft, mentor approves; mentor recommends, course manager decides | The team's mentor, coordinator, course manager | Yes |
Confidential data always has its own record type that students cannot open at all. Hiding a single field inside a record a student can read was found to be unreliable in the trials, so the design never relies on it.
How it is tested
A scripted confidentiality test runs at every one of the 14 checkpoints, logged in as every person in the cohort. Its expectations are written from the table above, independently of the tool's permission settings, so a wrong setting shows up as a difference.
| Part | What it checks |
|---|---|
| Settings audit | Every permission row matches the design flag by flag (a new row in Frappe allows writing by default; the audit catches that), no student or customer role on confidential record types, strict restrictions switched on |
| Visibility | For every person and every record type, the records the API returns are exactly the expected ones. Extra is a leak; missing is broken access |
| Probe battery | About 30 routes per forbidden record: lists, single records, version 2 API, counts, report views, group-by, search, link checks, linked documents, joined columns, comments, change history, file attachments, activity log, print view, PDF, export |
| Forbidden writes | The AI account writing scores or grades or editing a report; a student forging another team's report or signing off its own; a mentor changing another mentor's score or setting a final grade; a customer editing the sealed specification; a student writing the customer's decision |
| Dashboard reports | The dashboard's reports read the database directly, so they are refused to everyone except the coordinator and the course manager |
Result at the final checkpoint: all 14 checkpoints passed in one rebuild from scratch. At the final checkpoint: 49 active people x 23 record types, 1,564 records, 9,088 probe requests and 17 forbidden writes, with no leak and every write refused.
Known and accepted: a person can see that hidden records linked to a team exist ("1 hidden"), but nothing about them.