Security and privacy
The controlled workspace#
Evaluation runs inside a workspace owned by the organizer, with access scoped to the people and roles on that project.
- Organizer-owned project. In the current release each project belongs to one organizer, who owns its decks, evaluations and reports.
- Publish gate. Participants submit through a public page at
/e/<slug>, and that page is public only after the organizer publishes the project. Before that it returns 404 — the address does not resolve to a page for anyone who guesses it.
Access is enforced below the UI#
Access is not a property of the interface. Four guardrails decide what a request can see:
| Guardrail | How it works |
|---|---|
| Session boundary | Sessions live in httpOnly cookies. Client-side scripts cannot read the token directly |
| Database rules | Postgres row-level security. An organizer can read and write only the projects tied to their account |
| Server-only keys | Service-role and AI gateway keys stay on the server. Admin operations run only after an explicit admin check |
| Public gate | The /e/<slug> page opens only after publication; before that, 404 |
The distinction matters when someone asks whether a link can be shared sideways: hiding a button is not access control, and none of the four above is a UI behaviour.
Report delivery#
Reports move through the organizer's workspace rather than through accidental public access. A report reaches someone when the organizer chooses to share it. Participant-facing report sharing is post-MVP — today, delivery to teams is something the organizer does deliberately.
Training and ownership#
- Never trained on. Submissions are processed only for that program's evaluation and are never used to train models. Contractual, not a setting.
- The program owns the record. Reports, scores and the decision log belong to the program. Retention and deletion follow your policy, and a DPA is available.
- Education and institutional use. Student-data handling is structured to support an institution's obligations.
Procurement#
- PO and invoice accepted.
- Vendor registration forms and security questionnaires supported.
- Public sub-processor list, published rather than supplied on request.
- Education discount for university programs.
The question to ask every AI vendor#
A 2026 DataGrail review found that 63.6% of vendors advertising AI never name a third-party AI sub-processor in their legal documents (reported by VentureBeat). That is the gap your security questionnaire exists to close, and it is answerable before a sales call rather than after one.
EvalLens publishes its sub-processors with purpose and processing region for each, alongside the DPA and this page. Bring the questionnaire to the first conversation, not the last.
The human boundary#
AI prepares the analysis. The organizer reviews it, sets the final scores and decides how the report is used — the AI Total Score is a reference and does not rank participants by itself. That is a privacy property as well as a methodology one: the decision that affects a team's funding, selection and reputation is made by a person who can be asked about it.
Next steps#
- Prompt-injection safety — what happens when a deck tries to influence the evaluation.
- What EvalLens does not do — the limits to state before a committee asks.
- Collect submissions — the access settings on the submission page.