Data-flow sheet: Standing
1. What data exists and where it goes
| Data element | Read in browser | Sent to Common Benchmark | Sent to AI provider (opt-in step only) | Stored anywhere by us |
|---|---|---|---|---|
| Student name, LMS ID, SIS ID, email, section | Yes | Never | Never (replaced by "Student NN" before the request) | No |
| Per-student scores and totals | Yes | Never | Points for each sent answer, keyed to the placeholder | No |
| Question text and points possible | Yes | Never | Yes, for the questions the instructor selects | No |
| Student written answers | Yes | Never | Yes, de-identified, for the questions the instructor selects | No |
| Multiple-choice selections | Yes | Never | Never | No |
| The uploaded file itself | Yes (in memory) | Never | Never | No |
| Instructor's institution selection | Yes | Not in the beta (no accounts) | Never | No |
| Instructor's email (beta sign-in) | — | Cloudflare Access sign-in log (email, time, IP) | Never | Access logs, per Cloudflare retention |
"Never" means the application contains no code path that transmits the element; the request payload to the AI step is validated server-side and rejected if any student field is not a placeholder of the form Student NN.
2. The two steps
Step A — Score analysis (default). The instructor drops in a graded quiz export (Canvas Classic "Student Analysis" CSV, or any CSV with a column mapping). Parsing, pseudonymization, per-question statistics, and per-student standing run entirely in the browser (JavaScript, no server round-trip). The page can be printed; the printout carries a header stating that no student data was sent. Closing the tab discards everything.
Step B — Written-answer analysis (opt-in, per quiz). After reviewing the data-handling notes on screen and confirming that AI-tool use is permitted at their institution, the instructor selects which written questions to send. The browser builds a payload containing only: question text, points possible, placeholder student labels, answer text, and the instructor's score for each answer. Emails, phone numbers, and long digit strings inside answer text are additionally masked. The payload is sent over TLS to Common Benchmark's API, which validates it and forwards it to the AI model provider. The provider returns structured analysis (concepts missed, similar answers scored differently, ambiguous questions), which is passed back to the browser and displayed against the local placeholder-to-name lookup. Nothing is written to disk on our side.
3. Providers and locations
| Provider | Role | What it sees | Retention / training |
|---|---|---|---|
| Cloudflare | Hosting (Workers), TLS, and beta access control (Access, one-time email code) | Application files; request metadata; sign-in email for the beta allowlist; the opt-in payload in transit | Standard edge logs; no payload storage by our configuration |
| Anthropic (API) | AI model for the opt-in step | De-identified payload described in Step B | Not used for model training; retained up to 30 days for abuse monitoring, then deleted ◇ (verify against the provider's current API data terms before relying on this line) |
| Common Benchmark | Operator | Nothing persisted from either step; support email if the instructor writes to us | No student data at rest, ever, in the beta |
4. Controls the instructor sees
- A "here's what we read" confirmation screen before any analysis, including which student columns were detected and replaced.
- The institution's published AI policy, quoted verbatim with source link and date checked, shown on the import screen — as information, not as a determination that use is permitted.
- The opt-in step is off by default; it requires selecting questions and ticking a statement that includes "I've checked that this is allowed where I teach."
- Print/PDF output only; no data-sharing between users in the beta; the "ask your school to approve" email is sent by the instructor from their own address.
5. Security notes
- All traffic is HTTPS. The beta hostname sits behind Cloudflare Access; the API additionally verifies the Access token on each request, so it cannot be called without a signed-in allowlisted user.
- No database, no object storage, no analytics trackers in the beta. The only browser storage used is the appearance (light/dark) preference.
- Payload limits and validation are enforced server-side; the code is small enough to read in one sitting and is available to reviewers on request.
- Vulnerability reports: security@commonbenchmark.com.
6. Regulatory framing
FERPA. In Step A, education records never leave the instructor's device and no third party receives them. In Step B, the instructor transmits de-identified answer text; whether de-identified student work may be sent to a third-party AI tool is governed by the institution's own policy, which is why the application asks the instructor to confirm it. When an institution formally adopts the tool, Common Benchmark will sign the institution's data-processing agreement and act as a school official under its direct control.
Accessibility. WCAG 2.1 AA self-assessment and a VPAT/ACR are in preparation for the beta. HECVAT. A HECVAT Lite self-assessment is in preparation; the answers to the AI section follow from this sheet.