Student Privacy

Use student-privacy for a narrow, deterministic check around common student-record terms and student-ID phrases. It can block a request or mark it for redaction, and it has limited non-streaming response handling.

This policy is not a FERPA or COPPA compliance engine. It does not verify identity, parental consent, school context, jurisdiction, or the legal status of a record.

Outcome

Input evaluation

Condition	Verdict	Reason code
No student marker	allow	student_privacy.clean
Marker found, default action: redact	redact	student_privacy.redact
Marker found with action: block	block	student_privacy.block
Marker found with an under-13 age and age_gate: true	block	student_privacy.age_gate

Input details contain action, age_gate, under_13, the first matching keyword_hit if any, and student_id_like.

Output handling

Output behavior depends on action and response transport:

Configuration	Current non-streaming behavior
action: redact	The gateway's response-redaction pass replaces student-ID-like patterns in supported OpenAI JSON fields.
action: block	The gateway blocks supported responses containing student id, transcript, ferpa, or a student ID/identifier/number phrase with student_privacy.block.

An output block returns an HTTP 400 policy-violation response instead of the provider body.

Prerequisites

• Add student-privacy to the effective global or route chain.
• Choose block when delivery must stop on any detected input marker.
• Use redact only after testing the exact identifiers you expect the gateway to replace.
• Combine it with pii-detector or dlp-filter when broader personal or institutional data coverage is required.

Configuration

pack:
  name: student-privacy-example-1
  version: "1.0.0"
  enabled: true

policies:
  chain:
    - student-privacy

policy:
  student-privacy:
    action: redact
    age_gate: true

Supported fields

Field	Type	Default	Notes
action	string	redact	redact or block.
age_gate	boolean	false	When enabled, a detected student marker is forced to block if the first matching inline age expression is from 1 through 12.

Exact input detection

The evaluator joins message text and performs case-insensitive checks for:

Detector	Values
Fixed terms	student id, student_id, transcript, iep, 504 plan, grade, gpa, disciplinary, ferpa
ID phrase	student id, student identifier, or student number, optional :, #, or -, then at least four letters, digits, or hyphens
Inline age	age N or N years old, where N has one or two digits

The age expression does not trigger the policy by itself. It changes the action only when a fixed term or student-ID-like pattern also matched.

Examples:

Text	Default result	With age_gate: true
Help an age 12 learner with fractions.	allow	allow; age alone is not a marker
Student ID: AB-1234	redact	redact
Student ID: AB-1234, age 12	redact	block with student_privacy.age_gate
Send the transcript.	redact	redact

age_gate reads message text only. It does not consult account profiles, headers, verified dates of birth, or consent records, and it has no separate under-18 tier.

Redaction boundary

The input evaluator returns a redaction verdict for any fixed-term or ID match. The subsequent gateway redaction pass has one student-specific pattern: the student ID/identifier/number expression with an identifier of at least four characters.

Consequently:

• Student ID: AB-1234 can be replaced by the redaction pass.
• A keyword-only phrase such as send the transcript produces a redaction verdict but has no student-specific replacement target by itself.
• The policy result does not expose redaction targets; applied replacements are reported by the gateway redaction path.

Do not interpret a redact verdict as proof that every student-related word was removed. Verify the returned/forwarded payload.

For non-streaming output redaction, the same student-ID pattern is added to the gateway response redactor. If it changes a supported response, the observable gateway result uses nested reason pii.detected and final reason redact.applied, even though the selected policy kind can be student-privacy.

Output boundary

The output block detector is intentionally narrower than input detection. It does not currently check iep, 504 plan, grade, gpa, or disciplinary.

Streaming behavior also differs:

• action: redact enables buffered streaming redaction for matching student-ID patterns.
• action: block is not evaluated against accumulated SSE output by the current streaming path.

If output blocking is required, use non-streaming requests and verify stream: false in the client. Input blocking still occurs before either streaming or non-streaming provider calls.

Test input behavior

Create tests/blocks-under-13-student-record.json:

{
  "name": "blocks-under-13-student-record",
  "input": {
    "messages": [
      {
        "role": "user",
        "content": "Student ID: AB-1234 and age 12."
      }
    ]
  },
  "expected": {
    "verdict": "block",
    "reason_code": "student_privacy.age_gate"
  }
}

Use this config while running that fixture:

policy:
  student-privacy:
    action: redact
    age_gate: true

Then run:

kt policy lint --file policy-config.yaml
kt policy test --json

Add at least these companion cases:

• clean education content -> allow / ok;
• a student ID at age 13 -> redact / redact.applied;
• a keyword-only transcript request -> redact / redact.applied; and
• the same marker with action: block -> block / student_privacy.block.

kt policy test exercises the input evaluator. Verify response redaction, response blocking, and streaming behavior through a running gateway.

Rollout checklist

1. Inventory the actual student identifiers and record terms in your traffic.
2. Decide whether a marker should block or merely enter the redaction path.
3. Add positive, negative, age-gated, and substring false-positive fixtures.
4. Test non-streaming responses for both redaction and block behavior.
5. Test stream: true separately if the application enables it.
6. Deploy to a limited route or gateway and correlate representative request IDs with kt events tail --since 10m --json.
7. Review false positives around broad substrings such as grade and gpa.

Troubleshooting

Symptom	Likely cause	Resolution
An age-12 request was allowed	No student marker matched, age_gate was off, or a different earlier age expression was parsed.	Include a supported marker and test the exact message text.
A transcript request says redact but text is unchanged	Keyword detection caused the verdict, but the student redactor only has an ID-like replacement pattern.	Use block if the whole request must stop, or add a separately supported DLP/PII control.
Output with grade was not blocked	Output block detection uses a smaller marker set.	Test against the documented output terms or use another supported output control.
Streamed output was not blocked	action: block is not run against current SSE output.	Require stream: false for output blocking.
Policy never appears in results	It is not in the effective chain or is skipped by conditions/targeting.	Inspect the applied route chain and targeting context.
Team expects FERPA/COPPA certification	The policy is a deterministic text control only.	Complete the broader legal, consent, identity, retention, and operational controls outside this policy.

Next steps

• PII Detector — cover broader personal identifiers
• DLP Filter — add organization-specific sensitive terms and patterns
• Data Policies and Data Routing — route sensitive data to appropriate providers
• Config Testing — build a regression suite around real examples
• Investigate a Blocked Request — correlate a live result with its config version