Safety and Responsible AI Policy

Skyline is designed to help educators apply their own rubric consistently and explain a proposed result. It is not a substitute for professional judgment, student due process, accessibility accommodations, or school policy.

Teachers can inspect criterion evidence, edit feedback, change scores in tenth-point increments, require review, override a result with a reason, and maintain an audit trail. Schools should assign a responsible human owner for any workflow that affects a student record.

• Auto-post is off by default and must be affirmatively enabled by a teacher.
• The teacher chooses the confidence threshold and may disable auto-post at any time.
• Low OCR confidence, hard-to-read regions, or failed applicable verification force individual review.
• Open-ended and uncertain work is routed to stronger evaluation and review paths.
• Teacher calibration is conservative and cannot independently make a grade eligible for auto-post.
• Grades, approvals, overrides, posts, and errors are recorded for accountability.

A technical safeguard reduces risk but does not guarantee correctness. Customers must sample and monitor auto-posted work, especially after changing a rubric, answer key, course, model, or confidence policy.

A defensible suggestion should connect each awarded or withheld point to the teacher's rubric and specific evidence visible in the submitted work. Skyline aims to show: evidence used, why points were awarded, why points were withheld, and what could improve the result.

An explanation is a generated aid, not proof that every inference is correct. When evidence is missing, ambiguous, inconsistent, or inaccessible, the teacher should inspect the original work and record the final rationale.

Skyline must not be used as the sole decision-maker for:

• Suspension, expulsion, discipline, or allegations of misconduct.
• Special-education eligibility, disability diagnosis, accommodation decisions, or mental-health decisions.
• Admissions, scholarships, placement, promotion, graduation, or denial of an educational opportunity.
• Teacher evaluation, employment, credit, insurance, or other legally consequential decisions.
• Academic-integrity findings. Originality indicators are advisory and must be corroborated through a fair human process.

A customer may use Skyline as one source of information in a lawful process only when a qualified human independently reviews the evidence, applicable policy, accommodations, and the student's opportunity to respond.

Handwriting, language variety, disability, scan quality, culturally specific examples, and nonstandard solution methods can affect AI performance. Educators must not treat lower machine confidence as lower student ability.

• Apply approved accommodations and alternate-format rules before grading.
• Review multilingual work with a person competent in the language and subject when needed.
• Accept valid alternate methods when the rubric permits them.
• Periodically compare override and error patterns across relevant groups where lawful and appropriate.
• Provide a non-AI review path when a student or family raises a credible concern.

• Upload only the pages and information needed to grade the assignment.
• Remove unnecessary identifiers and highly sensitive information before upload.
• Confirm roster matches and student names before posting or sharing results.
• Send private family links only to the intended recipient through a school-approved channel.
• Do not post a family link publicly or forward it beyond the intended household.
• Use print packets, exports, and downloaded files as education records and store or dispose of them under school policy.

Users may upload only supported assignment files and must not upload executable code, malware, deceptive files, or content intended to attack the service. Users must use their own accounts, protect devices and credentials, sign out from shared devices, and report suspicious access promptly.

Schools should tell students when AI materially assists grading if required by policy or law. Students should have a clear way to ask what evidence was used, request correction of a transcription or roster mismatch, and obtain meaningful human review of a disputed result.

Correcting a score should not erase the original decision. Skyline preserves approval and override activity so the school can understand what changed and why.

A suspected privacy, security, or grading-safety incident should be reported promptly. We triage reports based on data exposure, number of affected people, severity of educational impact, exploitability, and whether an unsafe action is continuing.

We may temporarily disable auto-post, sharing, integrations, or affected features while investigating. We will preserve relevant audit evidence, correct unsafe behavior, notify affected customers as required, and document prevention steps.

Report suspected unauthorized access, exposed records, unsafe grading behavior, or a security vulnerability to security@silviaai.dev. Include the affected feature, date, steps to reproduce, and non-sensitive evidence. Do not include real student work unless we request it through an approved channel.

Good-faith security research must avoid accessing or changing another person's data, disrupting service, social engineering, automated high-volume testing, or publicly disclosing an unresolved vulnerability. We will acknowledge legitimate reports and coordinate remediation and disclosure in good faith.

We may restrict or suspend use that creates material risk, violates law, exposes student data, bypasses safeguards, or conflicts with this policy. We may update safeguards as threats, models, and law change. Material changes will be posted with a revised date and additional notice where required.