Back to home
TRUST & SECURITY

Confidence through clarity.

The Cyrine Trust Center. What we have checked, what we are improving, and what still needs evidence.

Evidence reviewed Manually maintained

In Progress

Security

Account and ownership checks restrict access to class data and private teaching materials. Recovery backups are encrypted. These safeguards are tested within a limited scope, not independently certified.

How it works & limitations for Security

Who can access data

Professors can access their own teaching materials and class responses. Student access and unrevealed answer keys are restricted in the tested application paths. Authorized service administrators can access data for operations and recovery; private storage does not exclude administrators.

Public evidence

This Trust Center publishes reviewed summaries, not student records, credentials, private system logs or detailed access configurations. Independent security assessment and full operational access review remain incomplete.

Verified

Specific checks

Automated checks cover professor ownership, student access boundaries and unpublished answer keys. Application database tables were confirmed to have Row Level Security enabled in the October 12 review.

Source tests and read-only configuration review · October 11–12, 2026

Still to verify

These checks are not an independent penetration test. Organization access, credential custody and ongoing operational controls need further verification.

In Progress

Privacy

Cyrine keeps account information, class participation and submitted answers to operate teaching sessions and analytics. Student records use internal IDs; optional display names are stored separately. IDs still represent people and are not anonymous data.

How it works & limitations for Privacy

Student and faculty information

Retained records include account and sign-in information, course and lecture details, participation times, selected answers and answer times. Professor accounts use email. Optional student display names support the professor’s class view; names are not required for analytics. The current Canvas-linked account flow still requires a verified-email Cyrine account.

Use and sharing

The owning professor can review participation and answers. Quality-assurance reporting is aggregate-first, but small groups may still allow someone to infer an individual’s activity. Supabase provides account, database and private-file services; hosting providers process delivery and request metadata. GitHub holds encrypted recovery archives. OpenAI processes the educational inputs described below.

Operational information

Application diagnostics are designed to record outcomes, times and request references rather than teaching content. Authentication and hosting providers may retain IP addresses, browser details and access logs. Their actual retention settings have not all been verified.

Verified

Specific checks

Request-boundary tests confirm that structured student identities, submitted answers, analytics and account profiles are excluded from the inspected AI generation requests. Selected lecture text and approved learning-outcome descriptions are used instead.

AI request construction tests and code review · October 11, 2026

Still to verify

Names or sensitive information can still appear inside teaching content. Audio is sent for transcription without automatic voice anonymization. Professors must review content; provider retention settings and complete legacy-data cleanup are not independently verified.

In Progress

Data Handling

Saved teaching content includes extracted text, filenames, page references, questions, outcomes, recordings and transcripts. Automatic age-based deletion is not active. Removing material does not erase every related assessment record or copy.

How it works & limitations for Data Handling

Documents, recordings and storage

The normal PDF, PowerPoint and Word flow extracts text in the browser and saves selected text and source references, not the original document. The last identified legacy original document was removed from active Storage on October 13 after backup and extracted-text verification. This does not erase copies in backups, the recovery environment or browsers. Recordings and transcripts may be stored in private cloud storage or the database, with working copies in the browser. Voices, syllabus text and transcripts may contain personal information.

What material deletion does

Professors can explicitly delete an owned stored material. The implemented flow removes its stored text versions and handles associated original-file cleanup, subject to shared references or retention holds. Questions, answer records, participation and learning-outcome links are preserved. Reused material copies and text quoted in questions are not automatically erased.

Retention is still being implemented

The approved retention periods are targets, not active expiry guarantees. Accounts, responses, materials, recordings, transcripts and historical versions can remain until a supported deletion or authorized cleanup occurs. Ending a lecture does not erase its records. Deleting a course can leave lectures, and removing a recording can leave historical audio or transcript copies. Account-wide erasure is not a self-service feature.

Backups and other copies

Encrypted backup artifacts have a configured rolling 14-day retention. Deletion from active storage does not rewrite existing archives. A restricted recovery environment also retains an earlier real-data copy; its cleanup remains outstanding. Restores require later deletion decisions to be reapplied, and automatic replay is not yet implemented. Offline browser copies, professor downloads and provider-held logs require separate handling; immediate erasure everywhere is not guaranteed.

Verified

Specific checks

Synthetic tests covered professor ownership, repeated material deletion, shared-reference and hold checks, preservation of assessment records, and prevention of stale re-upload under the same material ID.

Data inventory and material-deletion implementation review · October 11–12, 2026

Still to verify

Timed expiry, notices, full browser-cache cleanup, recovery-copy cleanup and complete restore-aware deletion remain unfinished. No new deletion or retention job is activated by this disclosure.

In Progress

AI Processing

OpenAI helps generate questions, suggest learning outcomes and transcribe audio. Reviewed AI requests exclude structured student identities, submitted answers, analytics and account profiles. Sensitive information inside teaching text or audio can still be transmitted.

How it works & limitations for AI Processing

What goes to OpenAI

Generation sends the professor’s prompt or selected lecture excerpts, requested question count, approved learning-outcome descriptions and, when needed, previous question prompts to reduce repetition. Syllabus analysis sends extracted syllabus excerpts. Transcript-based generation sends reviewed transcript text. Original PDF, PowerPoint and Word files are not sent through these reviewed paths. Responses include proposed questions, choices, correct answers, explanations, outcome suggestions or transcript text.

Recordings need special care

Transcription sends the raw audio. It can contain recognizable voices, spoken names, student contributions and background conversation. Text filtering cannot anonymize that audio. Professors should obtain any required permission, use appropriate teaching content, and remove sensitive passages before sending recordings or text. Generated questions and transcripts also need professor review for errors.

Provider retention and training

OpenAI states that API data is not used for training by default unless an account opts in. Chat abuse-monitoring logs may retain content for up to 30 days, with legal or safety exceptions allowing longer retention. Its transcription endpoint table lists no abuse-monitoring or application-state retention. This does not erase Cyrine’s own recordings or transcripts. Disabling stored chat responses does not establish Zero Data Retention. Cyrine’s account-specific sharing, retention and regional settings have not been fully verified; no zero-retention or data-residency guarantee is claimed.

OpenAI API data controls

Verified

Specific checks

Provider-boundary tests verified the educational request fields and excluded structured student identities, answers, analytics, account data and unnecessary course identifiers from inspected requests.

Application AI review · October 11; provider documentation checked October 12, 2026

Still to verify

Filtering is best effort, not complete anonymization. Unlabeled names and sensitive narrative can survive. Deleting a Cyrine record does not itself erase provider-retained data.

In Progress

Accessibility

Designed for access, with automated checks and further human testing needed.

Evidence & remaining work for Accessibility
Verified

Specific checks

The October 13 source audit checked 18 static or synthetic interface states with no reported violations in configured WCAG A/AA rules. Eighteen sampled design-token color pairs and seven synthetic focus and interaction checks passed. One radio-group check remains incomplete.

Static and synthetic checks, not a full browser audit · October 13, 2026

Still to verify

Manual keyboard, screen-reader, mobile, zoom and real classroom workflows still require assessment. Cyrine does not claim WCAG 2.2 AA conformance or accessibility certification.

In Progress

Reliability

Encrypted backups and recovery checks, with honest recovery limits.

Evidence & remaining work for Reliability
Verified

Specific checks

The October 13 scheduled monitor passed 11 public endpoint checks and backup freshness. The scheduled encrypted backup and its off-site artifact were confirmed. An October 8 hosted drill demonstrated limited database and synthetic-file recovery in 43.352 seconds, excluding project setup and a complete service rebuild.

Dated workflow evidence and isolated recovery tests · October 8–13, 2026

Still to verify

The latest archive has not had a full hosted restore. Complete service recovery and alert delivery remain unverified. Daily backups are a target, not a guaranteed maximum data-loss window. No uptime or full-service recovery-time guarantee is claimed.

In Progress

Canvas Integration

LTI foundations are tested. Institution-specific sandbox validation comes next.

Evidence & remaining work for Canvas Integration
Verified

Specific checks

Synthetic LTI 1.3 tests cover launch validation and role/course boundaries. Roster-service readiness and grade-passback planning have automated tests. Registration configuration endpoints were checked in the documented review.

129 focused synthetic integration tests and public backend endpoint checks · October 13, 2026 UTC

Still to verify

Genuine Canvas sandbox launches, registered key-URL reachability and institution approval remain pending. Live roster synchronization and grade passback are not verified. Grade passback remains a dry-run design. Dynamic Registration and Deep Linking are not implemented in this pilot profile. No Canvas partnership or 1EdTech certification is claimed.

Not Assessed

Educational Evidence

Learning impact needs classroom evidence, beyond software tests.

Evidence & remaining work for Educational Evidence

No learning-impact study available for this review

Evidence pending

No evaluated evidence is available here for improved learning, attainment or accreditation outcomes. Classroom pilots and a defined evaluation method are needed before making those claims.

READ-ONLY EVIDENCE

Live Verification

Published summaries

Results from existing checks, with their scope and limits. This page reads approved summaries every five minutes; it never runs security or load tests. Summaries are published after review, not streamed directly from test systems. A source-test pass applies to its listed revision, not automatically to the current deployment.

Loading approved summaries…

Security audit

Unavailable

Dependency advisory scan · source

No approved result available.

Checked
Unavailable
Tested version
Source revision unavailable
Scope & limitations for Security audit

Known package advisories only. Not a penetration test or a deployed-system scan.

Freshness window: 48 hours after the check. A fresh result still applies only to its tested version and scope.

Automated tests

Unavailable

GitHub regression suite · source

No approved result available.

Checked
Unavailable
Tested version
Source revision unavailable
Scope & limitations for Automated tests

Synthetic and source tests. The tested merge revision is not an attestation of the current deployment.

Freshness window: 24 hours after the check. A fresh result still applies only to its tested version and scope.

Availability monitoring

Unavailable

11 public endpoint checks + 1 backup-metadata check

No approved result available.

Checked
Unavailable
Tested version
Source revision unavailable · deployed version unverified
Scope & limitations for Availability monitoring

Expected authentication denials count as successful boundary checks. No authenticated classroom journeys, uptime percentage, restore guarantee or verified alert delivery. Deployed version was not attested.

Freshness window: 90 minutes after the check. A fresh result still applies only to its tested version and scope.

Accessibility checks

Unavailable

Trust Center static HTML · disclosures open and closed

No approved result available.

Checked
Unavailable
Tested version
Source revision unavailable
Scope & limitations for Accessibility checks

This result covers the Trust Center only. No real browser, screen-reader or mobile checks. Color contrast was excluded from this axe run; no WCAG conformance claim.

Freshness window: 168 hours after the check. A fresh result still applies only to its tested version and scope.

Synthetic classroom tests

Unavailable

Isolated local database · synthetic students

No approved result available.

Checked
Unavailable
Tested version
Source revision unavailable
Scope & limitations for Synthetic classroom tests

Single-connection PGlite, not production load or network latency. Supabase Realtime transport and larger groups were not tested. The listed source revision and migration fingerprint identify the dated local test inputs, not the current deployment.

Freshness window: 168 hours after the check. A fresh result still applies only to its tested version and scope.

ACCESSIBILITY & INSTITUTIONAL READINESS

Evidence for a considered decision.

In Progress

Internal evidence reviewed October 13, 2026. These dated assessments support review; they do not establish independent certification, WCAG conformance or institutional approval.

Accessibility evidence

18 static or synthetic interface states reported zero automated violations in configured rules, with one radio-group check incomplete. Eighteen sampled design-token color pairs and seven synthetic interaction checks passed.

October 13 source evidence. Browser-rendered contrast, real assistive technology and complete authenticated workflows were not tested.

Open findings & remediation progress

Changes were implemented for 14 of 15 findings, including focus, labels, contrast, announcements and dialog controls. Implementation is not closure: manual validation remains pending.

Content-language metadata (A13) remains open. Priority review includes live-question announcements, audio/transcript equivalence and busy-dialog exit. Screen readers, keyboard workflows, mobile, zoom and actual Canvas use still need assessment.

The draft includes all 15 findings and a 55-criterion review ledger. Product-wide conformance ratings remain undetermined.

Canvas & institutional readiness

LTI launch validation and hosted registration configuration have dated internal evidence. Actual instructor and student Canvas sandbox launches remain unverified.

Integration status & technical summary

LTI 1.3: synthetic launch and boundary tests; real Canvas acceptance pending. NRPS: tested source adapter, no live roster synchronization. AGS: dry-run planning only; live grade passback disabled.

The report summarizes architecture, identity mappings, data boundaries, operations, recovery limits and administrator handoff. Institution-specific configuration and sensitive technical records remain restricted.

Institutional review, manual accessibility assessment, full recovery evidence and alert-delivery verification remain open. No ASU approval, Canvas partnership or 1EdTech certification is claimed.

Reports use readable, self-contained HTML with headings, labelled tables and keyboard focus styles. Browser-generated PDFs are not assessed. No independent evaluation or institutional decision is implied by an internal test result.