Imported from Nmor/the-claude-council (skills/edtech-patterns/SKILL.md). Install upstream with npx skills add Nmor/the-claude-council --skill edtech-patterns. Copyright stays with the author.

EdTech Platform Patterns

Auto-fires on every file. Sister skills: ferpa-coppa-compliance (regulation), wcag-accessibility (the learner is the user), interaction-design (learning experience design), gdpr-ccpa-compliance (when learners are EU/CA minors), prompt-engineering + rag-design (when AI tutors are involved).

Purpose

Engineers building learning platforms (LMSs, MOOCs, K-12 tools, corporate training, adaptive tutors, assessment engines, credentialing systems) routinely re-invent interoperability, ship inaccessible-to-learners content, mis-model assessment, or import learning-analytics dashboards that quietly profile minors without parental knowledge. The cost of getting it wrong is high: a single integration break (Canvas → publisher → SIS) can take an entire district offline at exam time; an inaccessible assessment locks blind students out of their education; a biased proctoring system targets disabled or non-white learners disproportionately and lands the vendor in DOJ + OCR complaints.

This skill exists to make the standards-cited, accessibility- first, ethics-aware path the default — and to name the anti-patterns that turn well-intentioned edtech into discriminatory infrastructure.

Not legal advice; not pedagogical advice for any specific learner. Pair with district / institution / accreditor counsel on Section 504, IDEA, OCR + DOJ Title II/III obligations, and with learning scientists on assessment validity.

Standards Cited

Interoperability (1EdTech / IMS Global)

LTI 1.3 + LTI Advantage (Learning Tools Interoperability — 1EdTech Final Spec) — OAuth 2.0 + OIDC + JWT-based launch, replaces LTI 1.1's shared-secret model; Names + Role Provisioning (NRPS), Assignment + Grade Services (AGS), Deep Linking 2.0. LTI 1.1 / 1.2 deprecated; new integrations MUST be 1.3+.
OneRoster 1.2 (1EdTech) — Roster sync for SIS ↔ LMS ↔ tool; users / classes / enrollments / courses / academic sessions / demographics; CSV and REST API bindings; replaces PowerSchool's PSCB SIF where possible.
Caliper Analytics 1.2 (1EdTech) — Learning event vocabulary (Sensor API + Event model); contrasts with xAPI in that Caliper is opinionated about LMS-flavoured events while xAPI is open vocabulary.
QTI 3.0 (Question + Test Interoperability — 1EdTech) — XML-based item + test format, replaces QTI 2.x; widely used in K-12 + higher-ed assessment.
Common Cartridge 1.3 (1EdTech) — Package format for course content (HTML + LTI links + assessment + discussion-forum + web-link); the IMS Thin Common Cartridge variant is the practical floor.

Experience tracking

xAPI 2.0 / IEEE 9274.1.1-2023 — Statement-based "Actor → Verb → Object" event model with LRS (Learning Record Store) backing; supersedes SCORM for tracking learning outside the LMS (simulations, VR, mobile, workplace).
cmi5 (ADL) — xAPI profile that defines a SCORM-replacement contract (course structure + AU launch + completion + satisfaction); the deployable answer when an org wants xAPI benefits with SCORM-like packaging.
SCORM 2004 4th Edition (ADL) — Legacy but still dominant in corporate L&D; complete in scope (sequencing + navigation + rollup); strict packaging via Content Aggregation Model.
SCORM 1.2 (ADL) — Older but still ubiquitous; simpler (no sequencing); minimum support for any platform that ingests external content.

Credentialing

Open Badges 3.0 / W3C Verifiable Credentials (W3C Recommendation, 2023+ alignment) — Cryptographically signed credentials with JSON-LD; replaces Open Badges 2.0's baked-image model with a VC envelope; aligns with EBSI (European Blockchain Services Infrastructure) and CLR (Comprehensive Learner Record).
CLR Standard 2.0 (1EdTech) — Verifiable record of competencies + achievements; envelope for Open Badges + skill
- assertion data.

Accessibility (learner-specific)

WCAG 2.2 AA + AAA (W3C, Oct 2023) — Floor for any learning surface. AAA recommended for assessment paths (the educational record depends on the learner being able to demonstrate knowledge, not on whether the UI happens to work).
AccessForAll 3.0 (ISO/IEC 24751) — Personal needs + preferences (PNP) framework; lets learners declare accommodations (preferred-modality, captions, audio description, alt-text density, signing avatars, reading speed) and the system delivers matching DRD (Digitally Resource Description) variants.
Section 504 (29 USC §794 + 34 CFR Part 104) — US prohibition on disability discrimination in federally-funded programs; reasonable accommodations required.
IDEA (Individuals with Disabilities Education Act, 20 USC §1400+) — IEP (Individualized Education Program) + 504 Plan compliance for K-12.
DOJ Title II Web Accessibility Final Rule (April 2024, effective dates 2026-2027) — Mandates WCAG 2.1 AA for state + local government public-facing AND learner-facing digital content; bigger districts (≥50,000) effective Apr 24, 2026.
EAA (European Accessibility Act 2019/882, effective Jun 28, 2025) — Mandates WCAG 2.1 AA for ed-tech sold to EU.

Pedagogy + Universal Design

UDL 3.0 (Universal Design for Learning — CAST) — Multiple means of engagement / representation / action + expression; the framework that turns "accessibility add-on" into "designed-in flexibility".
IRT 2PL / 3PL / Rasch (Item Response Theory) — Psychometric foundation for adaptive testing; replaces "raw score / total" with item-difficulty + learner-ability latent estimation.
Bloom's Revised Taxonomy (Anderson + Krathwohl, 2001) — The cognitive-process dimension that learning objectives align to.

Proctoring / Integrity

NCME Standards for Educational + Psychological Testing (AERA / APA / NCME, 2014 + supplements) — Validity + reliability + fairness floor.
ABA Model Rules (for legal-ed remote bar exam) — Strict identity-verification + monitoring rules adopted post-COVID.

When to Fire

File globs:

**/lti/**, **/*lti*, **/launch.{ts,py,rb,go}, **/jwks*
**/scorm/**, **/imsmanifest.xml, **/cmi5*.json
**/xapi/**, **/*xapi*, **/lrs/**, **/statements/**
**/oneroster/**, **/csv-1.2/** (OneRoster CSV format)
**/caliper/**, **/sensor*.{ts,py,rb,go}
**/qti/**, **/*qti*, **/assessment-item-*.xml
**/badge/**, **/credential/**, **/clr/**, **/openbadges/**, **/verifiable-credentials/**
**/proctor*, **/proctoring/**, **/respondus/**, **/proctortrack/**, **/proctoru/**
**/grade-passback/**, **/grade-sync/**, **/ags/**
**/sis-sync/**, **/clever*/**, **/classlink*/**
**/canvas-api/**, **/schoology-api/**, **/moodle-api/**, **/blackboard-api/**, **/d2l-api/**, **/brightspace-api/**
Per-platform: **/lti-tool-provider*, **/lti-platform*

Keyword triggers (in diff, ticket, or prompt):

"LTI", "LTI 1.3", "LTI Advantage", "Deep Linking", "Names and Roles", "Assignment and Grade Services", "NRPS", "AGS"
"xAPI", "Experience API", "TinCan", "LRS", "Learning Record Store", "cmi5", "AU launch"
"SCORM", "SCORM 1.2", "SCORM 2004", "imsmanifest"
"Caliper", "Sensor API", "Caliper event"
"QTI", "assessment item", "test specification"
"Common Cartridge", "thin cartridge"
"OneRoster", "roster sync", "Clever sync", "ClassLink"
"Open Badges", "Verifiable Credential", "CLR", "Comprehensive Learner Record"
"proctoring", "remote proctor", "AI proctor", "ID verify", "session recording", "exam integrity"
"adaptive learning", "CAT", "computerized adaptive test", "IRT", "Rasch", "ability estimation"
"UDL", "Universal Design for Learning", "accommodations", "IEP", "504 plan", "AccessForAll"
"MOOC", "LMS", "VLE", "LXP", "SIS", "Student Information System"
"grade passback", "outcome service", "result service"
"gradebook", "transcript", "academic record"
"learning analytics", "early warning", "predictive analytics", "at-risk model"

Scope triggers:

New integration with Canvas / Schoology / Moodle / Blackboard / D2L Brightspace / Google Classroom / Microsoft Teams for Education / Echo360 / Panopto
Any system handling K-12 grades, IEPs, 504 plans, attendance
Any assessment with academic-record consequence
Any system targeting minors (ALWAYS auto-engage ferpa-coppa-compliance skill alongside)
Adaptive / personalized learning algorithms
Credential issuance (degrees, certificates, badges, micro-credentials)
AI-tutor / AI-grader / AI-proctor (adds ai-ethics-reviewer)

Core Patterns

Pattern 1: LTI 1.3 launch — OIDC + JWT, never the LTI 1.1 shared secret

LTI 1.1 / 1.2 used OAuth 1.0a HMAC-SHA1 with a shared secret. LTI 1.3 replaces this with OpenID Connect + asymmetric JWT. Every new tool integration MUST be 1.3+; legacy 1.1 only acceptable when the platform (LMS) doesn't yet support 1.3 AND the integration is documented as transitional.

LTI 1.3 launch flow (high level):

Resource link request — LMS sends iss, login_hint, target_link_uri, lti_message_hint, lti_deployment_id to the tool's OIDC login init endpoint.
OIDC auth request — Tool redirects to the LMS auth endpoint with response_type=id_token, scope=openid, nonce, state, prompt=none, plus the LTI-specific login_hint echo.
ID Token (LTI launch JWT) — LMS POSTs a JWT to the tool's target_link_uri. JWT is signed with the LMS's RSA private key; tool verifies via the LMS's published JWKS endpoint.
Claims validation:
- iss matches the registered platform
- aud matches the tool's client_id
- nonce matches the value sent in step 2 (replay protection)
- iat within ±5 minutes (clock skew tolerance)
- https://purl.imsglobal.org/spec/lti/claim/message_type is LtiResourceLinkRequest (or LtiDeepLinkingRequest)
- https://purl.imsglobal.org/spec/lti/claim/version is 1.3.0
- https://purl.imsglobal.org/spec/lti/claim/deployment_id matches the registered deployment

Reference implementation sketch (TypeScript):

import { jwtVerify, createRemoteJWKSet } from 'jose';

const jwks = createRemoteJWKSet(new URL(platform.jwksUri));

const { payload } = await jwtVerify(launchJwt, jwks, {
  issuer: platform.issuer,
  audience: tool.clientId,
  algorithms: ['RS256'],
  clockTolerance: '5m',
});

// Validate LTI-specific claims
if (payload['https://purl.imsglobal.org/spec/lti/claim/version'] !== '1.3.0') {
  throw new LtiError('lti_version_unsupported');
}
if (payload['https://purl.imsglobal.org/spec/lti/claim/deployment_id'] !== platform.deploymentId) {
  throw new LtiError('lti_deployment_mismatch');
}
// Nonce verification: look up `payload.nonce` in your one-time-use store; reject if reused
const nonceRecord = await nonceStore.consumeOnce(payload.nonce);
if (!nonceRecord) throw new LtiError('lti_nonce_replay_or_missing');

Per ferpa-coppa-compliance.md, NEVER ship LTI tools that default to "public roster"; respect the NRPS scope only when the platform admin has granted it, and treat the roster as educational records.

Pattern 2: LTI Advantage services — Names + Roles, Assignment + Grade, Deep Linking

LTI Advantage builds three services on top of 1.3:

NRPS (Names + Role Provisioning Services): fetch the context's member list (course roster). Scope: https://purl.imsglobal.org/spec/lti-nrps/scope/contextmembership.readonly. Returns members with user_id, name, email, roles (http://purl.imsglobal.org/vocab/lis/v2/membership#Learner, Instructor, ContentDeveloper, etc.).
AGS (Assignment + Grade Services): create line items in the LMS gradebook + post scores. Three sub-scopes: lineitem read/write, results read, score post.
Deep Linking 2.0: lets the tool return content selections to the LMS in a JWT signed by the tool; LMS embeds the selected items.

For each service:

Tool requests an access token from the platform's token endpoint using client_credentials + a signed client_assertion JWT (assertion JWT signed by the tool's private key; platform verifies via tool's published JWKS).
Platform returns a short-lived bearer token (typically 1 hour TTL).
Tool calls the AGS / NRPS REST endpoint with the bearer token in Authorization: Bearer <token>.

Tokens MUST be cached + reused until near expiry; minting a new token per request thrashes both sides.

Pattern 3: SCORM packaging + the manifest contract

SCORM packages are ZIP files containing imsmanifest.xml at the root. The manifest describes the content organization, the resources, and the SCORM CAM (Content Aggregation Model) metadata.

SCORM 1.2 manifest minimum:

<?xml version="1.0" encoding="UTF-8"?>
<manifest identifier="MANIFEST-1" version="1.2"
  xmlns="http://www.imsproject.org/xsd/imscp_rootv1p1p2"
  xmlns:adlcp="http://www.adlnet.org/xsd/adlcp_rootv1p2">
  <metadata>
    <schema>ADL SCORM</schema>
    <schemaversion>1.2</schemaversion>
  </metadata>
  <organizations default="ORG-1">
    <organization identifier="ORG-1">
      <title>Course Title</title>
      <item identifier="ITEM-1" identifierref="RES-1">
        <title>Module 1</title>
      </item>
    </organization>
  </organizations>
  <resources>
    <resource identifier="RES-1" type="webcontent"
      adlcp:scormtype="sco" href="index.html">
      <file href="index.html"/>
    </resource>
  </resources>
</manifest>

Runtime API:

SCORM 1.2: LMSInitialize, LMSGetValue("cmi.core.lesson_status"), LMSSetValue, LMSCommit, LMSFinish.
SCORM 2004: Initialize, GetValue("cmi.completion_status"), SetValue, Commit, Terminate + sequencing extensions (cmi.exit, cmi.suspend_data, cmi.location).

Common pitfalls:

suspend_data ≤ 64 KB (SCORM 2004); content authors routinely exceed this on save → silent data loss. Compress + bound at the authoring tool level.
Score scaling: SCORM 1.2 uses 0-100; SCORM 2004 uses cmi.score.scaled (−1 to 1) + cmi.score.raw. Mixing conventions causes gradebooks to show 0.83 when the learner earned 83.
lesson_status vs completion_status + success_status — SCORM 1.2 mashes them; 2004 splits. Map carefully when importing 1.2 content into a 2004-native player.

Pattern 4: xAPI 2.0 — Actor / Verb / Object statement model

xAPI describes any learning event as {actor, verb, object, result?, context?, timestamp, authority, version}. Statements flow to a Learning Record Store (LRS) over HTTPS POST /xAPI/statements.

Canonical statement:

{
  "actor": {
    "objectType": "Agent",
    "account": {
      "homePage": "https://lms.example.edu",
      "name": "user-abc123"
    }
  },
  "verb": {
    "id": "http://adlnet.gov/expapi/verbs/completed",
    "display": { "en-US": "completed" }
  },
  "object": {
    "objectType": "Activity",
    "id": "https://courses.example.edu/biology-101/module-3",
    "definition": {
      "name": { "en-US": "Cellular Respiration" },
      "type": "http://adlnet.gov/expapi/activities/module"
    }
  },
  "result": {
    "completion": true,
    "success": true,
    "score": { "scaled": 0.87, "raw": 87, "min": 0, "max": 100 },
    "duration": "PT22M15S"
  },
  "context": {
    "registration": "9f6c3e2a-1b4d-4e8f-a2c7-1234567890ab",
    "contextActivities": {
      "parent": [{ "id": "https://courses.example.edu/biology-101" }],
      "grouping": [{ "id": "https://courses.example.edu/biology-101/cohort-fall-2026" }]
    }
  },
  "timestamp": "2026-05-30T14:32:00Z",
  "version": "2.0.0"
}

xAPI 2.0 / IEEE 9274.1.1-2023 key changes from 1.0.3:

version field MUST appear in statements.
IRI-based identifiers normalized (strict comparison).
Signed statements use JWS over the statement JSON.
Removed the implicit 1.0.x versioning fuzziness.

LRS conformance: any LRS claiming compliance MUST pass the ADL LRS Conformance Test Suite. Self-built LRS is rarely the right answer — use Yet Analytics, Learning Locker (Apereo Foundation), Veracity, or build on a managed service.

Per ferpa-coppa-compliance.md, xAPI statements about minors ARE educational records under FERPA when sent to a school- official LRS, AND parental-consent-controlled under COPPA when sent from a directed-to-children service. The LRS retention policy MUST honor both regimes.

Pattern 5: cmi5 — the SCORM replacement that actually deploys

cmi5 sits on top of xAPI and standardizes the SCORM-shaped contract: a course is a CMI5 package containing cmi5.xml listing AUs (Assignable Units); the LMS launches each AU with ?endpoint=, ?fetch=, ?actor=, ?registration=, ?activityId= query params; the AU then POSTs xAPI statements to the LRS.

Minimum required statements per AU lifecycle:

launched — when the LMS launches the AU.
initialized — when the AU finishes loading.
passed / failed — pass/fail per the AU's mastery score.
completed — when completion criteria are met.
terminated — when the AU is closed by the learner or LMS.

The pass / completed distinction is the killer feature SCORM 1.2 lacked: a learner can complete (finished the content) without passing (didn't reach mastery), and vice versa.

Pattern 6: OneRoster 1.2 — the SIS ↔ LMS contract

OneRoster (REST API binding preferred; CSV fallback for legacy SIS) defines the canonical roster vocabulary:

users — students, teachers, administrators, parents, guardians; with identifier, givenName, familyName, email, username, role, enabledUser, dateLastModified, metadata.
orgs — districts, schools, departments.
classes — class sections (an instance of a course).
courses — the course catalog entry.
enrollments — relationship table linking users to classes with a role (student, teacher).
academicSessions — terms / semesters / grading periods.
demographics — separate endpoint, opt-in, FERPA-sensitive.

REST endpoints: paginated, last-modified-since incremental sync, OAuth 2.0 client credentials.

Per ferpa-coppa-compliance.md Pattern 6, the demographics endpoint is the highest-sensitivity surface: it carries race / ethnicity / disability / English-learner status. Default visibility is OFF; granted only via explicit district consent.

Reuse-first: every roster integration uses the OneRoster spec

a battle-tested library (@studentprivacy/oneroster-client for Node, OneRosterPy for Python) rather than hand-rolled parsing.

Pattern 7: Caliper Analytics — opinionated learning events

Caliper differs from xAPI by being more prescriptive: ~25 predefined Event types (NavigationEvent, AssessmentEvent, GradeEvent, MessageEvent, etc.) with strict schemas. Vendors emit Caliper events via a Sensor SDK to an Event Store.

When to choose Caliper over xAPI:

The data consumers are exclusively LMSs / LMS analytics → Caliper (interop with Canvas Data, Schoology Learning Analytics).
The data consumers include workplace learning, simulations, mobile, VR, informal learning → xAPI (broader vocabulary).
Both → emit Caliper for LMS, xAPI for everything else, bridge at the LRS / Event Store.

Pattern 8: QTI 3.0 — assessment items + tests

QTI 3.0 (2022 release) is the modern format for portable assessment content. An item is XML defining:

<itemBody> — the prompt (rich HTML + MathML + media).
<responseDeclaration> — the expected response type (choice / textEntry / inlineChoice / order / match / hotspot / drawing / mediaInteraction / etc.).
<outcomeDeclaration> — scoring outcomes (SCORE, MAXSCORE).
<responseProcessing> — scoring rules (template or custom XML).

A QTI test is a collection of items + a <testPart> with <assessmentSection>s controlling sequencing, time limits, shuffling, and adaptive routing.

QTI 3.0 mandates WCAG 2.1 AA for items; <accessibility> block declares the item's accommodation needs + variants. Per AccessForAll, learners with declared PNP automatically receive the matching DRD variant.

Pattern 9: Open Badges 3.0 — Verifiable Credentials, not images

Open Badges 2.0 baked a badge into a PNG/SVG via image metadata. Open Badges 3.0 is a W3C Verifiable Credential — JSON-LD with a cryptographic proof. The badge is no longer attached to an image; the image is a presentation artifact.

{
  "@context": [
    "https://www.w3.org/ns/credentials/v2",
    "https://purl.imsglobal.org/spec/ob/v3p0/context-3.0.3.json"
  ],
  "id": "urn:uuid:91537dba-3f97-4e6f-9b8a-1234567890ab",
  "type": ["VerifiableCredential", "OpenBadgeCredential"],
  "issuer": {
    "id": "https://issuer.example.edu",
    "type": ["Profile"],
    "name": "Example University"
  },
  "validFrom": "2026-05-30T00:00:00Z",
  "credentialSubject": {
    "id": "did:example:learner-abc123",
    "type": ["AchievementSubject"],
    "achievement": {
      "id": "https://issuer.example.edu/achievements/data-science-101",
      "type": ["Achievement"],
      "name": "Data Science Fundamentals",
      "criteria": { "narrative": "Completed all modules with score ≥ 80%." },
      "alignments": [{
        "targetFramework": "CASE",
        "targetCode": "DS.1.A.3",
        "targetName": "Statistical reasoning"
      }]
    }
  },
  "proof": {
    "type": "DataIntegrityProof",
    "cryptosuite": "eddsa-rdfc-2022",
    "created": "2026-05-30T00:00:00Z",
    "verificationMethod": "https://issuer.example.edu/keys/2026",
    "proofPurpose": "assertionMethod",
    "proofValue": "..."
  }
}

The cryptographic proof is verifiable WITHOUT calling back to the issuer; the learner owns the credential and can present it to employers, transfer institutions, and credential aggregators (LinkedIn, Credly, EBSI).

Pattern 10: AccessForAll + UDL — accommodations as a first-class capability

UDL (CAST) frames accessibility as three planes:

Multiple means of engagement — recruit + sustain interest; self-regulation; choice; relevance.
Multiple means of representation — perception; language + symbols; comprehension (multiple media; captions; AD; signed video; tactile equivalents).
Multiple means of action + expression — physical action; expression + communication; executive function (multiple response modes; assistive-tech support; scaffolding).

AccessForAll (ISO/IEC 24751) operationalizes UDL by letting learners declare PNP (Personal Needs + Preferences) once; the system delivers DRD (Digitally Resource Description) variants that match. Example PNP entries: display: high-contrast, color: blue-yellow, audio: required-captions, reading-rate: slow, signing: ASL, input: switch-control.

For assessment specifically, ALL accommodations from the IEP / 504 Plan MUST be honored:

Extended time (1.5×, 2×, unlimited)
Read-aloud (screen reader OR human reader)
Scribe (typing assistance)
Calculator (basic / scientific / graphing)
Frequent breaks
Separate setting (proctor-only)
Magnification, color overlay, large print

Hard rules:

Accommodation flags are stored ON the learner's profile, NOT in the assessment URL.
Accommodation determination is the school's responsibility, NOT the platform's — but the platform MUST faithfully apply what the school has set.
An assessment that ignores a declared accommodation is a Section 504 / ADA violation AND an academic-record-integrity violation (the score reflects the disability, not the knowledge).

Pattern 11: Adaptive testing — IRT, not "more wrong → harder"

Adaptive testing means the next item depends on the learner's prior performance. Naïve "tree of branches" implementations are psychometrically meaningless; the validated approach is Item Response Theory (Rasch / 2PL / 3PL):

Each item has parameters: difficulty (b), discrimination (a), guessing (c).
Each learner has a latent ability (θ).
The probability of a correct response is a logistic function of (θ − b).
After each response, the system updates the posterior estimate of θ via MLE / EAP / MAP, then selects the next item to MAXIMIZE information at the current θ (often Fisher Information Maximization).
Termination: when SE(θ) drops below a threshold OR a fixed item count is reached.

CAT (Computerized Adaptive Testing) requires:

A calibrated item bank (typically 5-10× the test length; items calibrated via field-testing with classical statistics
- IRT).
Content balancing (the algorithm prefers high-information items, but the test still needs to cover the blueprint).
Exposure control (Sympson-Hetter; randomesque) so popular items don't appear too often.
Compromise detection (track which items have been seen by many learners in quick succession — possible cheating).

For competency-based microlearning, simpler models (knowledge tracing, BKT, DKT) may be appropriate; for high-stakes assessment, IRT is the floor.

Pattern 12: Proctoring — minimize, never weaponize

Remote proctoring routinely produces these harms:

Bias against non-white learners (face-detection models perform worse on darker skin tones; flagged for "looking away" disproportionately).
Bias against disabled learners (tremor, eye-movement conditions, assistive-tech use flag as "suspicious").
Bias against learners in shared housing (background noise flags; family member entering room flags).
Privacy harms (recording a learner's bedroom; demanding 360° room scan; logging keystrokes including passwords typed outside the exam window).
Data-protection harms (biometric scans retained indefinitely; outside the FERPA / GDPR / BIPA / Illinois Biometric Information Privacy Act lawful basis).

Hard rules:

Default to NO proctoring. Pivot to project-based assessment, open-book exams, oral defense, formative assessment, where feasible.
When proctoring IS required, use the LEAST invasive method that achieves the validity goal:
- Human proctor (live or async review) ≻ AI flags + human review ≻ AI auto-flag with no human review (NEVER use this).
Bias-audit every proctoring vendor: demand published TPR/FPR by skin tone, gender, age, disability status; reject vendors who can't produce one.
Biometric retention: 30 days max post-exam; explicit consent; separate from FERPA records (BIPA requires written informed consent in Illinois).
Accommodations apply: extended time, breaks, separate setting ALL pass through to the proctor session.
Learner has the right to challenge ANY proctoring flag before academic consequence.

Pattern 13: Learning analytics — ethics + transparency

Learning analytics (LA) dashboards predicting "at risk" learners are deployed widely. Hard rules:

Transparency: the learner (and parent, for minors) MUST see what data is collected, what models are trained, what predictions are made about them, and how to challenge.
Purpose limitation: data collected for engagement- monitoring CANNOT be used for admissions, scholarship, or disciplinary decisions without separate lawful basis.
Bias audit: at-risk models trained on historic outcomes bake historic discrimination into the predictions. Audit TPR/FPR by protected class quarterly.
Action discipline: a prediction of "at risk" must lead to SUPPORT (advisor outreach, tutoring offer), never to surveillance escalation or pre-emptive academic penalty.
GDPR Article 22 + automated-decision-making: a prediction with "legal or similarly significant effect" requires human review + the right to obtain human intervention + the right to contest.

Anti-Patterns

Anti-Pattern 1: Shipping LTI 1.1 in 2026

The OAuth 1.0 HMAC-SHA1 model is end-of-life; tools that require it will fail platform certification + lose customers as LMSs deprecate 1.1 support (Canvas + Brightspace already discourage; Moodle 4.x prefers 1.3+; Blackboard Ultra is 1.3-only). New tools MUST be 1.3+ from day one.

Anti-Pattern 2: SCORM scoring drift across LMSs

The same SCORM 1.2 package scoring differently in different LMSs is the most common bug. Causes: cmi.core.lesson_status ambiguity between "completed" and "passed"; LMSs treating SCORM_DOC's "completed" as "100/100"; content using cmi.score.raw without min/max. Fix: always set min/max; emit both completion + success status in 2004; test against the Rustici Cloud SCORM Engine for canonical behavior.

Anti-Pattern 3: xAPI statements pointing at user-typed

identifiers actor.mbox = mailto:user@example.com works but leaks PII across the wire. Use actor.account with a stable opaque identifier instead; reserve mbox/mbox_sha1sum for single-LMS deployments where the LRS is inside the FERPA boundary.

Anti-Pattern 4: Roster sync via direct DB access

SIS vendors sometimes offer "direct DB credentials" as a faster path than OneRoster. NEVER take this path: it produces zero schema stability, zero authorization controls, zero audit trail, and the contract becomes "we sync everything in the SIS including fields no one consented to share." Use OneRoster 1.2 REST OR Clever/ClassLink-mediated provisioning.

Anti-Pattern 5: AI tutor / AI grader without dataset

provenance Building a tutor or grader on a foundation model with no clarity on what training data the model saw fails on two fronts: (a) FERPA-protected data may have been in the training set (the model can leak educational records); (b) the model may have memorized copyrighted textbook content and reproduce it. Use only models with documented training-data provenance; prefer RAG over fine-tuning when grounding is essential; never let the AI grader's score be the SOLE basis of a grade — always require human review per ai-ethics veto.

Anti-Pattern 6: Generic "early warning" dashboard with no

intervention pathway Dashboards that flag learners as "at risk" without naming WHO intervenes, WHAT the intervention is, and HOW the learner can challenge the label produce surveillance theater + harm. Every EWS implementation MUST have a closed intervention loop + audit log of actions taken per flag.

Anti-Pattern 7: Adaptive assessment with no validity evidence

"Our algorithm picks easier questions when you get one wrong" is not adaptive assessment; it's a guessing-friendly UX. Validated CAT requires a calibrated item bank, IRT modeling, documented validity + reliability evidence. Without those, the "adaptive score" has no psychometric meaning, MUST NOT contribute to academic records, and MUST be labeled "for practice only."

Anti-Pattern 8: Proctoring without accommodation support

Proctoring vendors that don't expose accommodation flags (extended time, frequent breaks, separate setting, scribe, read-aloud) at the session level are not deployable in American K-12 or higher-ed. The IEP / 504 / OCR audit trail WILL find them.

Anti-Pattern 9: Caliper + xAPI emitted from the same handler

without bridge Running two parallel event pipelines, one for each specification, produces drift: event counts differ between LRS + Event Store, dashboards disagree, and reconciliation is impossible. Emit canonically into ONE store + bridge to the other format at read time, OR emit both from the same source-of-truth event with a contract test that ensures field-level parity.

Anti-Pattern 10: Treating MOOCs as exempt from accessibility

"It's a free course on the open web, so accessibility is nice-to-have." DOJ Title II Final Rule (April 2024) explicitly applies to public-college MOOCs; ADA Title III applies to private platforms offering courses to the US public; EAA applies in Europe. The accessibility floor is not optional and extends to AT-rendered video transcripts, captions, audio description, MathML for STEM content, and PDF tagging.

Anti-Pattern 11: Hardcoded grade scales + locale assumptions

"A is 90+" works in the US, not in Germany (1.0-5.0 inverted), not in the UK (degree classifications), not in IB (1-7), not in China (100-point with passing thresholds varying by school type). Grade representation MUST be a typed GradingSchemeRef with locale + institution context, NEVER a freeform string assumed to be "A-F."

Anti-Pattern 12: COPPA non-compliance via "school as agent"

without DPA A vendor relying on the school-as-agent VPC exception (per ferpa-coppa-compliance Pattern 4) WITHOUT a signed DPA that documents the school's role + the vendor's data-handling + ad-free + retention obligations is not actually using the exception — the vendor is still on the hook for COPPA VPC. This is the dominant K-12 audit finding in 2024-2025 FTC investigations.

Verification Checklist

When this skill activates, verify:

Cross-References

~/.claude/skills/ferpa-coppa-compliance/SKILL.md — student- privacy regulation; always paired with this skill for K-12
~/.claude/skills/wcag-accessibility/SKILL.md — WCAG 2.2 AA
- AAA implementation
~/.claude/skills/gdpr-ccpa-compliance/SKILL.md — when EU / CA minors are learners
~/.claude/skills/interaction-design/SKILL.md — learning experience design discipline
~/.claude/skills/ux-research/SKILL.md — formative + summative user research with learners (with IRB / parental consent)
~/.claude/skills/rag-design/SKILL.md — when AI tutors ground answers in course materials
~/.claude/skills/prompt-engineering/SKILL.md — AI-tutor system prompts
~/.claude/skills/ml-model-selection/SKILL.md + mlops-patterns — adaptive engines + at-risk models
~/.claude/skills/security-review/SKILL.md — LTI launch signature verification, JWT validation, OAuth client-credential rotation
~/.claude/skills/observability-patterns/SKILL.md — Caliper
- xAPI event pipelines
~/.claude/skills/api-design/SKILL.md — designing additional REST endpoints alongside standards-mandated ones
~/.claude/skills/research-methods/SKILL.md — validity + reliability evidence for assessment + adaptive engines
~/.claude/rules-library/common/a11y.md — global WCAG floor rule
~/.claude/rules-library/common/security.md — global security umbrella (LTI JWT, OAuth, JWKS rotation)
~/.claude/rules-library/common/audit-logging.md — access to educational records is audit-logged
~/.claude/rules-library/common/data-retention.md — minor-data retention bounded; biometric retention 30 days max
~/.claude/agents/accessibility-reviewer.md — pairs on UI / assessment a11y review
~/.claude/agents/ai-ethics-reviewer.md — AI tutor + grader
- EWS bias review

Why This Skill Exists

EdTech is the second-most-regulated software domain after healthcare, with the additional twist that the user base is disproportionately minors with reduced legal capacity to consent or contest. Recent enforcement + litigation patterns that this skill aims to prevent:

2024 DOJ Title II Final Rule — class-action against community colleges and state universities for WCAG-non- conformant LMSs, captioning failures, and assessment platforms locking out blind students.
Edmodo FTC settlement ($6M, May 2023) — COPPA + targeted advertising on student data; the kind of pattern that recurs across edtech if FERPA + COPPA aren't engineered defensively.
CDE / state-AG actions on proctoring vendors (2020-2024) — bias against students of color, students with disabilities, and students in shared housing leading to refunds + injunctions.
HEOA + State Authorization Reciprocity Agreement (SARA) audits — institutional liability when third-party course materials don't meet accessibility + verification of student identity standards.
District-AG investigations of "free-for-schools" tools — vendors that monetized "free" K-12 use via behavioral advertising or data resale faced state AG + FTC + class- action exposure 2023-2025.
Open Badges 2.0 deprecation pressure — credentials baked into images that can be tampered with, copied, or lose the linked-data context; W3C VC Open Badges 3.0 became the only credentialing pattern enterprise + government accepts by 2025+.
EU AI Act high-risk classification (effective Aug 2026) — education + assessment systems classified as "high risk"; mandates conformity assessment + risk management + bias testing + human oversight + transparency for AI graders, proctoring, and admissions tools.

The cost of getting this right: choosing standards-conformant libraries + paying for accessibility audits + investing in IRT modeling + signing DPAs + auditing AI components quarterly. The cost of getting it wrong: losing district contracts, class-action litigation, OCR investigations, DOJ consent decrees, and — most importantly — locking learners out of their own education for reasons unrelated to their knowledge or capability.

Not legal advice. Not pedagogical advice for any specific learner. This skill provides engineering patterns; the validity of any assessment, the appropriateness of any accommodation, and the legal sufficiency of any consent framework requires institution counsel, learning-science review, and (for assessment) psychometric expertise.

Learning hooks

Per ~/.claude/rules/common/continuous-learning-mandate.md:

Signals to watch:

LTI 1.1 / 1.2 integration shipped in new code (rule 1 violation — LTI 1.3+ floor)
LTI launch handler missing nonce one-time-use store (replay attack vector)
LTI launch handler skipping JWKS verification + caching (key rotation + signature validation gap)
AGS / NRPS access token minted per request rather than cached (rate-limit + cost issue)
SCORM package missing min/max on cmi.score.raw (scoring drift across LMSs — Anti-Pattern 2)
SCORM 2004 package only setting lesson_status (missing completion/success split)
xAPI statement using raw email in actor.mbox (Anti-Pattern 3 — PII over wire)
xAPI 2.0 statement missing version field (spec violation)
OneRoster sync via direct DB credentials (Anti-Pattern 4)
Roster demographics fetched without district consent (FERPA + COPPA violation)
QTI items shipped without <accessibility> block (a11y default-off)
Adaptive assessment without IRT validity evidence shipped (Anti-Pattern 7)
Proctoring vendor selected without published bias audit (Pattern 12 weakening)
Accommodation flag (IEP / 504) bypassed in proctor session (Section 504 violation)
AI tutor / grader shipped without documented model provenance (Anti-Pattern 5)
AI grader producing record-affecting outputs without human review gate (ai-ethics veto)
Early-warning dashboard without closed intervention loop (Anti-Pattern 6)
Open Badges 2.0 (baked-image) emitted for new credentials (Pattern 9 weakening)
MOOC shipped without WCAG 2.2 AA (Anti-Pattern 10 — DOJ Title II / EAA)
Hardcoded "A is 90+" grade scale (Anti-Pattern 11 — locale assumption)
School-as-agent VPC claimed without DPA in place (Anti-Pattern 12 — FTC 2024-2025 finding pattern)
Caliper + xAPI emitted in parallel without contract test (Anti-Pattern 9 — drift)

Refinement candidates:

New row in Standards Cited when 1EdTech publishes a new spec major version (LTI 2.0, QTI 4.0, OneRoster 1.3, Caliper 2.0)
Tightening of accommodation-passthrough verification when OCR / DOJ enforcement adds a specific failure pattern
New row in Anti-Patterns when an FTC / OCR / DOJ enforcement action names a new edtech recurrence
New cross-reference when EU AI Act conformity-assessment patterns become public (Aug 2026+)
New pattern entry when a credentialing standard (CLR, EDC, EBSI digital wallet) ships a production binding
Tightening of proctoring rules when bias-audit methodologies improve (NIST FRVT-style protocols for ed proctor)
New row when state student-privacy laws expand (TX SB-820 II, IL SOPPA amendments, more state-MTLs-equivalent edtech laws)

edtech-patterns

About