Evidence reviews and scorecards · Complete web edition

Phase Six Verification Investigation Design

This is the complete web edition of the canonical project document. Its narrative, recommendations, evidence tables and source links are presented below without abridgment.

2,412words

14tables

50text blocks

Download the Word document

Project, Research, Analysis, Editorial by Patrick Russell; assisted by ChatGPT (OpenAI).

Published as a web edition September 13, 2026

California Education and AI Verification Investigation Design

Phase Six From public scorecard to testable evidence

Responsible person	____________________________________________
Organization or role	____________________________________________
Review date	____________________________________________

Decision

The California comparison is ready for verification, but the institutions should not be tested as if they were the same kind of school. The investigation should preserve two tracks. The K through 12 track should examine development, parent authority, age appropriate AI use, relationships, and preparation for independent study. The college and university track should examine independent scholarship, disciplinary judgment, course design, faculty governance, and degree level outcomes. Both tracks should measure what students can do without AI after assistance has ended.

Mother of Divine Grace School belongs only in the K through 12 track. It is a private Catholic distance learning and home education program, not a university. Thomas Aquinas College remains a separate four year college level comparison.

Purpose and boundaries

Verify whether published educational principles appear in representative programs, courses, assignments, assessments, and student support practices.
Distinguish AI access and adoption from durable learning, independent judgment, attention, relationships, mental fitness, and wellness.
Treat COVID closure exposure as one historical factor whose effects may differ by prior condition, family capacity, disability, friendships, and instructional support.
Use public evidence first. No information request packet is part of this phase. Interviews, restricted records, or student level research would require separate authorization, consent, privacy review, and institutional approval.

Questions the investigation should answer

Question area	Testable question
Formation	Do students repeatedly read, remember, reason, calculate, write, discuss, revise, and defend work in ways that make their own thinking visible?
AI purpose	Is each AI use tied to a defined learning problem, or is access being treated as the objective?
Independent transfer	After assistance ends, can students perform a comparable task without AI and explain their reasoning?
Human relationships	Does the design preserve contact with parents, teachers, peers, tutors, advisers, counselors, and community?
Wellness	How do attention, self regulation, belonging, distress, sleep, family burden, and help seeking change?
Governance	Who approves tools and rules, how are parents or students informed, and how are errors or misconduct findings reviewed?
Historical change	Which differences predate COVID, which track closure exposure, and which emerge during the later AI period?

Institution sample

All ten institutions should remain in the documentary review. A smaller matched sample should be used only when direct measurement becomes possible. The sample should represent each institutional form rather than selecting only the schools with the most public AI material.

Institution	Type	Reason for inclusion	Minimum review unit
UC Berkeley	Public university	Large institution with mature AI infrastructure	General education and two contrasting disciplinary courses
UCLA	Public university	Six public teaching and learning AI principles	Faculty fellow or workshop course and a comparison course
Stanford University	Private research university	Human centered AI literacy and broad implementation	AI integrated course and protected human performance
USC	Private research university	University wide ChatGPT Edu access	Courses with different instructor permissions
Caltech	Private research university	Strong verification and actual ability concerns	Writing intensive and technical course
Loyola Marymount University	Private Catholic university	Whole person mission with developing AI practice	Core curriculum and professional program
Thomas Aquinas College	Private Catholic college	Integrated Great Books formation model	Seminar laboratory tutorial composition and thesis evidence
Mother of Divine Grace School	Private Catholic K through 12	Parent directed developmental formation	Elementary middle and high school sequences
Harvard Westlake School	Private grades 7 through 12	Independent school with visible AI coursework	Middle and upper school assignments
High Tech High	Public charter K through 12	Project based work and public exhibitions	Elementary middle and high school projects

Two track comparison

Dimension	K through 12 track	College and university track
Educational purpose	Development toward attention memory language reasoning character and later independence	Independent scholarship disciplinary judgment professional competence and civic responsibility
Authority	Parents guardians teachers school leaders and governing boards	Students faculty departments academic senates integrity offices and administration
AI permission	Age specific notice consent supervision approved tools and a non AI alternative	Course specific rules disclosure data controls authorship standards and appeal
Verification	Oral narration recitation handwriting calculation observation projects and age appropriate delayed checks	Examinations oral defense source analysis original writing calculation laboratory performance and delayed transfer
Relationships	Parent child teacher peer counselor and community contact	Faculty student peer adviser residential professional and community contact
Outcome horizon	Development within grade bands and preparation for the next stage	Course mastery retention transfer degree progress and post completion capability

Matched comparison design

The strongest feasible design should compare similar students or courses under clearly different AI conditions. Random assignment may be appropriate in some low risk instructional pilots. When it is not feasible, matched comparison groups and repeated pre and post measures should be used. A simple before and after comparison without a credible comparison group should be treated as descriptive evidence.

Condition	Definition	Reason
Protected human work	No generative AI during the central mental act	Establish unaided baseline and protect foundational practice
Bounded student AI	Hints questions feedback or source navigation after a first attempt	Test assistance that preserves effort and explanation
Open student AI	Broad chatbot or generation access with disclosure	Measure substitution dependence and authorship risk
Teacher facing AI	Planning diagnostics examples or feedback suggestions reviewed by the educator	Test efficiency and instructional quality without delegating judgment

Evidence that should be collected

Evidence layer	Examples	What it establishes
Public institutional record	Mission curriculum handbooks AI rules privacy terms board records teaching guidance archived webpages and public reports	Authority scope implementation date and change history
Course or grade record	Syllabi lesson sequences assignments rubrics tool permissions and assessment design	What students are expected to do and how performance is verified
Visible student performance	Public exhibitions portfolios theses demonstrations presentations and deidentified work when lawfully available	Quality process revision explanation and independent contribution
Outcome record	Unaided assessments delayed recall transfer attendance persistence and completion	Learning that remains after support ends
Human condition record	Attention belonging distress sleep friendships help seeking family burden and teacher or parent contact	Mental fitness and wellness without reducing them to one score
AI system record	Tool model purpose prompt timing frequency output retention data location and human review	Actual exposure rather than the label AI assisted

Measures of a well formed mind

The investigation should report separate measures. A composite score may summarize patterns, but it should never hide a decline in memory, authorship, relationships, or wellness behind gains in speed or completed output.

Domain	Observable capacity	Minimum method
Attention and self regulation	Sustained task time distraction recovery first attempt completion and appropriate help seeking	Observation task logs and validated age appropriate scales
Memory	Immediate recall delayed recall and retrieval without AI	Parallel quizzes oral recall or narration
Reading	Comprehension inference source comparison and accurate quotation	New passage assessment and oral explanation
Reasoning and judgment	Claim evidence counterargument uncertainty and decision quality	Written analysis oral defense and unfamiliar problem
Writing and authorship	Original composition revision ownership citation and ability to explain choices	Draft history oral defense and unaided writing
Mathematical and scientific thought	Calculation proof modeling laboratory reasoning and error diagnosis	Worked problems demonstration and practical task
Relationships and belonging	Parent teacher peer contact trust help seeking collaboration and isolation	Surveys interviews observation and contact records
Mental fitness and wellness	Distress sleep motivation agency emotional regulation family strain and functioning	Validated measures reported separately with support pathways
Independent transfer	Comparable performance on a new task after AI support is removed	Delayed no AI assessment with the same learning objective

AI use record

Every observed AI use should be recorded with enough detail to reproduce the educational condition. Brand name alone is insufficient because models, settings, system instructions, data terms, and classroom rules change.

Field	Required record
Purpose	Tutor feedback research planning generation assessment companion or administration
Timing	Before first effort during practice after a draft during assessment or after assessment
Control	Student initiated teacher assigned teacher facing or automated
Guardrails	Hints only source grounding refusal of assessed answers disclosure age limit and human review
Data	Inputs personal information retention training use hosting jurisdiction and vendor access
Verification	Oral explanation source check draft history parallel unaided task and delayed transfer

COVID wellness and historical comparison

COVID should be examined as a change in learning conditions rather than a universal diagnosis. School closure, emergency remote instruction, loss of services, reduced friendship contact, family stress, and unequal adult supervision may have had different effects on students with different starting conditions. The design should test whether students who already faced dysregulation, weak family support, few friendships, disability related barriers, or mental health difficulties were less able to maintain prior learning and wellness than students with stable routines and strong support.

Period	Minimum variables	Interpretive use
Before 2020	Achievement attendance attention wellness disability support friendships and family conditions	Starting condition
Closure period	Closure duration actual instructional contact service continuity technology access adult supervision and household stress	Exposure and protection
Recovery period	Return timing attendance learning support social reconnection symptoms and family burden	Persistence or recovery
AI period	Tool access timing purpose safeguards dependence teacher practice and independent performance	Later educational condition

Institution specific verification priorities

Institution	Priority
UC Berkeley	Trace central guidance into representative courses and compare assisted output with later unaided mastery
UCLA	Test whether the six teaching and learning principles appear in assignments assessments and student outcomes
Stanford University	Measure whether human centered AI literacy predicts verification metacognition well being and transfer
USC	Compare courses with different ChatGPT Edu rules while controlling for discipline instructor and student preparation
Caltech	Test whether oral in person process based and delayed assessment practices preserve actual ability
Loyola Marymount University	Determine whether the whole person mission changes AI permission mentoring assignment design and measured wellness
Thomas Aquinas College	Document protected human acts and test whether any permitted AI use strengthens rather than substitutes for them
Mother of Divine Grace School	Examine elementary middle and high school sequences separately including parent oversight social support and age specific AI boundaries
Harvard Westlake School	Verify schoolwide AI expectations and compare middle and upper school authorship research and oral explanation
High Tech High	Verify individual mastery within collaborative projects and public exhibitions including any AI assisted work

Analysis and evidence rules

Rule	Application
Separate adoption from benefit	Tool access frequency satisfaction and completed work do not establish learning
Preserve the baseline	Record preexisting achievement attention disability family and wellness conditions before attributing change
Use comparable tasks	Assisted and unaided conditions should test the same learning objective at similar difficulty
Measure after assistance	Immediate output should be followed by delayed recall transfer or oral defense
Report variation	Publish results by age grade course discipline prior preparation support condition and AI use pattern
Avoid a single wellness score	Report attention distress belonging sleep agency relationships and family burden separately
Distinguish absence of evidence	A missing public record is an unresolved gap rather than proof that a practice does not exist
State causal limits	Describe association unless assignment comparison quality and attrition support a stronger conclusion

Minimum evidence ratings

Rating	Evidence level	Minimum meaning
3	Strong comparative evidence	Credible comparison or randomized design low attrition comparable measures and delayed unaided outcomes
2	Repeated quantitative evidence	Multiple cohorts or courses with pre and post measurement but important selection or implementation limits
1	Descriptive evidence	Public example survey observation case report or before and after result without a credible comparison
0	Not established	No relevant public evidence located or evidence cannot answer the question

Privacy consent and research review

Public documentary analysis can proceed without collecting private student records. Any later study involving identifiable education records, surveys, interviews, classroom observation, or experimental assignment should be reviewed by the responsible institution before data collection. The review should determine which privacy, consent, parental notice, child protection, research ethics, and data security requirements apply.

Use the least identifiable data that can answer the question and define deletion dates before collection.
For minors, separate parent or guardian permission from the child or adolescent assent process when required.
Do not place sensitive wellness disclosures in general AI systems or ordinary classroom prompts.
Do not use AI detection output as the sole evidence of misconduct. Preserve notice, corroboration, human review, and appeal.
Publish aggregate findings only when small groups and unusual combinations cannot reveal individuals.

Work sequence

Step	Work	Minimum result
1	Freeze the record	Archive the cited public policies curricula tools terms and dates for all ten institutions
2	Select review units	Choose representative grades courses assignments and AI conditions within each institutional type
3	Code implementation	Record purpose timing authority guardrails data exposure assessment and human relationships
4	Establish baseline	Document prior learning support wellness COVID exposure and relevant student or course differences
5	Measure performance	Use immediate and delayed unaided tasks oral explanation and visible process evidence
6	Analyze differences	Compare conditions report variation test sensitivity and state causal limits
7	Update the scorecard	Revise formation AI readiness and outcome evidence scores with traceable evidence
8	Publish findings	Report methods results limitations responsible person corrective actions and next review date

Evidence record

Field	Entry
Institution and educational level	____________________________________________________________
Program course grade or assignment	____________________________________________________________
AI condition and tool version	____________________________________________________________
Protected human capacity	____________________________________________________________
Comparison and baseline	____________________________________________________________
Immediate and delayed measures	____________________________________________________________
COVID family friendship and wellness factors	____________________________________________________________
Finding evidence rating and limitation	____________________________________________________________
Responsible person and next review date	____________________________________________________________

Expected outputs

An archived public evidence record for all ten institutions with dates and source links.
Separate K through 12 and college or university findings with no cross level ranking of students.
Course and grade case studies that distinguish protected human work bounded AI open AI and teacher facing AI.
A COVID and wellness analysis that tests prior vulnerability and support rather than assuming one uniform effect.
An updated scorecard supported by direct outcome evidence and a clear list of unresolved questions.

Public standards and sources

These sources establish the minimum research evidence privacy and human oversight standards for the verification phase. Institution specific sources remain listed in the Phase Five scorecard.