Inspect the evidence.
Know the limits.

LEARNALITY - RESEARCH AUDIT AND DEVELOPMENT DECISIONS
October 14, 2026 • Version research-audit-1
Scope: first US grades 9-12 school pilot; proposed future Pre-K-12 pathway.
Status: targeted literature and source-code audit. No Learnality student validation study completed.

THE MAIN FINDING
There is a defensible project here: collect student voice, inspect defined learning tasks, preserve required curriculum and access arrangements, trial teaching supports, measure subsequent learning, and revise with educator review. Published papers do not establish that a short annual test can identify every student's ideal mixture or perfectly adapt any curriculum. Continuous percentages, appealing roles, AI and immutable records do not supply that missing evidence.
A useful adaptive system would estimate what helps this student with this objective in this context, with uncertainty. It must remain willing to conclude that a broadly useful teaching practice helps many students and that a personality-specific rule adds no benefit.

WHAT THIS REVIEW DID
Three coordinated audits examined measurement/development, instruction/personalization, and engagement/presentation. Sources were sought through primary journal publications, author-hosted papers and official assessment/development guidance. Search themes included learning-style matching and newer opposing evidence, retrieval/spacing, self-regulation, adaptive instruction/AI, profile validity, repeated testing, developmental assessment, gamification, decorative distraction, and perceived versus demonstrated learning. Exact sources, populations, design, access status and limitations are recorded in the annexes.
Search date: October 14, 2026. Foundational and newer publications were included; this is not a claim that every paper published through that date was captured. No registered protocol, complete database export, duplicate independent screening or PRISMA flow was performed. Abstract-only or inaccessible documents are flagged, not presented as fully appraised evidence. Vendor reports are identified. No invented effect sizes, school samples, national norms or power calculations are supplied.

READ THE SOURCE AUDITS
measurement-research-audit.txt - score meaning, self-report bias, individual profiles, retakes, annual comparisons, fairness/access and young-child assessment.
instruction-research-audit.txt - learning-style matching including newer contrary evidence, effective study/instruction, adaptive systems and AI limitations.
engagement-research-audit.txt - gamification, perceived learning, rewards, distractions and personality-test inspiration.
These are complementary audit notes, not three independent systematic reviews. Repetition of a citation is not additional evidence. The same underlying studies across reviews must not be counted repeatedly.

CLAIM-TO-EVIDENCE AND IMPLEMENTATION REGISTER

1. “These percentages objectively describe how much of each Learnality you are.”
Decision: not supported for the current instrument. Two ordinal self-report items per proposed role provide maximum-endorsement indices. They are not fractions of personality, success probabilities, mastery percentages, norm percentiles or calibrated latent-trait estimates. Separately displayed dimensions need not sum to 100; that mathematical choice does not establish eight independent psychological factors. Current original item wording/factor structure and score precision remain unvalidated. See measurement audit M1-M2 and its local source audit.
Action: keep student voice separate from scored artifacts; show exact item coverage and context. Test whether the eight-role structure survives expert review and empirical analysis; revise or discard it if it does not.

2. “Matching curriculum to a learning personality reliably improves learning.”
Decision: not established for Learnality. Historical critiques and later direct tests challenge simple preference matching. A 2024 synthesis found small average benefits but inconsistent interaction evidence and methodological concerns. It does not validate Learnality. Primary source: https://doi.org/10.3389/fpsyg.2024.1428732 . The instruction audit records its detailed limitations.
Action: compare profile-guided teaching with good existing practice and broadly useful supports. Distinguish an average intervention benefit from a genuine, reproducible difference in who benefits and a useful routing policy. Use prospective held-out outcomes; do not pick a method because its explanation sounds like the student's label.

3. “Independent tasks identify learning ability rather than prior knowledge or access barriers.”
Decision: current tasks provide bounded observations. A rule, table, plan or explanation can depend on reading, arithmetic, familiarity, language, access, opportunity and help. The new algebra bank samples one objective only; common templates do not establish alternate-form equivalence. More questions alone do not establish broader validity.
Action: define claims/evidence/tasks and rival explanations before expansion. Use prerequisite checks, cognitive interviews, multiple content families, rubric exemplars and actual artifacts. Genuine collaboration requires genuine interaction; a simulated peer does not establish collaborative skill. Validate item/occasion/rater sampling for the intended decisions.

4. “Retakes and annual changes measure growth.”
Decision: immutable histories are implemented; calibrated growth is not. Higher scores can reflect exposure, different difficulty, help or changed self-report reference standards. Preference changes and objective learning changes are distinct. See measurement audit M2, M5 and M1.
Action: separate reflection revisits from new task runs; preserve all versions and exposure. Establish appropriate linking/comparability and response-shift interpretation before growth percentages. Retakes propose a plan revision; they must not erase mastery, change grades, cancel already issued work or silently alter lawful supports.

5. “Students should receive different tests that fit their style.”
Decision: practice can vary; outcome comparability cannot be assumed. Assessment access may vary when the target construct is preserved. If writing itself is the objective, replacing writing with drawing changes the claim. A shared criterion string is not proof of equal difficulty.
Action: preserve target coverage, cognitive demand, scoring criteria and allowed accommodations; create expert-reviewed equivalent prompts, common anchors where appropriate, independently reviewed open artifacts and actual delayed/transfer outcomes. Measure rater agreement and investigate subgroup functioning. No easy-test routing from a low reflection score.

6. “The student never needs a teacher.”
Decision: intuitive self-administration is a usability goal, not a result already demonstrated. Independent learning often needs subject-specific instruction, scaffolded planning and feedback. Access support and classroom professional judgment remain necessary. No paper establishes that one unsupervised interface suits every age.
Action: observe independent use in the actual high-school pilot and record requested UI help separately from conceptual help. Keep available adult/access support. Pre-K requires its own short, authentic adult-supported observation/play design; see NAEYC guidance in measurement audit M6. Do not resize a high-school questionnaire and claim a cross-age scale.

7. “Animations keep disengaged learners learning.”
Decision: an engagement hypothesis, not an established Learnality effect. Gamification evidence is mixed by outcome/design; attractive irrelevant content can distract. Perceived fluency and actual learning can diverge. See engagement audit E1-E5.
Action: optional, non-score-dependent companion choices, progress feedback at boundaries and reduced motion. Compare simplified/richer designs with unchanged assessment content; measure disengagement, accessibility, fatigue, persistence and retained learning separately. Do not claim measured dopamine or infer learning from time-on-site.

8. “16Personalities provides the scientific model.”
Decision: inspiration for a welcoming experience only. A vendor's reliability report concerns its own instrument and use. Its coefficients, types, items and norms do not validate Learnality or instructional matching. See engagement audit E6.
Action: original visual identities and student-friendly strategy vocabulary; no inherited psychometric authority. Describe the roles as approaches to develop. Avoid fixed labels that limit a subject or imply diagnosis.

9. “AI can adapt any curriculum perfectly.”
Decision: not established. AI-generated explanations, questions, scoring and pacing require their own evaluations. Existing tutoring/adaptive intervention studies have specific populations, tasks, safeguards, comparators and outcome conditions; transfer to this product is not automatic. See instruction audit.
Action: begin with a bounded reviewed unit and verified source/objective lineage. Check factual accuracy, omitted objectives, construct fidelity, answer keys, access, resource/time constraints and human review. Preserve a blocked/infeasible outcome. No identifiable student data may be sent to an external model without approved arrangements. No AI personality score is created from chat fluency or engagement.

THE TEST THAT SHOULD BE DEVELOPED
The following is our design inference and proposed research sequence, not an externally validated recipe or fixed sample-size rule.
A. Student voice: short contextual reflection, goals and optional access/opportunity context. Never grade self-description. Investigate wording, desirable answers, comprehension and response mapping.
B. Starting knowledge: curriculum prerequisites and familiarity. Report this separately from learning after instruction.
C. Learning episodes: reviewed instruction, varied tasks and clearly logged help. Preserve first attempts and supported revisions as one evidence sequence, not multiple independent successes.
D. Reasoning artifacts: hypotheses, models, source judgments, explanations, controlled comparisons, plans, actual collaboration and applicability/non-applicability of principles. Every claim has an observable criterion and an explicit unobserved state.
E. Common outcomes: independent mastery, later retention and unfamiliar transfer under comparable demand/access. Exact intervals, earlier quizzes, restudy, absence and outside help are recorded. A 7 to 10 day return is the current timing convention, not a universally optimal delay established by research.
F. Support responsiveness: repeated prospective comparisons using fresh appropriate content and controlled/counterbalanced assignment where feasible. Distinguish prerequisites, content, time, teacher and order effects from support effects. Learning is not reversible; an A/B “washout” cannot erase what was learned.
G. A changing evidence profile: dated observations, coverage, disagreement and uncertainty. A self-report change is valid student feedback but not sufficient evidence to replace the curriculum automatically.
The annual sitting starts the year's evidence collection. It cannot be the only observation used for precise ongoing adaptation. More data are not always better: purposeful sampling, fatigue/access review and minimization matter.

VALIDATION PROGRAM REQUIRED BEFORE STRONG CLAIMS
Stage 0 - Intended-use specification. Freeze permitted/prohibited claims and consequences. Measurement, curriculum, accessibility and developmental experts review the construct blueprint and wording. Roles are provisional; review must be able to change them.
Stage 1 - Response processes and administration. Cognitive interviews and authentic usability across grades 9-12, prior attainment, language, access/device contexts and engagement patterns, including non-completers. Revise misleading prompts/interfaces before scoring scale development.
Stage 2 - Task and scoring development. Independently review keys; train raters; double-score planned artifacts; estimate relevant task/occasion/rater variation, precision, missingness and dependence. Examine scale structure and fairness for intended individual uses. Select study sizes and reporting thresholds from decision requirements and planned analyses with a qualified specialist; no universal “three opportunities” rule supplies validity.
Stage 3 - Recommendation experiment. Predeclare common primary outcomes, meaningful benefit/error criteria, analysis, attrition handling and stopping/revision rules. Compare strong current practice, broadly supported instruction and profile-guided plans. Analyze heterogeneity with adequate design; hold out students/sites/time appropriately. Report mixed/null results and burdens, not just completion or satisfaction.
Stage 4 - Longitudinal comparability. Investigate alternate forms, changing prerequisites/constructs, exposure, score drift, subgroup functioning and missing longitudinal records. Establish exactly what can be compared within/between grades before annual-growth or norm claims.
Stage 5 - Independent review and bounded release. Publish a technical manual, source ledger, validation findings, populations, versions, uncertainty and prohibited decisions. Monitor drift and revalidate consequential item/model/interface changes. Extend younger ages only through separate development and linking evidence.
School authorization, consent/governance, privacy/security and research review remain separate prerequisites. Papers and software checks do not replace them. No real school records, students or research recruitment were used in this audit.

CHANGES THIS AUDIT REQUIRES NOW
• Update the science area with this audit and the newer, qualified matching evidence.
• Withdraw obsolete arbitrary role-highlighting thresholds from the active research rationale; preserve them only as clearly retired historical proposals.
• Continue to label every current percentage as self-report endorsement and every task/plan as bounded/provisional.
• Stop treating new modules or passing source tests as progress in educational validation. Source behavior, usability, psychometric evidence and instructional benefit are separate statuses.
• Prioritize the technical blueprint/content review and a preregistered pilot protocol before expanding stronger claims or multiplying questionnaires.
• Keep assignment checks and histories, but regard them as safeguards for prospective evaluation, not evidence that the policy works.

PUBLIC CLAIM PERMITTED TODAY
“Learnality is a research-informed project for student reflection, defined learning challenges and educator-reviewed support planning. Current percentages summarize self-reported strategy endorsement. Its original assessment and personalization policy have not demonstrated validated individual matching or calibrated annual growth.”

WHAT REMAINS UNRESOLVED
A literature audit can identify the evidence and design requirements. Only Learnality's own appropriately authorized studies can establish its score precision, fairness, usability, alternate-form equivalence, individual recommendation benefit, annual comparability or performance in younger children. A claim that the product is fully validated, universally accurate or perfectly curated remains unjustified.

LEARNALITY - MEASUREMENT RESEARCH AUDIT
Reviewed October 14, 2026 • Private development research
Scope: score meaning, self-report, retakes, annual growth, fairness/access, developmental extension.

Finding
The present app can collect student voice and task-specific evidence. It cannot yet identify an objectively established eight-part learning personality, estimate the best teaching mixture for each student, or report calibrated annual learning-ability growth. Replacing categories with continuous percentages does not establish these claims. None of the external sources below validates Learnality’s item bank or recommendation rules.

Review method and boundaries
Read blend-evidence-design.txt, learning-profile.js, personality-reflection.js, subject-readiness.js and assessment-coverage-and-validation.md. Searched official measurement and developmental resources and primary study publications. Opened the listed sources; access limitations are stated. This is a targeted evidence audit, not a registered exhaustive systematic review. Sources were selected for direct relevance, opposing evidence, population applicability and transparent design. No school data, educational calibration, expert content panel or efficacy experiment has been performed by this audit.

SOURCE LEDGER

M1. AERA / APA / NCME (2014), Standards for Educational and Psychological Testing.
URL: https://www.testingstandards.net/uploads/7/6/6/4/76643089/9780935302356.pdf
Type/population: professional consensus standards, not an intervention study or student sample. Full official PDF opened and relevant passages inspected.
Relevant locations: Standards 1.12-1.15 (printed pp. 26-27), 1.24 (p. 30); chapter 3; 5.6, 5.12-5.20 (pp. 103-107); educational growth discussion (p. 185).
Evidence: profile interpretations and small item subsets need their own support; routing people to different treatments requires differential-outcome evidence where feasible. Access barriers and subgroup score meaning matter. Interchangeable alternate forms, common scales over time and changed versions require comparability evidence. Growth models have differing assumptions and can be distorted by missingness.
Limits: these standards specify responsibilities; compliance with software checks is not psychometric validity. They provide no Learnality sample size or universal reliability threshold.
Implication: freeze intended interpretations and demonstrate the evidence needed for each consequential decision before release.

M2. Lira et al. (2022), Large studies reveal how reference bias limits policy applications of self-report measures.
URLs: https://doi.org/10.1038/s41598-022-23373-9
https://pmc.ncbi.nlm.nih.gov/articles/PMC9649615/
https://pubmed.ncbi.nlm.nih.gov/36357481/
Type/population: primary three-study investigation. Indexed full article retrieved: Mexican high-school cohorts; U.S. grades 8-12 school data; U.S. senior-year questionnaire/task measures followed to college graduation. Full-page opening encountered a browser challenge, so accessible indexed article text was used and the access failure is retained.
Evidence: stronger peer performance corresponded with lower self-ratings and higher implicit standards; behavioral measures did not show the same reference-bias pattern. A changing reference group can change questionnaire answers without equivalent behavioral change.
Limits: adolescent self-regulation/grit/conscientiousness, not Learnality’s eight roles; observational associations do not directly measure each individual’s bias. Childhood and other contexts need investigation.
Implication: self-report is useful student voice, but cannot be the sole annual-growth signal, school-ranking measure or automatic teaching prescription. Retake displays need context and original responses.

M3. Rogowsky, Calhoun & Tallal (2015), Matching Learning Style to Instructional Method: Effects on Comprehension.
URL: https://www.apa.org/pubs/journals/features/edu-a0037478.pdf
DOI: https://doi.org/10.1037/a0037478
Type/population: primary study; PDF fully opened. College-educated native-English-speaking adults ages 25-40, excluding reported neurological/learning impairments. The aptitude investigation involved 121 people. The instructional investigation randomized 61; its categorical style analysis used 41, a materially smaller group.
Design: style questionnaire, counterbalanced reading/listening aptitude measures, randomized audiobook versus e-text instruction, common written comprehension immediately and two weeks later.
Evidence: no statistically significant preference-by-method relationship supporting the meshing claim on immediate or delayed comprehension.
Limits: small categorical cells, selected adult population, narrow content and written outcomes. A nonsignificant result does not prove every possible personalization is ineffective or settle accommodations.
Implication: Learnality must test actual routing benefit; preferred presentation cannot be treated as demonstrated effectiveness.

M4. Rogowsky, Calhoun & Tallal (2020), Providing Instruction Based on Students’ Learning Style Preferences Does Not Improve Learning.
URL: https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2020.00164/full
DOI: https://doi.org/10.3389/fpsyg.2020.00164
Type/population: primary study, publisher article opened and methods inspected; fifth graders ages 10-11 at one rural Pennsylvania school. Attendance/data availability reduced analysis samples; there was not one identical sample for every comparison.
Design: alternate reading/listening passages, random counterbalanced form/modality order, written comprehension questions and preference inventory.
Evidence: no differential modality benefit according to assessed auditory/visual preference. Preference group differences in overall comprehension did not establish a matching advantage.
Limits: one school, narrow reading/listening domain, immediate outcomes and particular inventory; not a universal curriculum trial or a delayed-learning study.
Implication: younger-population evidence also challenges simplistic preference matching. It does not invalidate tailored prerequisite instruction or verified accessibility supports.

M5. Hausknecht et al. (2007), Retesting in selection: a meta-analysis of coaching and practice effects for tests of cognitive ability.
URL: https://pubmed.ncbi.nlm.nih.gov/17371085/
DOI: https://doi.org/10.1037/0021-9010.92.2.373
Type/population: research synthesis, primary article abstract opened; personnel-selection cognitive testing, not a Learnality school sample. The abstract reports 107 samples and 134,436 participants.
Evidence: scores improved on retesting; effects were greater with coaching and identical forms.
Limits: employment selection and cognitive tests differ from educational learning episodes. The aggregate effect cannot be used to subtract a correction from a Learnality student’s score. Abstract-level review does not establish all moderator details.
Implication: log exact form/content exposure, interval, help and practice. Fresh-looking questions alone are not proven equivalent forms; a higher retake score cannot automatically mean growth.

M6. NAEYC, DAP: Observing, Documenting, and Assessing Children’s Development and Learning.
URL: https://www.naeyc.org/resources/position-statements/dap/assessing-development
Type/population: official developmental-practice guidance for birth through primary grades, not a randomized efficacy experiment. Page opened and assessment guidance read.
Evidence: assessment should be ongoing, authentic, culturally/linguistically responsive and embedded in routines, play and work samples. Children demonstrate skills unevenly across settings. Familiar assessors, family information and multiple methods matter; digital assessment should be limited for young children. Screening is not diagnosis.
Limits: guidance does not validate a particular app or define a universal developmental scale.
Implication: Pre-K needs an adult-supported observational mode and separate developmental task development. A teacher-free high-school questionnaire cannot simply be resized for preschool; observer reports must remain distinguishable from child performance.

M7. Roll & Barhak-Rabinowitz (2023), Measuring self-regulated learning using feedback and resources, OECD.
URL: https://www.oecd.org/en/publications/innovating-assessments-to-measure-and-support-complex-skills_e5f3e341-en/full-report/component-16.html
Type/population: official assessment-design chapter; no single experimental student sample. Full chapter opened.
Evidence: describes using digital learning environments and resource/feedback opportunities to elicit evidence of self-regulated learning under an evidence-centered design. Resource-use actions need interpretation in task context and in relation to affect and effort.
Limits: design rationale is not validation of Learnality, nor permission to equate a hint click with a trait. Behavioral traces can have rival explanations.
Implication: explicitly link claim, observable evidence and task opportunity; record why support was used rather than awarding a universal strategy score for interactions.

M8. Pashler et al. (2008 volume; online 2009), Learning Styles: Concepts and Evidence.
URL: https://www.psychologicalscience.org/journals/pspi/j.1539-6053.2009.01038.x/
DOI: https://doi.org/10.1111/j.1539-6053.2009.01038.x
Type/population: commissioned research review; official journal summary opened. Diverse prior style theories and studies, not a single cohort or new Learnality trial.
Evidence: preferences exist, but the review found virtually no adequate evidence of the interaction needed to justify matching instruction to assessed styles.
Limits: historical review, not an exhaustive review through 2026; does not establish that all individual differences or personalization are unhelpful. The direct studies above provide later, bounded tests.
Implication: eight strategy labels are proposed vocabulary. Calling them a blended personality supplies no missing empirical evidence.

M9. NCES (2020), ECLS-K:2011 Third-, Fourth-, and Fifth-Grade Psychometric Report, NCES 2020123.
Official landing page opened:
https://ies.ed.gov/use-work/resource-library/report/technicalmethodological-report/early-childhood-longitudinal-study-kindergarten-class-2010-11-third-grade-fourth-grade-and-fifth
Report URL: https://nces.ed.gov/pubs2020/2020123.pdf
Type/population: technical report on direct and indirect assessments in the ECLS-K:2011 grade 3-5 data collections. The PDF failed to open; do not claim a full methods audit of this report. The landing page verifies its scope and March 2020 publication.
Implication: potential implementation example for a later linking-study design review. No specific equating method or numerical reliability result is imported into Learnality from this inaccessible PDF.

LOCAL AUDIT - OUR OBSERVATIONS, NOT EXTERNAL RESEARCH RESULTS
1. learning-profile.js transforms ordinal 0-4 endorsements into an independent maximum-point percentage per role. Missing, invalid and “varies” are excluded and completeness is required. The numeric transform is transparent, but equal psychological spacing and trait precision are unestablished. Two items per role provide extremely sparse evidence. The current one-decimal display can imply more resolution than the response combinations actually offer.
2. Reflection items are desirable strategy statements. Several overlap in checking/revision and metacognitive control. There is no evidence that exactly eight distinct factors exist. “Team Synthesizer” includes independently explaining after discussion, which is not itself evidence of synthesis with others. Context, opportunity, language, social desirability and prior teaching can influence answers.
3. Subject readiness separates starting knowledge, learning, transfer and delayed evidence and labels forms uncalibrated. Different task phases are not directly subtractable into a gain. Equivalent criteria still need content/difficulty review and rater studies.
4. Historical coverage audit contains a retired threshold rule, preceded by a supersession warning. A consolidated current technical manual should prevent old and current scoring descriptions from being read as simultaneously operative.
5. Immutable records support auditability; they do not establish score validity. Browser/client help logs cannot prove absence of outside help.

RELEASE GATES - PROPOSED LEARNALITY REQUIREMENTS
These are our operational research requirements inferred from the evidence and intended product decisions, not universal numeric cutoffs supplied by a paper.

A. Construct and claim gate
Publish a versioned technical blueprint: exact intended inference, grade/domain, observable behavior, rival explanations, scoring criteria, consequences and prohibited uses for every reported dimension. Have measurement, subject, developmental and accessibility experts review it. Retain a possible result that the eight-role structure must be revised or removed.

B. Response-process gate
Conduct cognitive interviews across the actual pilot population: grades 9-12, language and access variation, lower-engagement students and varied prior attainment. Observe item interpretation, response mapping and task strategies. Revise ambiguities before expanding the bank. No fixed participant count is declared sufficient in advance without the design and decision requirements.

C. Scoring/precision gate
Double-score appropriately sampled artifacts using trained raters and explicit criterion decisions. Examine disagreement and task/occasion/rater variation. Study scale structure, conditional precision, missingness and item dependence. Set reporting/stopping criteria from the actual decision error costs, not a convenient universal reliability number. Automated scoring stays pending until it is evaluated against defensible human criteria and subgroup evidence.

D. Retake and growth gate
Preserve every original response and instrument/scoring version. Log previous content, practice, support, timing and administration conditions. Establish form comparability and longitudinal score meaning before numeric growth claims; investigate item/function drift, response shifts and missingness/attrition. Until then show exact dated observations and self-report changes, not improvement in general learning ability.

E. Fairness/access gate
Verify keyboard, screen reader, motion, text, device and low-connectivity access in authentic administrations. Separate target knowledge from irrelevant reading, motor, language, sensory or interface demands. Review accommodations by intended construct; do not remove writing when writing is the objective. Analyze relevant subgroup score meaning and prediction with sufficient, appropriately authorized data. Treat insufficient subgroup evidence as uncertainty, not proof of no bias. Access needs never depend on a role percentage.

F. Teaching decision gate
Pre-register prospective comparisons of high-quality existing instruction, broadly evidence-informed supports and profile-guided plans using common independent mastery, delayed retention and fresh transfer outcomes. Account for subjects, prerequisite knowledge, classes/schools, adherence, time, prior exposure and attrition. Hold out independent people/sites/time for evaluation; examine who benefits differentially and whether the complete routing policy improves outcomes. Engagement, confidence and visual reward are separate outcomes. Publish mixed/null results and educational harms as well as benefits.

G. Developmental extension gate
Keep the first pilot within grades 9-12. For Pre-K and early elementary, create short authentic play/observation protocols with educators/families and language-sensitive evidence collection. Validate their distinct interpretation and adult involvement. Do not promise one identical independent online test, common eight-part personality scale or cross-age percentage growth from Pre-K through high school without developmental linking evidence.

H. Publication/operating gate
Maintain an accessible claim-to-evidence ledger including negative findings, local validation status, versions and conflicts of interest. Revalidate consequential changes to tasks, scoring, animations or recommendation rules. No marketing phrase may imply perfect matching, validated diagnosis, objective personality percentages, replacement of professional judgment or guaranteed retention.

Permitted present claim
“Learnality is a research-informed project for student reflection, defined learning challenges and educator-reviewed support planning. Its current percentages summarize self-reported strategy endorsement. Its assessments and adaptation rules require school-based validation before claims of accurate individual curriculum matching or calibrated growth.”

LEARNALITY - INSTRUCTION, PERSONALIZATION AND AI EVIDENCE AUDIT
Research audit dated October 14, 2026. Draft scientific design guidance; not validation of Learnality.

SCOPE AND SEARCH TRANSPARENCY
Targeted critical review, not an exhaustive systematic review or meta-analysis. Searched for primary experiments, newer learning-style evidence, classroom retrieval research, readiness-based adaptive systems, self-regulation training, official instructional guides and AI tutoring trials. Exact searches included learning style matching Rogowsky; Clinton-Lisell Litzinger 2024; Bastani generative AI guardrails; Kestin AI tutoring RCT; Cognitive Tutor Algebra I scale/addendum; RAND personalized learning; Roediger Agarwal classroom retrieval; Cepeda optimal spacing; Azevedo Cromley self-regulation; IES preschool mathematics and adolescent literacy. Searches used official publishers, author/institution repositories and IES/RAND. Blogs, social-media posts and product marketing were not used to establish efficacy. Selected sources include adverse/null findings. Full-text availability is recorded below; inaccessible papers are not represented as fully read.

THE CENTRAL DISTINCTION
There are three different claims: (1) an instructional practice helps students on average; (2) instruction adjusted to subject prerequisites, progress or access needs helps in a particular program; (3) this student's Learnality profile predicts which method helps them more than alternatives. Evidence for (1) or (2) cannot establish (3). Changing categories to continuous percentages does not bridge that gap. All eight named Learnality dimensions, their measurement, decision rules and longitudinal interpretation require direct evidence.

SOURCE LEDGER
Summaries are deliberately short. Exact source links permit independent checking. Source-derived summaries below stay under 100 words per source; subsequent design requirements are our proposed implications, not findings attributed to individual papers.

I01. Pashler, McDaniel, Rohrer & Bjork, Learning Styles: Concepts and Evidence (2008 volume, online 2009), review.
https://doi.org/10.1111/j.1539-6053.2009.01038.x
https://www.psychologicalscience.org/journals/pspi/j.1539-6053.2009.01038.x/
Access: official APS synopsis opened; prior Learnality evidence design read. Boundary: historical review, not a complete current review.
Finding: Preferences exist, but a curricular matching claim needs randomized alternative methods and common outcomes demonstrating learner-by-method differences. The review did not find an adequate general matching evidence base. It does not establish that all personalization is useless. Newer evidence below must qualify any statement that matching has “no evidence.”

I02. Rogowsky, Calhoun & Tallal (2015), Matching Learning Style to Instructional Method: Effects on Comprehension, primary study.
https://www.apa.org/pubs/journals/features/edu-a0037478.pdf
https://doi.org/10.1037/a0037478
Access: complete 15-page publisher PDF opened; methods, sampling, immediate/two-week outcomes inspected.
Finding: 121 college-educated English-speaking adults aged 25-40 completed aptitude tests; 61 received randomized audiobook/e-text instruction, with 41 clearly categorized participants in matching subgroups. Preference did not significantly predict comprehension aptitude or matching benefits at immediate/two-week tests. Restricted adult sampling, exclusions and small subgroups limit generalization. A null finding does not prove every possible learning-profile interaction is zero.

I03. Clinton-Lisell & Litzinger (2024), Is it really a neuromyth? A meta-analysis of the learning styles matching hypothesis.
https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2024.1428732/full
https://doi.org/10.3389/fpsyg.2024.1428732
Access: full article opened; methods, coding, aggregate effects, crossover, moderator, sensitivity and limitations inspected.
Finding: 21 studies/101 effects yielded g=.32 (95% CI .07-.57), I²=91%. Directional crossover appeared in 11/42 outcomes; most crossover studies failed coded quality criteria. Five studies met standards; 85% excluded ambiguous-style participants. Removing two quasi-experiments retained a positive aggregate. Quality moderators were not significant; publication bias tests were negative, not proof of absence. Authors questioned widespread accommodation. These modality findings do not validate eight Learnality dimensions. Directional crossover tallies are not automatically statistically established, replicated individual policy benefits.

I04. Roediger, Agarwal, McDaniel & McDermott (2011), Test-enhanced learning in the classroom: Long-term improvements from quizzing, primary classroom experiments.
https://pdf.retrievalpractice.org/guide/Roediger_Agarwal_etal_2011_JEPA.pdf
https://doi.org/10.1037/a0026252
Access: complete 14-page article opened; manipulation, participant exclusions and long-term limitations inspected.
Finding: Within-student sixth-grade social-studies experiments at one suburban public school found benefits from low-stakes quizzing, including comparisons with restudy and delayed tests. Many absentees and special-education/gifted students were excluded; some long-delay estimates were too unstable to report. Reused facts/items limit broad-transfer interpretation. Retrieval is both an assessment and learning exposure; these results do not support personality-specific quizzes or unlogged repeated testing.

I05. Cepeda et al. (2008), Spacing Effects in Learning: A Temporal Ridgeline of Optimal Retention, primary experiment.
https://doi.org/10.1111/j.1467-9280.2008.02209.x
https://escholarship.org/uc/item/0kp5q19x
https://laplab.ucsd.edu/articles/Cepeda%20et%20al%202008_psychsci.pdf
Access: author/institution indexed manuscript text inspected; direct PDF open failed. Not claimed as complete visual inspection.
Finding: Over 1,350 online participants learned factual material with manipulated study gaps and later delays up to one year. Optimal gaps depended on the retention interval; longer gaps were not always better. Completion declined at long delays. These factual-memory findings support spaced review but do not identify each child's optimal schedule or justify a universal seven-day retention rule.

I06. IES/WWC (2007), Organizing Instruction and Study to Improve Student Learning, official evidence synthesis/practice guide.
https://ies.ed.gov/ncee/wwc/PracticeGuide/1
https://ies.ed.gov/ncee/wwc/Docs/PracticeGuide/20072004.pdf
Access: official recommendation page opened; indexed guide text inspected.
Finding: Recommendations cover spaced study, worked examples/problem solving, integrated graphic/verbal and abstract/concrete representations, retrieval, learning judgments and explanatory questions, with differing evidence ratings. The guide is a dated synthesis, not a Learnality validation study. A practice's rating and target situation must travel with its recommendation rather than treating every suggestion as equally proven.

I07. Azevedo & Cromley (2003 conference report; related 2004 journal study), self-regulated learning training with hypermedia, primary randomized study.
https://files.eric.ed.gov/fulltext/ED477464.pdf
https://eric.ed.gov/?id=EJ685011
https://doi.org/10.1037/0022-0663.96.3.523
Access: full 26-page conference manuscript opened; 2004 journal abstract inspected. Conference and published report must not be counted as two independent trials.
Finding: 131 undergraduates were randomized to 30-minute self-regulation training or no training before circulatory-system learning. Training improved conceptual mental-model changes, with planning/monitoring strategy evidence. This was one topic, older hypermedia and immediate pre/post measurement. It supports teaching self-direction, not assuming students can manage an unstructured adaptive curriculum unaided or extrapolating adult effects to Pre-K.

I08. Klahr & Nigam (2004), The Equivalence of Learning Paths in Early Science Instruction, primary experimental comparison.
https://doi.org/10.1111/j.0956-7976.2004.00737.x
https://www.cmu.edu/dietrich/psychology/pdf/klahr/PDFs/KlahrNigam.PsychSci.pdf
Access: author-hosted publisher manuscript indexed text inspected; direct PDF retrieval failed.
Finding: Among 112 third/fourth graders studying control of variables, more learned the target procedure with direct instruction than discovery. Students who learned it through either route performed similarly on later science-poster evaluation. Comparing only successful learners is not an unconditional causal transfer comparison. One skill and age band cannot decide every instructional method; nevertheless self-driven should not mean guidance-free.

I09. Muralidharan, Singh & Ganimian (2019), Disrupting Education? Experimental Evidence on Technology-Aided Instruction in India, primary lottery experiment.
https://doi.org/10.1257/aer.20171112
https://assets.aeaweb.org/asset-server/files/9271.pdf
https://alejandroganimian.com/files/20190216_Disrupting_education.pdf
https://www.socialscienceregistry.org/trials/980
Access: full 95-page author manuscript/appendix opened through author's latest-manuscript link; publisher abstract and 68-page supplement also inspected.
Finding: Lottery access for 619 Delhi middle-school students produced math/Hindi gains over 4.5 months (intent-to-treat .37/.23 SD). Instruction adapted to preparation, bundling computer work, human support and additional time. Grade-level school math exams did not improve overall; weakest students did not improve school exams in either subject. Similar absolute independent-test gains do not demonstrate personality-by-format matching. Readiness remediation can improve prerequisite knowledge without immediately delivering grade-level attainment; setting, language and implementation constrain US transfer.

I10. Pane et al. (2014) Cognitive Tutor Algebra I at Scale, author addendum and official WWC review.
https://doi.org/10.3102/0162373713507480
https://www.rand.org/content/dam/rand/pubs/working_papers/WR1000/WR1050/RAND_WR1050.pdf
https://ies.ed.gov/ncee/wwc/Docs/InterventionReports/wwc_cognitivetutor_122116.pdf
https://ies.ed.gov/use-work/awards/effectiveness-cognitive-tutor-algebra-one-implemented-scale
Access: full 8-page author addendum opened; official WWC tables/methodological summary inspected.
Finding: Large US school-randomized blended Algebra I study had no first-year effect; authors' fully adjusted second-year high-school estimate was +.21 SD (p=.04), while simpler adjusted analysis was not significant. Post-assignment student selection and attrition complicate inference. WWC reported a nonsignificant overall algebra result and reservations. Report both analyses; a program average, even positive, does not validate Learnality's routing or remove educator support.

I11. RAND (2015), Continued Progress: Promising Evidence on Personalized Learning, primary evaluation report.
https://www.rand.org/content/dam/rand/pubs/research_reports/RR1300/RR1365/RAND_RR1365.pdf
Access: complete 57-page report opened; matched comparison methods, grade counts and limitations inspected.
Finding: Evaluation of 62 supported schools used matched virtual comparisons rather than randomized school adoption. Positive aggregate growth was promising; matching cannot eliminate unobserved selection. Early-adopter schools and a predominantly elementary sample limit generalization to typical high schools. Multi-component school models do not isolate profile matching, software, teaching or additional support effects.

I12. RAND (2017), Informing Progress: Insights on Personalized Learning Implementation and Effects, primary evaluation report.
https://www.rand.org/pubs/research_reports/RR2042.html
https://www.rand.org/content/dam/rand/pubs/research_reports/RR2000/RR2042/RAND_RR2042.pdf
Access: full 59-page report opened; achievement, school variability and implementation limitations inspected.
Finding: One-year matched comparison covered 32 schools/about 5,500 students, mostly secondary settings newer to implementation. Math estimate .09 SD was significant; reading .07 was not. School results varied, including negative estimates. Constraints included resources, time, student pacing and curriculum. This is mixed quasi-experimental evidence for whole-school models, not perfect customization or a causal eight-dimension assignment rule.

I13. Bastani et al. (2025), Generative AI without guardrails can harm learning: Evidence from high school mathematics, primary randomized trial.
https://doi.org/10.1073/pnas.2422633122
https://pmc.ncbi.nlm.nih.gov/articles/PMC12232635/
https://github.com/obastani/GenAICanHarmLearning
Access: publisher search-indexed full sections inspected; direct publisher/PMC opens hit cookie/403/reCAPTCHA. Methods/limits from primary text; not represented as successfully fetched complete paper. SSRN author abstract also inspected.
Finding: Nearly 1,000 Turkish grades 9-11 students in four 90-minute sessions received ordinary GPT-4, safeguarded tutoring or control. Assisted practice improved, but unrestricted AI reduced unassisted exam performance; safeguards largely mitigated the harm without establishing a general independent-learning advantage. Short-term, single-school mathematics and 2023 model/version limit generalization. Tutor solutions/common-error guidance were problem-specific. Assisted success is not evidence of retention.

I14. Kestin et al. (2025), AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design, primary randomized crossover trial.
https://www.nature.com/articles/s41598-025-97652-6
https://doi.org/10.1038/s41598-025-97652-6
Access: full publisher article opened; eligibility, design, outcomes and supplied-solution safeguards inspected.
Finding: 194 eligible Harvard introductory-physics students experienced two lessons with swapped AI-home/in-class active-learning conditions. AI yielded higher immediate posttest learning and shorter median time; the design bundled tutor, setting and self-pacing. Eligibility depended on participation/completion; no broad annual retention result was established. Correct solutions/scaffolds were engineered. Undergraduate two-topic efficacy does not establish autonomous US high-school or Pre-K success, nor profile-specific effects.

I15. IES/WWC (2008), Improving Adolescent Literacy, official practice guide.
https://ies.ed.gov/ncee/WWC/PracticeGuide/8
https://ies.ed.gov/ncee/wwc/docs/practiceguide/adlit_pg_082608.pdf
Access: official recommendations and guide sections indexed text inspected.
Finding: For grades 5-12, strong evidence ratings support explicit vocabulary, direct comprehension-strategy instruction and specialist interventions for struggling readers; discussion/motivation recommendations receive moderate ratings. Modeling, guided practice and feedback lead toward independence. Learnality should preserve necessary instruction and refer literacy needs to educators/specialists rather than interpreting them as a personality weakness.

I16. IES/WWC (2013), Teaching Math to Young Children, official practice guide.
https://ies.ed.gov/ncee/wwc/practiceguide/18
https://ies.ed.gov/ncee/wwc/Docs/practiceguide/early_math_pg_111313.pdf
Access: official recommendation page and evidence-rating definitions inspected.
Finding: Preschool/Pre-K/K recommendations emphasize developmental progressions and teaching connected to children's knowledge. Number/operations has moderate evidence; several other recommendations, including progress monitoring, have minimal evidence in this guide. Age-specific tasks and adult observation matter. This does not validate independent touchscreen personality assessment in young children or imply that recommendations labeled minimal are worthless.

PROPOSED PRODUCT RULES FROM THIS AUDIT (DESIGN INFERENCES)
1. Keep student voice, observed strategy, objective readiness and causal support response as separate evidence panels. Never sum them into one personality certainty score.
2. Use validated curricular prerequisites/errors and ongoing independent outcomes for routing before preference-based method guesses. A preference may choose a feasible starting option; it must not lower subject coverage or lock out representations.
3. Give broadly helpful practices as the default with content-specific conditions: retrieval with feedback, spaced opportunities, appropriate examples, accurate explanations and progressive guidance. These are not special benefits restricted to matching avatars.
4. Compare a future profile-guided policy against a strong evidence-informed default, not only a weak control. Otherwise gains could reflect better teaching generally rather than useful personalization.
5. Protect independent attempts. Separate practice assisted by AI/teachers from unassisted common-criterion learning checks. Keep help/exposure/date logs; prior assessment itself changes future learning.
6. Build constrained, expert-approved content units with prerequisite maps and equivalent assessment blueprints. Any-curriculum ingestion needs factual and standards review; AI-generated variants do not automatically retain cognitive demand or measurement equivalence.
7. Support self-direction by explicit goal setting, worked demonstrations, guided practice, checkpoints, reflection, useful feedback and teacher escalation. Fade scaffolding from observed success, not by assigning a supposedly independent personality.
8. Personality language should be provisional and empowering. Describe “strategies you are using in this context,” with growth opportunities, and avoid claims that a student has an immutable brain type or cannot learn through a low-index strategy.
9. Delay and transfer are different. A new-context item is not automatically far transfer. Specify distance, prerequisite burden, response mode and common rubric; broad transfer claims need much wider sampling.
10. Pre-K requires a separate developmental program and validation, including adult-facilitated/play-based observations, language/access checks and burden limits. A reworded high-school questionnaire is not an annual common scale.

RESEARCH STILL REQUIRED FOR LEARNALITY
• Content experts must define role claims independently of branding, audit rival explanations and demonstrate adequate task coverage across subjects.
• Cognitive interviews must verify interpretation, especially confidence/frequency/missing/“depends” responses, reading burden and avatar-label effects.
• Pilot scoring studies must examine item/occasion/rater error, alternate-form difficulty, response access and subgroup fairness. Two reflection items per role cannot provide fine-grained prescription accuracy.
• Prospective support trials need common delayed/transfer outcomes, credible comparison policies, logged exposures and preregistered learner-by-support analyses. More clicks, completion or preference ratings cannot establish learning.
• Repeated individual trials need matched fresh content, order/carryover checks and uncertainty-aware decisions. A single winning method on one topic is not a stable student-wide truth.
• Curriculum policy evaluation must test retained coverage, rigor, access, teacher burden, misrouting and harms, with external replication and held-out schools/classes/time.
• Annual growth and unlimited retakes need stable anchors/linked measures and practice-effect handling before claiming comparable progress over grades. A changed reflection does not overwrite demonstrated knowledge or needed accommodations.
• US school-day use across grades 9-12 and broad subjects remains a separate efficacy question; international after-school trials, undergraduate physics and selected elementary schools do not resolve it.
• Source updates are necessary as AI and educational evidence change. Provider/model/prompt/content versions belong in records. Do not describe this targeted audit as “fully researched forever.”

PUBLIC CLAIM BOUNDARY
Permitted: “Learnality is informed by research on learning and assessment, and is being developed to help students and educators evaluate strategies, readiness and supports.”
Not established: validated learning-personality percentages; reliable identification of each student's best method; perfect adaptation of any curriculum; equivalent personalized tests; guaranteed engagement; verified Pre-K-12 growth scale; teacher-free AI learning; or Learnality-specific learning gains.

LEARNALITY - ENGAGEMENT, PERSONALITY PRESENTATION AND LEARNING EVIDENCE
Research audit • October 14, 2026 • Target: school-mediated US grades 9-12
This is a targeted evidence review, not a preregistered exhaustive systematic review. Source access and limits are explicit. Recommendations below are design inferences requiring local testing, not validated Learnality effects.

E1. Sailer & Homner (2020; online 2019), The Gamification of Learning: a Meta-analysis
https://doi.org/10.1007/s10648-019-09498-w
Access: full publisher HTML, including methods, robustness and limitations.
The synthesis distinguishes cognitive, motivational and behavioral outcomes. It reports positive pooled effects, with heterogeneous implementations and less stable motivational/behavioral results in the more rigorous subset. Game fiction did not significantly distinguish cognitive or motivational effects. Some behavior measures were collected during intervention. This does not establish that Learnality avatars improve retention, that animation intensity has a dose benefit, or that completion measures equal learning.
Design inference: test optional cosmetic progress against a simpler interface using common learning outcomes and access measures; measure engagement separately.

E2. Dai, Xu & Xing (2025), Gamified learning impact: a meta-analysis of game element combinations on students’ learning outcomes
https://doi.org/10.1007/s11423-025-10493-y
Access: publisher abstract/metadata; subscription full text not inspected. Not treated as a fully appraised source.
The abstract reports a synthesis of randomized/quasi-randomized trials and positive average outcomes with domain/duration moderators. Its reported best combination is not an experimental demonstration that those elements will be optimal for every student or Learnality. Detailed bias/dependence assessment remains outstanding.
Design inference: newer positive evidence belongs in the ledger, but no avatar prescription or ranked game-element policy is adopted from abstract-only evidence.

E3. Sundararajan & Adesope (2020), Keep it Coherent: A Meta-Analysis of the Seductive Details Effect
https://doi.org/10.1007/s10648-020-09522-4
Access: publisher abstract/metadata; full subscription text not inspected.
The synthesis reports that interesting irrelevant material can hinder learning, with effects moderated by design and methodology. This is a reason to investigate distraction, not proof that every decoration is harmful.
Design inference: celebrate at task boundaries, permit reduced motion, keep celebrations out of evidence items, and evaluate comprehension/interruption as well as enjoyment. Full-text appraisal is still needed before detailed mechanism claims.

E4. Deci, Koestner & Ryan (1999), A Meta-Analytic Review of Experiments Examining the Effects of Extrinsic Rewards on Intrinsic Motivation
https://selfdeterminationtheory.org/wp-content/uploads/2014/04/1999_DeciKoestnerRyan_Meta.pdf
Access: author-hosted PDF retrieved; indexed primary-PDF abstract inspected; full PDF was not text-appraised. This audit does not claim a complete reanalysis of the reward literature.
The authors distinguish expected tangible/contingent rewards from positive feedback and report different effects on intrinsic interest/free-choice behavior. This older, contested literature is not a direct avatar trial or a measurement of student dopamine.
Design inference: use informative progress and voluntary customization, avoid making instructional access depend on cosmetics, and test sustained willingness to learn. Do not claim “dopamine dumps,” addiction mechanisms or universal motivation gains.

E5. Deslauriers et al. (2019), Measuring actual learning versus feeling of learning in response to being actively engaged in the classroom
https://doi.org/10.1073/pnas.1821936116
Access: full primary-paper HTML; methods and randomized crossover design inspected.
In introductory Harvard physics, students experienced active and passive instruction across two class topics. Actual learning and perceived learning diverged: active instruction produced better test performance while feeling less effective. Population/content are specific; this does not prove identical results in every school or identify an individual’s best format.
Design inference: enjoyment, ease and confidence are student voice; assess retained understanding independently. Product adaptation must not chase perceived fluency alone.

E6. 16Personalities, Reliability and Validity
https://www.16personalities.com/articles/reliability-and-validity
Access: vendor’s complete public article, including consistency and retest descriptions.
This is the developer’s own report about its English personality instrument. Its coefficients and samples cannot be borrowed as evidence for Learnality, child assessment, academic mastery or curriculum matching. A consistent personality questionnaire and a useful instructional assignment policy are different claims requiring different evidence.
Design inference: borrow only the idea of an inviting experience and understandable language. Create original artwork/items. Do not borrow types, scores, item content, norms or accuracy claims. The eight Learnality roles have not been demonstrated to be distinct psychometric factors.

ENGAGEMENT STUDY TO RUN, NOT A CLAIM OF SUCCESS
Compare simplified and richer cosmetic interfaces while holding task content, access, instructions and feedback rules constant. Randomize/counterbalance presentation where appropriate; record novelty, device, accessibility settings and prior exposure. Collect completion, voluntary return, interruption, fatigue, student interviews, and independent immediate/delayed/transfer performance as separate outcomes. Assess learners who discontinue, not only completers. Explain effortful learning without rewarding correct guessing or shaming support use. Fix a primary outcome and analysis before examining results; no retrospective claim of a best costume for a type.

ACCESS REQUIREMENTS
Cosmetics are optional and independent of ability/grades. Keyboard and non-drag alternatives, reduced motion, readable layouts and lossless pause/resume require actual student and assistive-technology checks. An attractive screenshot is not accessibility conformance or a study of disengaged students. Pre-K animation/reading demands need a separate development and supervision design.

CURRENT LIMIT
No Learnality engagement experiment, neural measurement, randomized avatar study or sustained participation study has occurred. Evidence supports careful evaluation of the idea; it does not establish that the current animations or role language improve learning.

LEARNALITY - ASSESSMENT TECHNICAL BLUEPRINT
Version blueprint-development-1 • October 14, 2026
First intended population: U.S. grades 9-12 through participating schools.
Status: proposed development specification. No expert approval, factor validation, school administration, calibrated score or instructional benefit is established.

1. INTENDED USE AND SCORE CONTRACT
The assessment begins an evidence record used by students and assigned educators to propose and evaluate bounded teaching supports. It is not an annual once-and-for-all determination of a student's best teaching mixture. Annual sittings, contextual reflection and recurring authentic work jointly inform later decisions.
Permitted current outputs: exact self-report responses/endorsement points; task-bound closed outcomes; independently reviewed criterion judgments; support/exposure/missingness; proposed trials with rationale; comparable outcomes only to the extent actually demonstrated.
Prohibited outputs: objective percentage of a personality, permanent exclusive type, global intelligence/learning ability, diagnosis, placement or eligibility decision, inferred disability/attention from behavior, calibrated growth without linking, optimal instructional weights, guaranteed engagement/retention, or automatic removal of accommodations/required content.
The eight strategy names are provisional student-facing vocabulary. They are not established independent factors. Retain the option to merge, rename or remove roles after expert and empirical review. Cosmetic avatar choice supplies no assessment evidence.
Current reflection percentages equal endorsement points/maximum available points after complete numeric answers to both role items. Missing/invalid/depends states are explicit. Equal psychological spacing of ordinal answers is unestablished. All eight can be high together; do not normalize to total100 or blend preferences into criterion mastery.

2. EVIDENCE LAYERS AND UNIT OF OBSERVATION
A Student voice: reported frequency, contextual preferences, aspirations, opportunity and optional access context. No grading.
B Strategy artifact: response showing a specified process in a specified task, with rival explanations and limitations.
C Curricular learning: prerequisite knowledge, acquisition, independent mastery, delayed retention and bounded transfer for an explicit objective.
D Support responsiveness: prospective comparison of allowable supports on common outcomes. An effect requires an appropriate causal design; a single improvement after help is not an individual treatment effect.
Keep A-D separately visible. A policy may use approved evidence from multiple layers but must retain the source and uncertainty of each; no unvalidated composite certainty score.
One scenario with calculation/category/explanation is a task family with dependent components, not three independent measures. Repeated revisions are an attempt sequence, not additional independent successes. Rater judgments, occasions, learners and content families have distinct denominators.

3. CLAIM / EVIDENCE / TASK MATRIX
Every row below has a bounded claim, current mappings, additional task opportunity, example analytic evidence and rival explanations. New opportunities and criterion examples are proposals for review, not new questions or validated rubrics.

S1 PATTERN SCOUT - proposes and tests a candidate rule against relevant examples in this task.
Current student voice: reflection-01,09.
Current artifacts: exp-pattern-first/test/limit/challenge/rule/revise; exp-pattern-delay-counter/explain. Core m2 supplied-rule application is not independent rule discovery.
Needed opportunity: another unrelated content family requiring a testable hypothesis, informative counterexample and warranted revision; permit more than one defensible rule when observations underdetermine it.
Analytic example: hypothesis fits supplied observations; proposed new case has contrasting predictions; revision acknowledges which alternative was rejected without claiming unique proof. Judge these separately; missing opportunity is unobserved.
Rivals: arithmetic/reading familiarity, memorized puzzle, answer-option cues, teaching already given, perceptual access and prior exposure. Prohibited inference: general pattern giftedness or an optimal diagram learning method.

S2 MODEL MAKER - represents specified relationships accurately and translates between representations.
Voice: reflection-02,10.
Artifacts: exp-model-build/explain/translate and closed model-order/error/change. Core m3 comparison of supplied representations is narrower than constructing a model.
Needed: authentic diagram/data/model artifact plus independently evaluated translation in a second subject. Offer text-equivalent access without assuming a drawing interface measures the same construct.
Analytic example: required entities and directional/conditional relations present; unsupported relations absent; explanation preserves dependency when representation changes. Example current criterion model-connections expects Source→Relay→Receiver and Receiver→Source with distinct message/acknowledgement labels.
Rivals: writing/motor drawing skills, convention familiarity, software access, working-memory demand and reading. Prohibited: visual learner classification or assessment of artistry when relationships are the target.

S3 EVIDENCE TRACKER - connects a claim with relevant evidence and warranted limits.
Voice: reflection-03,11.
Artifacts: m5 made-up survey/absence/source claims; exp-pattern-limit and experiment-bounds can supply related but overlapping evidence, not independent role factors.
Needed: unrelated authentic source bundle with credibility, selection/confounding, changed evidence and a justified revised claim; objective and subject source knowledge explicit.
Analytic example: distinguishes surveyed from responding population; separates observed association from causation; states a specific limitation; revises proportionately rather than changing for its own sake.
Rivals: statistics vocabulary, topic familiarity, literacy and test-taking clues. Prohibited: global skepticism/credulity or permanent personality truth from one source problem.

S4 EXPLANATION BUILDER - communicates accurate necessary reasoning, sufficient for the objective.
Voice: reflection-04,12.
Artifacts: m2/m3/m6 constructed explanations; exp-model-explain; exp-calibration-explain/transfer-explain; subject-rate-* constructed artifacts.
Needed: blind analytic review across contrasting objectives, and evidence that the explanation supports warranted understanding/application; not merely fluent prose.
Analytic example: necessary mechanism present; conclusion warranted; distinction/assumption explicit. Separate reasoning accuracy from communicative sufficiency. Current archive-version distinguishes saved Monday snapshot from unsaved Tuesday edits. Grammar/length is excluded unless language production is the curricular objective.
Rivals: expressive language/assistive technology, rehearsed explanation, imitation/AI/outside help, writing fluency. Prohibited: eloquence=understanding or automatic instructional effectiveness of self-explanation for this learner.

S5 EXPERIMENT PILOT - plans an informative comparison and updates from its result under defined conditions.
Voice: reflection-05,13.
Artifacts: exp-experiment-control/difference/conclusion/revise/next/design; submitted prediction is student voice, not correctness evidence.
Needed: safe authentic prediction→chosen test→actual observation→revision sequence; second topic/occasion. School-approved materials only; no improvised physical experiments.
Analytic example: comparison changes focal factor while controlling relevant others; outcome/measurement specified; update uses results and limitations. Current experiment-design expects Q/no-Q comparison at new temperature with other factors held constant and seven-day growth measured.
Rivals: taught science procedures, numeracy, resource opportunity, reading a supplied study instead of conducting one. Prohibited: scientific identity, general hands-on superiority, or causal inference from uncontrolled attempts.

S6 ROUTE PLANNER - plans, monitors and revises a multi-step task against its explicit constraints.
Voice: reflection-06,14.
Artifacts: m4-plan/revision exhibition problem; exp-calibration-restudy is a choice, not evidence a chosen support worked.
Needed: multi-day authentic assignment with declared goal, feasible steps, checkpoint evidence and reasoned adaptation; preserve actual constraints and teacher help.
Analytic example: all required subtasks covered within time/resources; checkpoint inspects relevant progress; retains a valid plan or revises an invalid one with correct recalculation. Change itself is not a reward.
Rivals: calculation skill, organizational opportunity/support, device access, executive-function demands and external prompts. Prohibited: executive-function diagnosis, moralized effort score, or assuming independent learning requires no instruction.

S7 TEAM SYNTHESIZER - integrates contributions in genuine collaboration while showing individual understanding.
Voice: reflection-07,15. The latter independent explanation after discussion does not itself demonstrate synthesis.
Current artifacts: NO live human collaborative task in current bank; objective claim remains unobserved.
Needed: school-approved structured paired/group task with attributable contributions, distinct ideas, integration/reasoning record and separate individual follow-up. Alternatives and privacy/access reviewed before use; a chatbot or made-up peer is not an equivalent human collaboration measure.
Analytic example: identifies another contribution accurately; incorporates/contrasts it with reasons; resulting solution explicitly links ideas; independently applies relevant principle afterward. Score individual evidence separately from group product. Dominance/talk volume is not skill.
Rivals: partner expertise/status/language, friendship, task distribution, communication access and opportunity. Prohibited: introversion/extroversion, social worth or fixed group-work requirement.

S8 TRANSFER NAVIGATOR - recognizes applicability and limits of a learned principle in a specified unfamiliar setting.
Voice: reflection-08,16.
Artifacts: m3 changed rule/application; m6 boundary/checking; exp-calibration-transfer/transfer-explain/delay-transfer; subject-rate-* near-transfer artifacts.
Needed: explicitly varied surface features, deeper structural mapping and non-applicable case; independently rated prerequisites and novelty; more than one domain for broader scope.
Analytic example: identifies relevant correspondence; applies principle accurately; rejects attractive but invalid mapping; states conditions. Label near/further transfer descriptively from a reviewed task taxonomy rather than automatically calling changed wording far transfer.
Rivals: prior familiarity, reading/numeracy, cueing, exposure and shared task scaffold. Prohibited: universal transferable learning capacity from one analogy.

C1 STARTING KNOWLEDGE - demonstrates prerequisites for an explicit unit before its learning episode.
Current: m1 orientation/prerequisites; subject-readiness FormA/B three prerequisite components each.
Need: school-approved prerequisite map across target objectives; review ceiling/floor and gaps. Artifacts show relevant component accuracy and explanation, not course readiness from one topic.
Rivals: prior opportunity, language, item demand, unlogged assistance. Prohibited: personality/ability, course placement or whole-grade readiness.

C2 ACQUISITION / MASTERY - meets a defined objective after specified instruction under recorded conditions.
Current: m2 Route Lab taught rule/application, StudyLab made-up conditional systems, subject-rate immediate tasks.
Need: adequate coverage of objective content and demand with reviewed keys/rubrics and appropriate independent attempts. Starting/immediate phases have different tasks and cannot simply be subtracted into a gain score.
Analytic current subject criterion rate-method: change in output/change in input with output units per input unit; rate-boundary separately limits interpretation. Missing/uninterpretable not zero. One calculation and its ratio are dependent.
Prohibited: assisted success=independent mastery, universal subject proficiency, arbitrary masterypercentile.

C3 RETENTION - retrieves/applies a specified taught objective after recorded elapsed interval and subsequent exposure.
Current: delay-a/b; exp-pattern-delayed and exp-calibration-delayed; StudyLab delayed; subject-rate return. A 7 to 10 calendar day gap is a timing convention, not universally optimal interval.
Need: server administration/lesson anchors, school-calendar evidence where school days matter, authentic/simulated separation, logged interim quizzes/study and fresh comparable prompts. First unaided response and supported continuation distinct.
Rivals: intervening instruction, test exposure, timing/device errors, absence, outside help. Prohibited: long-term durability from early simulation, retained ability from submission counts, forgetting rate from non-equivalent forms.

D1 SUPPORT RESPONSIVENESS - specified support improves common independent outcomes in the defined context compared with a credible alternative.
Current: StudyLab four diagram/system/order assignments supply exploratory raw comparisons; hint success and format preference are NOT established effects.
Need: prospective randomized/counterbalanced comparable fresh units, fixed support versions/dose, prerequisites and actual exposure, common delayed/transfer outcomes and appropriate uncertainty. Test individual-policy benefit beyond average support effects.
Rivals: content/order/teacher/time/place differences, prior knowledge, carryover and differential dropout. Prohibited: winner after one A/B topic=best lifetime method or computed proportion of each optimal teaching style.

4. CURRENT BANK AND COVERAGE REGISTER
Core bank 1.0:6 initial missions+2 returns,77steps;49closed keys includes practice;38 scored closed components. Seven dimension tags do not equal eight independent roles: application→Pattern and checking→Explanation/Experiment are shortcuts, Team=null. These mappings must not drive objective role scores.
Advanced advanced bank 1.0:4initial families+2later missions,54steps,21closed components,12 criterion-bearing open artifacts. Distinct model/pattern/experiment content remains bounded and supplied; no broad representativeness.
StudyLab: a chosen assignment yields32steps,16closed components,4open explanations; four counterbalanced variants are four assignment possibilities, not four independent occasions per learner. Conditions/content/order and self-paced exposure remain confounded until studied.
Reflection learnality-reflection-16-v1/frequency-5-v1:16items,2per proposed role. Current independent endorsement display is descriptive only.
Subject subject-rate-1:2draft forms,34steps total,20dependent closed components and6open artifacts. Partial HSF.IF.B.6 rate from table/unit/constant-rate scope; graph estimation and wider functions absent. Same blueprint is not established equivalence. Authenticated criterion review currently covers these six artifacts, not every core/advanced open artifact.
All counts require re-audit after bank changes. Use source-derived inventory, never multiply steps by inferred reliability.

5. ADMINISTRATION, SAMPLING AND VERSION RULES
Register item/task/family/form/instrument/rubric/scoring/support/policy versions and hashes; approval status and intended grade/domain; objective/source/prerequisite/access references; explicit criteria and permitted interpretations. Content changes that alter demand or target create new versions and trigger review; old runs keep old keys/rubrics.
A session fixes authorized student, cycle, item/form assignment, presentation order and accommodation plan before presentation. Attempts retain first response, later revision, support kind (UI/access/conceptual), lesson/check exposure, device event/receipt times and pause/context. Outside help remains unverified unless genuine administration supports a narrower statement.
Fresh run IDs preserve prior artifacts and reviews; reflection revisit is distinct from task retake. Do not reuse an exposed family as clean unexposed evidence. Identical appearance changes are not alternate psychometric forms.
Balanced content sampling follows claim scope, dependencies, time/access burden and conditional precision; measurement experts select representative families/occasions and feasible coverage. Do not impose an arbitrary universal number of tasks, raters or opportunities. Adaptive stopping/routing rules require simulation and empirical decision-precision evaluation, not convenience.
Show unobserved/missing/skipped/access interruption/not applicable/incomplete and contradictory evidence separately. An unsubmitted criterion is not zero. A wrong relevant submitted answer may be incorrect; uninterpretable artifact remains unobserved. Explain denominator and scope in every report.
Open review requires trained rubric-specific raters, exemplar anchors including edge cases, independent blinded judgments before adjudication and preserved revisions. Proposed0/1/2examples here guide artifact review, not equal-interval total-trait scoring.
Prefer common independent outcomes for varied practice; differing accessible response modes require target-construct review and comparability evidence. Required accommodations persist regardless of score; a writing objective cannot silently become a drawing objective.
Annual reports show exact dated evidence until linking/form drift/response-shift evidence supports comparisons. Retakes propose plan diffs; issued work, grades and prior mastery are not erased.

6. DEVELOPMENT ACCEPTANCE GATES
Content/construct: subject, measurement and accessibility experts approve bounded claims, tasks, criteria and rival explanations; record disagreement and revision, including possible role removal.
Response process: cognitive interviews confirm intended reasoning and response meanings across actual users; independent navigation/access usability measured, including non-completers.
Scoring/precision: keys independently checked; raters studied; task/occasion/rater/dependence and conditional uncertainty evaluated for intended decisions. No universal numeric reliability threshold certifies validity.
Fairness/form/growth: authentic access and score meaning across relevant subgroups; sample/exposure-limited form comparability and longitudinal invariance/linking as applicable; insufficient data means restricted claim.
Instructional policy: held-out prospective benefit and harms versus strong evidence-informed practice, common independent retention/transfer and preserved coverage/rigor.
Public release: expert review/technical manual, supported population/version claims, legal/security governance and tested operational workflow. Code tests prove behavior only. Younger grades require a separate developmental blueprint and evidence, not universal factor extrapolation.

7. SOURCE BASIS AND LIMITS
This blueprint is a design inference from research-audit.txt, measurement-research-audit.txt, instruction-research-audit.txt, engagement-research-audit.txt and user-requirements-audit.txt. Exact populations/full text gaps/conflicting findings remain in those ledgers.
Professional framework: AERA/APA/NCME Testing Standards (2014), https://www.testingstandards.net/open-access-files.html . No numeric cutoff imported.
Evidence-centered digital assessment rationale: Roll/Barhak-Rabinowitz OECD2023 chapter in measurement ledgerM7. It does not validate interaction traces.
Self-report reference bias: Lira etal2022 https://doi.org/10.1038/s41598-022-23373-9 ; does not quantify bias in each current item.
Matching: older Pashler review plus newer qualified Clinton-Lisell/Litzinger2024 https://doi.org/10.3389/fpsyg.2024.1428732 . Neither validates8roles or policy.
Instruction/AI/sampling/development: instruction ledgerI04-I16 and measurementM5-M6 guide bounded hypotheses; externalaverage effects are not Learnality benefits. No new literature search or learner study was performed in authoring this blueprint.

LEARNALITY - PROSPECTIVE PILOT AND VALIDATION PROTOCOL
Version protocol-development-1 • October 14, 2026
Scope: participating U.S. schools, grades9-12, bounded reviewed curricular units.
Status: study-design draft, NOT a registered protocol, approved research program, completed study or invitation to collect real student records. No fixed adequate sample size, validated scales or guaranteed educational benefit is claimed. Qualified school, measurement, statistical, subject, accessibility and ethics expertise must finalize it before recruitment.
Companion specification: assessment-technical-blueprint.txt; user-requirements-audit.txt; four research-audit files.

1. QUESTIONS AND INTERPRETATIONS TO FREEZE
Q1 Do the tasks elicit the proposed bounded strategies rather than reading/access/familiarity artifacts? Are eight roles distinct enough to justify independent interpretations, or should vocabulary/scales change?
Q2 How precise and repeatable are task/rubric interpretations across tasks, occasions, raters and relevant populations? Can alternate prompts/access modes be compared for intended outcomes?
Q3 Can high-school students navigate assessment and report without routine UI instruction, with adult/access support available? Who disengages or experiences barriers?
Q4 Does a frozen evidence-guided instructional policy improve independent curriculum learning, delayed retention and transfer beyond strong ordinary/evidence-informed teaching, while preserving required content, rigor, lawful access and teacher feasibility?
Q5 Does profile information predict differential support response beyond prerequisites/context, and do those differences yield useful held-out decisions? An average intervention gain does not answer this.
Q6 What changes can annual/retake reports legitimately compare after exposure, form, context and developmental changes?
Not investigated by onepilot: universal optimal curriculum blend, brain/personality diagnosis, broad national norms, Pre-K-12 linked scale, high-stakes placement, universal teacher-free learning or guaranteedcommercial accuracy.

2. GOVERNANCE BEFORE LEARNER COLLECTION
Name legal operator, sponsor/principal investigator, scientific lead, school authorized representative, privacy/security owner, complaints/escalation path and independent reviewers; declare commercial/funding conflicts. Finalize study sites/states/populations, protocol and agreements rather than treating a student checkbox as authorization.
School/DPA terms identify purposes/data/processors/storage/security/access/rights/retention/deletion/incident responsibilities. Determine appropriate research/ethicsreview and school/parent/eligible-student permissions/assent under applicable institutional and legal requirements with qualified review. Do not claim one universal waiver or consent route. Include under-13 eligibility handling if anypotentialhigh school participant makes it relevant; do not collect exact birth date automatically without necessity.
Distinguish service participation, optional research, publicity and secondary model training. No penalties, lost curriculum access or coercivecostumes for declining research. Document accessible withdrawal and what can/cannot be removed after de-identification/aggregation per reviewed agreements.
Minimize identifying data: school-owned ID mapping separated from research key; approved educational records/accesscontext only. Sensitive demographic/access variables collected only when justified/authorized for fairness and protected; no psychiatric/religion/income questions for personality depth. Do not infer disability from actions.
No identifiable records sent to unapproved AI/mail/analytics services. The current site is not evidence that hosting, legalconsent, security or backup deletion is ready. Synthetic rehearsal and security/privacy acceptance must precede collection.
Retention periods and deletion procedures are explicit before collection, including raw responses, audio if proposed, logs, exports, backups and research-derived data. Audit access and verify deletion. Immutable lineage means append-only correction while retained, not permanent retention.

3. STAGED DESIGN AND STOP/REVISE GATES
StageA - expert construct/content review
Panel includes high school subject teachers across selected curricula, educational measurement/statistics, accessibility/language expertise and student advisers. Review blueprint claim/task/criterion/rivals, prerequisite maps, representation demands, cognitive complexity, source accuracy and report wording. Experts independently assess then reconcile documenteddisagreement. Testkeys independently computed. Create approved content/rubric exemplars and explicitly reject unsupported role inferences.
Gate: a boundeddomain/version and set of permissibledecisions are approved in writing; panel can discard current eight-factor proposal. No fixed agreement number imposed here.

StageB - cognitive interviews and authentic usability
Purposive feasibilitysampling across grades9-12, prior attainment, languages/readingdemands, disability/access/device contexts and students prone todiscontinuation, subject toapprovedprivacy. Separate task reasoning interviews from ordinary unprompted administrations: think-aloud itself changes conditions.
Study interpretation of reflection frequency/depends/missing, socially desirable wording and label/avatar deference; observe strategies and whetheroptions cue answers. Log requested UI help, access help and conceptual help separately; record interruption/fatigue/time burden and barriers. Include non-completers through approved follow-up, not only delighted completers. Test simplified/richer interface with constant tasks and feedback where feasible; learning and enjoyment as separate outcomes.
Gate: report revised wording/content, unresolved barriers, assistance/burden and administration instructions; rereview consequential changes beforelargerscoring study.

StageC - bank/rater/form development
Use a planned matrix of content families/tasks/occasions/forms and raterassignment suitable for individual decisions. Include missing/uninterpretable/artifact edge cases, prior attainment/access/language variation and fresh families; do not treat correlated components as independent items.
Train raters using exact criterion exemplars; independently/blindly double-score a prespecified selection before adjudication. Preserve all first judgments; estimate disagreements bycriterion and relevant subgroup. Use appropriate categorical/ordinal agreement and uncertainty, adjudication frequency and severity drift; not raw agreement alone. Generalizability/multilevel/ordinal measurementmodels may be appropriate if assumptions/data support; modelchoice requires specialistreview.
Investigate structure of self-report separately from behavioral criteria, local dependence, conditional precision, floor/ceiling, reference bias and missingness. Do not insist confirmatoryeightfactorsfit; cross-validate alternative/reduced structures after exploration without reusing validation dataset to select it. Test-responsebehavior and role-language effects remain alternativeexplanations.
Forms: experts confirm construct/coverage/demand, assign counterbalanced/randomforms using reviewed design, record all exposures; use common anchors where appropriate and contaminationplanned. Estimate comparability and uncertainty across eligiblepopulations; sharedcriteria/template/numberchanges alone not equating. Accessmode equivalence investigated only when sameconstruct preserved.
Gate: technical report narrows interpretation to supported populations/forms/conditions; thresholds/reporting precision/stopping rules selected from decisionrisk and measured conditionaluncertainty. No arbitrary reliabilitycutoff or "threeobservations" certification.

StageD - prospective instruction/policy comparison
Begin with a reviewed bounded unit and explicit mandatory objectives, prerequisite checks, school policy/calendar/resources/access, common criteria and independent checks. Freeze source, instruction variants, policy versions, teacher review rules and intervention dose before outcome access.
Candidate design contrasts (i) credible current practice, (ii) strong broadly evidence-informed support, (iii) constrained evidence-guided policy. Specify comparison arm relevance; do not create weak control that denies needed instruction/accommodations. All arms preserve required school curriculum, required supports and meaningful standards. Policyincludes teacherapproval/overrides; analyze this actual policy, not hypothetical perfect adherence.
School/class/unit randomization level is selected to address contamination and operational constraints. Within-student freshmatchedunit comparisons may support exploratory responsiveness, but learning is not reversible and no assumed washout erases carryover. Content/order/period balancedwhere feasible; presentation/support assignment fixed andlogged before exposure. Wholeclass discussion, teacher changes, peer spillover, time/place and AIhelp must be modeled/documented.
Use blinded common outcome raters where feasible, same approved construct/demand and access. Contrast independent first attempt learning with supported practice and assisted revisions. Score a delayed common-criterion outcome at a prespecified curricularly relevant interval/window anchored toactual teaching; retain actualdelay and logged intervening exposure. Seven to ten days is the current timing convention, not an empirically universal optimum. Define novel transfer demand/structural distance explicitly with required prerequisites; changing wording is not fartransfer.
Evaluate profile×support predictions beyond prerequisites/context and whether frozenpolicy gains survive held-out people/sites/time. Beneficial default practice offered broadly when useful; do not insist there is an individualizedwinner.
Gate: interpretable prospective policy benefit, uncertainty, coverage/rigor/access and harms evaluated against predeclared decision criteria; bounded/null/harmful findings reported. A novel profile policy may be abandoned or simplified toreadiness guidance.

StageE - annual linking and replication
Separate reflection response-shift/context changes from knowledge/strategy performance. Repeated study includes grade/subject/development, stable anchors where appropriate, new equivalent families, practice/coaching/access changes and interval history. Evaluate population/form/rubric/scoring drift and longitudinal missingness. Use linking/model assumptions and cross-validation justifiedby actual data; no numericannual growth beforecomparability.
External sites/subjects and later cohorts test frozen policy generalization; failures constrainclaims. Youngergrades require newdevelopmental blueprints, family/educator observation and language/culture/accessdesign; later linking only if empirically defensible. Do not extrapolate high school norms to preschool.

4. PREREGISTRATION AND ANALYSIS PLAN
Register after governance/finaldesign and before examining confirmatory outcomes: intended claims, eligibility/recruitment/exclusions, units/randomization/concealment, interventions/overrides, bank/form/rubric/policy versions, primary outcome/contrast/assessment window, secondary/harm/fairness/engagement outcomes, sample design, analysis-code plan and stop/revision rules. Publication ledger distinguishes preregistered confirmatory, exploratory and post hoc changes.
Primary outcome proposal tofinalize: independently rated delayed common-criterion performance for mandatory objectives, with exact aggregation/weighting and interpretability specified by curriculum/measurement experts before data. Immediatelearning/transfer/retention are not interchangeable. Do not create an unvalidated combinedpersonality-learning endpoint.
Essential secondary outcomes: commonindependent immediate/transfer performance; coverage and source accuracy; scaffold/help/exposure; unsupportedrouting/teacher override, burden/time/cost; participation/attrition/fatigue/access/barriers; student confidence/enjoyment separately; adverse content/learning/access outcomes. Prefer approved minimalcontext variables; no intrusive attention telemetry.
Estimate intention-to-treat effects for assigned policy where design supports it, adjusting only prespecified baseline/objective variables and clustering/nesting/randomization structure. Adherence/per-protocol analyses secondary with assumptions and selection bias explicit. Prespecify methods for bounded/ordinal criteria, multiplicity, effect size/uncertainty, baseline imbalance and outcome/rater dependency; don't use significance alone as shipping gate.
Interactions: few theory/blueprint-grounded prespecified support×evidence tests with shrinkage/hierarchicalmethods only if suitable; separate averagebenefit, conditionalbenefit and actualpolicy value. Avoid hunting eight roles×manysupports until something wins. Compare added value over prior attainment and universal good teaching. Review fairness/calibration/benefit uncertainty across authorized subgroups; small subgroups are uncertainty, not proof no harm/bias.
Missingness/attrition: classify absent/unsubmitted/withdrawal/accessbarrier/simulated/out-of-window separately; do not convert tozero or omit quietly. Report reasons/rates byarm/site/subgroup/time; specify followup and sensitivity/imputation appropriate toestimand and plausible missing mechanisms. Complete-case-only gains are insufficient when dropoutrelated tolearning/access. Original withdrawal rights govern handling.
Leakage: partition learner/class/school/time/content families as needed by intended generalization, not random response rows. Rater training/exemplar selection, item calibration, routing/model tuning, threshold selection and outcomeanalysis kept separate from final held-out evaluation. Preserve keys/outcome access restrictions; any tuning aftervalidation turns those data intodevelopment data and requires fresh holdout. Log prior form exposure and secondary assistance.
Carryover: distinguish teaching content, practice/quizzes, support, peer discourse and order/period. Fresh content and counterbalance reduce but do notproveabsence of carryover. Prespecify period/orderinteraction and sensitivity checks; assess plausibility of assumptions before individual effect claims. Repeated individual trials need appropriatelymodeled matched families and uncertainty; one contrast not a stable learner trait.
Harms: monitor loss of required coverage, misconception, unjustified weakness labels, delayed needed instruction, increasedburden/disengagement, inaccessible adaptation and independent performance fallingdespite assisted practice rising. Predetermine escalation, content pause, repair and participant/school notification consistentwith reviewed governance; don't invent numerical stop thresholds beforeexpert/designreview.

5. SAMPLE SIZE AND DECISION PRECISION - REQUIRED SPECIALIST TASK
No invented universal N or minimally adequate count is supplied. A qualified measurement/statistical specialist must document intended decision error costs, meaningful benefit/misrouting/harm bounds, primary contrast and interaction/policy value estimands, scoring variance/dependence, class/school intraclass correlations and cluster availability, task/occasion/raterstructure, baseline predictiveness, attrition/contamination, form/subgroup coverage and longitudinal follow-up.
Use justified external/pilot parameters with uncertainty/sensitivity or simulations where analytical approximations insufficient. Compute feasibility/power/precision for actual design and held-outpolicyvalidation, including fairness estimates; smallfeasibility/cognitive interview samples cannot establish precise individual routing. Choose recruitment/allocation only afterschool feasibility, burden and governance review. If feasible data insufficient, narrowclaims and deferdecisions instead of claiming nominal power validates them.

6. OUTPUTS AND RELEASE GATES
Required study outputs: frozen blueprint and version registry; expert decision ledger; cognitive/usabilityreport includingnon-completers; rater manual/exemplars and agreement report; bank/form/fairness/precision technical report; preregistration + deidentified analysis code where governance allows; outcomes/uncertainty/harms/overrides/attrition/carryover/limitations; replication/annual comparability reports; independent technical review.
Commercial claim ledger maps each public statement to population/version/study and allowed consequence. Educational validity, secure operations, legal authorization and independent usability remain separate signoffs. Software passing tests do not satisfy any research gate. Model, content, interface or scoringchanges material todecision meaning trigger re-review/revalidation.
Current status: research-informed only. No schools recruited, students studied, ethicsapproval/registration, actual sample calculation or independent panel report exists through thisprotocoldraft. Private engineering may continue on immutablelineage/reviewed content/evidence/QC, but claims beyond this boundary must remainrestricted.

7. REFERENCES / BASIS
This protocol is a proposed design informed by local research ledgers, not a ready-made method endorsed by those sources. AERA/APA/NCME2014 Testing Standards establishes responsibilities for intendeduses, profiles, precision, fairness/form/growth; https://www.testingstandards.net/open-access-files.html . IES guides and directinstruction/adaptive/AItrials in instructionledgerI04-I16 inform strong comparators and separate assisted/independentoutcomes, not Learnality efficacy. Lira2022/retake evidence in measurementM2/M5 motivates response-shift/exposure safeguards; NAEYCM6 developmentalextension. Matchingcontraryfindings (instructionI03) mustremain in technicaldiscussion. EngagementE1-E6 supports testing ratherthan assumingavatarbenefits. Fulltextgaps/populationlimits are retained in originalledgers. No new literature search or external contact occurred for thisdraft.