Verification and pilot materials · Complete web edition
Phase Ten Protected Human Performance Protocol
This is the complete web edition of the canonical project document. Its narrative, recommendations, evidence tables and source links are presented below without abridgment.
Project, Research, Analysis, Editorial by Patrick Russell; assisted by ChatGPT (OpenAI).
California Education and AI Protected Human-Performance Protocol
Phase Ten | Testing what learners can do after AI assistance ends
Protocol date: September 12, 2026 • Informational research design • No outreach or personal attribution
Protocol field | Entry |
|---|---|
Responsible person | ________________________________________ |
Institution or program | ________________________________________ |
Course, grade, or activity | ________________________________________ |
Review period | ________________________________________ |
Phase Ten conclusion
The central educational question is not whether artificial intelligence can improve a product while it is available. It is whether the learner can later retrieve knowledge, explain reasoning, verify truth, create, judge, and act responsibly without the tool. This protocol therefore separates assisted practice from immediate unaided performance, delayed retention, and transfer to a new problem. It also prevents an institution from receiving evidence credit merely for publishing a policy, licensing a tool, offering a course, or producing impressive group work.
The protocol samples capacities associated with a well-formed mind; it does not reduce the whole person to one score. Knowledge, judgment, intellectual honesty, self-command, moral responsibility, attention, and human relationship must be interpreted together, with limits stated. No wellness or integrity consequence may be automated from the protocol.
Required decision before testing
Question | Required answer |
|---|---|
What is being formed? | Name the exact capacity: recall, explanation, writing, source judgment, quantitative reasoning, coding, creativity, moral judgment, attention, or another defined ability. |
What may AI do? | State the tool, account, mode, prompts, permitted stages, disclosure rule, data restrictions, accessibility, and non-AI alternative. |
What must remain human? | Identify the first attempt, reading, retrieval, analysis, composition, oral defense, practical work, or judgment that assistance may not replace. |
What evidence changes the score? | Predefine the outcome, scoring rule, comparison, delay, subgroup analysis, and threshold for confirmation or revision. |
Five-condition performance sequence
Condition | Timing | Administration | Interpretive purpose |
|---|---|---|---|
C0 — Unaided baseline | Before the intervention | Same-domain task without AI; ordinary approved accommodations remain. | Starting knowledge and skill; baseline equivalence. |
C1 — Bounded practice | One to four learning sessions | AI condition is recorded exactly: hints, feedback, critique, source-bounded help, or open generation. | Assisted performance, process, tool reliance, and human support. |
C2 — Immediate unaided | Within 24 hours | Parallel task with AI and AI-enabled search removed. | Whether the learner can reproduce the core mental act. |
C3 — Delayed retention | Normally 2–4 weeks | Equivalent task with no rehearsal immediately beforehand. | What persists after short-term performance effects fade. |
C4 — Novel transfer | Normally 4–8 weeks | New text, problem, dataset, scenario, or audience requiring the same underlying concept. | Whether learning generalizes beyond the practiced form. |
AI-use conditions
Code | Condition | Meaning |
|---|---|---|
H0 | Human-only | No student-facing AI; ordinary non-AI resources and approved accommodations are documented. |
H1 | Teacher-facing AI | AI may assist planning or feedback preparation, but the student does not interact with it. |
S1 | Bounded student use | A defined tool and function supplies hints, critique, feedback, comparison, or source-bounded assistance without replacing the target act. |
S2 | Open student use | Broad generation or agentic help is permitted under disclosure and verification rules. |
Administration rules
- Use parallel—not identical—tasks across C0, C2, C3, and C4 so memory of an answer is not mistaken for transfer.
- Hold content difficulty, time, scoring, language, and ordinary accommodations as constant as practicable.
- Do not withdraw disability-related assistive technology merely because AI is removed; distinguish accommodation from generation of the target work.
- Record teacher, tutor, parent, peer, and AI assistance separately. Human mediation is an intervention component, not background noise.
- Do not tell students that a protected check is a trap for detecting misconduct. Its purpose is to measure learning, not infer guilt from a performance gap.
Developmental administration
Band | Protected format | Immediate | Retention | Transfer | Boundary |
|---|---|---|---|---|---|
K–2 | Oral, manipulatives, drawing, reading aloud, brief handwriting; 5–15 minutes. | Minutes to 2 days | 3–7 days | Concrete variation | Parent permission and child assent when research applies; no open chatbot condition. |
Grades 3–5 | Recall, oral narration, handwriting, arithmetic, source recognition; 10–25 minutes. | Within 1 day | 1–2 weeks | New example or short text | Adult-supervised S1 only; preserve foundational reading, writing, and calculation. |
Grades 6–8 | Explanation, structured writing, logic, multi-step mathematics, source comparison; 20–40 minutes. | Within 1 day | 2–3 weeks | New problem, text, or claim | Age, consent, privacy, parent notice, and non-AI alternative are threshold conditions. |
Grades 9–12 | Extended writing, oral defense, research verification, quantitative and creative work; 30–60 minutes. | Within 24 hours | 2–4 weeks | New domain example or audience | Increase student agency while retaining disclosure, privacy, and protected assessments. |
Higher education | Discipline-authentic written, oral, laboratory, clinical, mathematical, coding, or design task. | Within 24 hours | 2–4 weeks | New case, dataset, source set, or professional scenario | Course purpose and faculty authority determine the protected act; accommodations remain. |
K-through-12 threshold protections
- Instructional use and research participation are different. Ordinary classroom assessment may not require research consent, but systematic investigation may require institutional review, parental permission, and affirmative child assent.
- A child's silence or failure to object is not affirmative assent. Dissent must be respected when assent is required.
- Use the least data necessary. Separate identity from analytic records, limit retention, prohibit secondary model training unless specifically authorized, and prevent vendors from using results for advertising or profiling.
- No child should lose instruction, opportunity, or accommodation for declining an optional AI or research condition; provide a genuine non-AI alternative.
- Wellness questions must not become covert diagnosis, discipline, or emotional surveillance. Concerning disclosures require a human safeguarding process defined in advance.
Method sources HHS Subpart D | HHS children research FAQ | Education technology privacy
Core domain 1 Memory and foundational knowledge
Element | Protocol |
|---|---|
Construct | Accurate retrieval of knowledge required for later reasoning rather than mere recognition or ability to relocate information. |
C0 | Short free recall, cued recall, oral narration, or calculation from previously taught material. |
C1 | Record whether AI retrieves facts, organizes them, prompts recall, or merely checks a student-produced answer. |
C2/C3 | Use parallel questions with no AI-enabled search. Score correct ideas, important omissions, misconceptions, and confidence calibration. |
C4 | Require use of retained knowledge inside a new explanation, argument, calculation, or practical problem. |
Failure signal | High assisted accuracy followed by weak unaided recall or inability to use the knowledge in a new task. |
Core domain 2 Reading and source judgment
Element | Protocol |
|---|---|
Construct | Close reading, claim identification, source evaluation, corroboration, context, and recognition of uncertainty or fabrication. |
Protected task | Provide a short primary text and a conflicting or misleading source. Require annotation, claim-evidence mapping, credibility judgment, and citation verification without AI. |
Transfer | Use a new topic and source type so familiarity with content cannot substitute for the evaluation method. |
Scoring | Accuracy of claims, quality of evidence, source provenance, counterevidence, uncertainty, and correction of false or unsupported statements. |
Core domain 3 Writing and authorship
Element | Protocol |
|---|---|
Construct | Idea formation, organization, evidence, sentence-level control, revision judgment, voice, and truthful disclosure of assistance. |
Protected task | Use an in-class or supervised paragraph, memo, proof explanation, or essay segment tied to the same concept as the assisted work. |
Authorship check | Ask the learner to explain the thesis, reconstruct one decision, identify what changed after feedback, and revise a new passage without AI. |
Scoring | Reasoning and evidence are scored separately from fluency. A polished assisted product cannot offset inability to explain or reproduce its core argument. |
Method sources Retrieval-practice study | Google-effects study
Core domain 4 Oral explanation and dialogue
Element | Protocol |
|---|---|
Construct | Coherent explanation, responsive reasoning, listening, objection handling, self-correction, and ownership of conclusions. |
Protected task | Three to eight minutes of questioning: explain the central idea, answer an objection, apply it to a new example, and identify one limit or uncertainty. |
Safeguard | Use a short standardized prompt set and allow appropriate communication accommodations; do not equate verbal speed, accent, or extroversion with understanding. |
Evidence | Record rubric scores and brief notes, not facial-emotion inference or automated personality judgments. |
Core domain 5 Mathematical and scientific reasoning
Element | Protocol |
|---|---|
Construct | Representation, method selection, multi-step reasoning, estimation, units, evidence, experimental logic, and error detection. |
Protected task | Solve or design a parallel problem without AI, showing intermediate reasoning. In science, interpret new data or defend a laboratory decision. |
Transfer | Change surface features while preserving the underlying concept; include one plausible but flawed solution for diagnosis. |
Scoring | Separate final answer, conceptual model, procedure, error recognition, and explanation. Partial reasoning remains visible. |
Core domain 6 Coding and computational work
Element | Protocol |
|---|---|
Construct | Problem decomposition, data structures, algorithm choice, debugging, testing, security, and explanation of generated or reused code. |
Protected task | Write or repair a small program without generative assistance, explain each important choice, predict behavior, and add a test for an unseen case. |
Transfer | Alter requirements or data so copied code cannot pass without conceptual adaptation. |
Scoring | Correctness, reasoning, tests, debugging process, security, efficiency appropriate to level, and accurate disclosure. |
Core domain 7 Creative production
Element | Protocol |
|---|---|
Construct | Observation, imagination, technical control, aesthetic judgment, revision, intention, and ability to explain influences and choices. |
Protected task | Produce a bounded sketch, passage, composition, design variation, or performance from a new prompt without generation, then explain and revise one decision. |
Interpretation | Do not reward novelty alone. Evaluate intentionality, craft, coherence, responsiveness to constraints, and truthful provenance. |
Core domain 8 Moral, civic, and professional judgment
Element | Protocol |
|---|---|
Construct | Recognition of persons, duties, consequences, rights, goods, uncertainty, conflicts, and responsibility for action. |
Protected task | Present a new dilemma. Require a decision, reasons, strongest objection, affected parties, uncertainty, and conditions that would change the judgment. |
Scoring | Assess factual accuracy, principled reasoning, recognition of tradeoffs, consideration of persons, intellectual humility, and responsibility—not agreement with a preferred political conclusion. |
Limit | A written scenario samples judgment; it does not prove character or future conduct. Avoid psychological labeling. |
Core domain 9 Attention, self-regulation, and human help-seeking
Element | Protocol |
|---|---|
Construct | Sustained attention, tolerance of productive difficulty, planning, monitoring, correction, completion, and appropriate recourse to teachers, parents, peers, or tutors. |
Behavioral evidence | Time on task, voluntary pauses, abandoned attempts, correction after feedback, help source, and completion under equivalent conditions. |
Self-report | Use brief, neutral questions about effort, confidence, frustration, perceived learning, belonging, and preferred help. Self-report supplements behavior; it does not replace it. |
Safeguard | Do not use webcam affect detection, covert monitoring, keystroke surveillance beyond what is essential, or automated mental-health inference. |
Human relationship indicators
- Frequency and quality of teacher, tutor, parent, and peer interaction before and after AI use.
- Whether AI prompts a student to seek human clarification or becomes a substitute for asking a person.
- Belonging, friendship, family support, and trusted-adult access, interpreted as contextual conditions rather than automatic causes.
- For home education, distinguish healthy parent mediation from excessive burden and distinguish chosen home learning from crisis schooling.
Common scoring rubric
Score | Performance anchor |
|---|---|
0 — Not demonstrated | No relevant response, wholly unsupported response, or inability to begin without prohibited assistance. |
1 — Fragile | Isolated correct elements with major misconceptions, missing reasoning, or inability to apply the idea. |
2 — Developing | Core idea is partly correct; reasoning is incomplete but recoverable with limited human prompting. |
3 — Independent | Accurate, coherent, appropriately supported performance without AI; errors are minor and do not undermine the central result. |
4 — Transfer | Independent performance adapts the concept to a genuinely new case, addresses limits or objections, and self-corrects when warranted. |
Dimensions scored separately
Dimension | Evidence |
|---|---|
Knowledge | Accurate retrieval and use of required facts, concepts, texts, procedures, or principles. |
Reasoning | Visible chain from evidence or premise to conclusion; appropriate method and counterargument. |
Verification | Source provenance, corroboration, uncertainty, error detection, and correction. |
Authorship | Ability to explain, reproduce, and revise personally owned intellectual decisions. |
Transfer | Application to a new problem, text, dataset, scenario, or audience. |
Self-regulation | Planning, persistence, monitoring, productive help-seeking, and correction. |
Communication | Clarity appropriate to the discipline and age, scored without penalizing protected disability or language accommodations. |
Decision rule
Report each dimension and condition separately. Do not collapse moral judgment, wellness, knowledge, and fluency into a single score. A claim of protected independent learning requires at least PH-2 design evidence and acceptable C2 performance. A claim of durable learning requires acceptable C3 performance. A claim of transfer requires acceptable C4 performance and a credible comparison. The protocol does not prescribe one universal numerical cut score; institutions must set and justify age- and discipline-appropriate thresholds before examining results.
Comparison and validity design
Design element | Minimum standard |
|---|---|
Preferred comparison | Random assignment to H0, H1, or S1 when educationally and ethically feasible; otherwise use a planned crossover or matched comparison. |
Baseline equivalence | Compare C0 performance, prior achievement, age or grade, language, disability, and relevant course preparation. |
Common treatment | Use the same learning objective, time, instructor access, task difficulty, and scoring; document unavoidable differences. |
Fidelity | Record the actual AI tool, version, settings, prompts, outputs, duration, human assistance, and deviations from the assigned condition. |
Attrition | Report who did not complete C2, C3, or C4 and whether missingness differs by condition or student group. |
Blinding | When practicable, remove names and condition labels from scored work; use at least two raters for subjective high-value samples. |
Subgroups | Predefine analyses by prior performance, disability, language, age, family support, and existing distress only when sample size and privacy permit. |
Causal language | Use causal claims only when design and implementation warrant them. Otherwise describe associations and plausible alternatives. |
Time and transfer windows
Claim | Minimum evidence |
|---|---|
Immediate mastery | C2 within 24 hours under H0 conditions. |
Retention | C3 after 2–4 weeks for most learners; shorter developmentally appropriate window may be justified for young children. |
Near transfer | C4 uses new surface features in the same domain. |
Far transfer | C4 uses the underlying principle in a different context; claim only when the new context is specified and scored. |
Durability | More than one delayed point is preferable; a single end-of-course product is not durable-learning evidence. |
Method sources WWC standards and protocols | Stanford SCALE 2026 review
Wellness and COVID context module
Wellness is neither a proxy for academic performance nor a diagnosis. The module asks whether AI changes the conditions under which learning occurs and whether effects differ for students who entered COVID disruption with stable habits and support versus those already experiencing family dysregulation, isolation, weak friendships, disability-related barriers, or distress.
Context field | Minimum record |
|---|---|
Pre-COVID trajectory | Comparable academic, attendance, support, or wellness indicators before 2020 when available; do not reconstruct from memory alone. |
Disruption exposure | School closure, emergency remote instruction, family illness or loss, device and internet access, housing or financial disruption. |
Prior condition | Learning habits, attention, friendships, family stability, disability support, and distress before or at the start of disruption. |
Recovery conditions | Return to in-person learning, tutoring, counseling, family capacity, peer reconnection, attendance, and continuity of instruction. |
AI-era exposure | Tool, frequency, duration, purpose, age of first use, teacher or parent mediation, and whether use replaced or supplemented human support. |
Current indicators | Attention, sleep, workload, frustration tolerance, belonging, loneliness, friendship, trusted-adult access, and help-seeking. |
Interpretation boundaries
- Do not attribute a current difficulty to COVID merely because the student experienced school closure.
- Do not treat stable homeschooling before 2020 as equivalent to emergency crisis schooling.
- Do not infer that AI caused improved or worsened wellness from simultaneous adoption; examine timing and alternatives.
- Do not label engagement, convenience, satisfaction, or product completion as mental fitness.
- Report when samples are too small to support subgroup conclusions and suppress identifiable combinations.
Stop conditions
- Pause an activity if it creates more than minimal unanticipated risk, exposes sensitive student information, or conflicts with required accommodations.
- Escalate credible self-harm, abuse, or immediate-safety disclosures through trained human procedures; never delegate the judgment to an AI system.
- Do not use performance discrepancies, detector scores, or writing-style changes as sole evidence of misconduct.
Application to the ten California institutions
Institution | Practice unit | Protected check | Delayed or transfer task | Priority control |
|---|---|---|---|---|
UC Berkeley | Berkeley Law writing or doctrinal task | Closed-resource issue spotting plus oral defense | New fact pattern after 2–4 weeks | Authorized exceptions, citation accuracy, faculty variation |
UCLA | AI Literacy Module plus Faculty Fellows course | Supervised correction and concept explanation | New misleading output or source set | Module completion, course design, behavior beyond awareness |
Stanford | COLLEGE 102 policy memo | Individual understanding check | Apply policy principles to a new assessment | Proposal adoption, individual versus group learning |
USC | ChatGPT Edu course workflow | Reproduce reasoning and correct one flawed output | New prompt or professional scenario | Tool-only versus AI-plus-faculty support |
Caltech | Ethics-and-writing or technical course | In-class analysis, oral conference, or proctored problem | New scientific, quantitative, or ethical case | Cross-division transfer, trust, detector practice |
Loyola Marymount | AI-free and AI-integrated paired assignments | AI-free anchor plus supervised revision | New source or ethical problem | Reasonable alternatives, co-constructed rules, protected tools |
Thomas Aquinas College | Composition and senior thesis | Existing seminar, demonstration, composition, and defense | New text, objection, or proof after delay | Rubrics, exceptions, accessibility, later responsible use |
Mother of Divine Grace | Grades 6–9 logic and composition | Narration, supervised writing, and parent or teacher dialogue | New argument or text before any AI comparison | Parent mediation, age rules, family burden, friendships |
Harvard-Westlake | AI ethics, technology, and society course | Human-authored essay segment and oral objection | New source-evaluation and ethical scenario | Course rule, division rules, consent, privacy, grade differences |
High Tech High | Exhibition-based project plus elementary comparison | Individual process conference or construction task | New design problem without AI | Full policy, individual mastery, group authorship, school variation |
Minimum evidence package for each case
- Current rule, authority, effective date, approved tool, privacy review, disclosure standard, accessibility, and non-AI alternative.
- Learning objective, prompt, readings or data, AI condition, human support, drafts, outputs, final work, rubric, and evaluator record.
- C0 through C4 results by separate dimension, with de-identified condition and subgroup information where justified.
- Implementation deviations, attrition, student and educator experience, wellness context, and stated limitations.
- Scorecard update identifying which Phase Nine claims move from missing or restricted to partial, verified, PH-3, or PH-4.
Field recording form
Record | Entry |
|---|---|
Institution / course / grade | ____________________________________________________________ |
Capacity being formed | ____________________________________________________________ |
AI condition and exact tool | ____________________________________________________________ |
Protected human act | ____________________________________________________________ |
C0 baseline task and score | ____________________________________________________________ |
C1 practice and assistance | ____________________________________________________________ |
C2 immediate task and score | ____________________________________________________________ |
C3 retention task, date, score | ____________________________________________________________ |
C4 transfer task, date, score | ____________________________________________________________ |
Human support and accommodations | ____________________________________________________________ |
Context and wellness notes | ____________________________________________________________ |
Deviations and missing evidence | ____________________________________________________________ |
Dimension record
Dimension | C0 | C2 | C3 | C4 | Evidence note |
|---|---|---|---|---|---|
Knowledge | __ | __ | __ | __ | ____________________________ |
Reasoning | __ | __ | __ | __ | ____________________________ |
Verification | __ | __ | __ | __ | ____________________________ |
Authorship | __ | __ | __ | __ | ____________________________ |
Transfer | — | — | — | __ | ____________________________ |
Self-regulation | __ | __ | __ | __ | ____________________________ |
Communication | __ | __ | __ | __ | ____________________________ |
Evidence judgment
Question | Judgment |
|---|---|
Did assisted performance improve? | Yes / No / Mixed / Insufficient |
Did immediate unaided performance improve? | Yes / No / Mixed / Insufficient |
Was learning retained? | Yes / No / Mixed / Insufficient |
Did learning transfer? | Yes / No / Mixed / Insufficient |
Were authorship and verification preserved? | Yes / No / Mixed / Insufficient |
Did human relationships or wellness change? | Improved / Worsened / Mixed / No detected change / Insufficient |
Phase Nine status update | V / P / M / R / S; PH-0 / PH-1 / PH-2 / PH-3 / PH-4 |
Evidence governance and reporting
Requirement | Rule |
|---|---|
Protocol registration | Record outcomes, time windows, comparisons, exclusions, and subgroup plans before reviewing results. |
Data minimization | Collect only what the educational question requires; separate identifiers and limit retention. |
Human review | Educators or trained reviewers examine scored work and any integrity or wellness concern. |
Accessibility | Provide ordinary accommodations and explain how they differ from AI generation of the target act. |
Transparency | Tell learners what is measured, what AI is recorded, who can see it, and what consequences are possible. |
Version control | Preserve the policy, tool version, model, settings, prompt, and date; AI systems can change during a study. |
Negative and null results | Report no effect and adverse effects, not only successful pilots. |
Limits | State design weakness, attrition, sample size, implementation variation, and unmeasured alternatives. |
What the protocol may support
- A limited judgment about a defined learning activity, student group, AI condition, and time window.
- A comparison of AI alone, AI plus a human educator, teacher-facing AI, and protected human-only work.
- A verified Phase Nine status change and a decision to expand, narrow, redesign, pause, or prohibit the defined use.
What the protocol may not support
- A claim that one test fully measures a well-formed mind, character, faith, mental health, or human worth.
- A misconduct finding based solely on a performance gap, detector, style change, or AI-generated probability.
- A campus-wide, schoolwide, or causal claim from one case or while baseline, implementation, attrition, or alternatives remain unresolved.
Phase Eleven handoff
Phase Eleven should turn this protocol into a prospective pilot scorecard for Berkeley Law, Stanford, USC, Thomas Aquinas College, and High Tech High, while retaining Mother of Divine Grace as a separate K-through-12 benchmark. It must separate executable design from unavailable results and identify the evidence needed for PH-3 or PH-4.
Method sources WWC | Student privacy | HHS child protections | SCALE