Current study page: Learning as a Measurable System

The earlier reference below is retained for continuity.

Learning evidence lab¶

Canonical target: ML01, with ML04, ML07, ML15 and ML16.

Distinguish performance while support is available from retained capability after that support is removed. A correct assisted answer establishes neither independent recall nor transfer. Conversely, slow practice performance is not automatically evidence that learning failed.

Reading and reconstruction¶

Read the introduction and the knowledge/reasoning and assessment discussions in How People Learn II, especially its treatment of context, prior knowledge and evidence. This is source SRC-ML-HPL2, a consensus synthesis, not evidence that this particular lab improves outcomes. Reconstruct its argument in a diagram linking learner, task, environment and assessment. Mark which links are empirical findings and which are proposed design choices. Do not treat a brain image as a direct instructional prescription.

Investigation¶

Choose a benign unfamiliar capability: interpreting residual plots, identifying clauses, or tying a practical knot under supervision. Define success before practicing. Specify accessible alternatives and prior experience.
Make three matched task sets: baseline, practice and delayed test. Keep the delayed set unseen. Vary surface details without silently changing difficulty.
Collect an unaided baseline, including confidence and error types. A blank response and a confident misconception require different interventions.
Practice with feedback. Record duration, attempts, feedback and AI assistance. Separate worked-example reconstruction from independent retrieval.
Test on an unpracticed example, then again after a declared delay. Explain which differences might reflect task difficulty, fatigue, repeated exposure or scoring.

The experiment is a personal learning audit, not a causal trial proving that one study method is superior. A comparison of two methods needs credible assignment, comparable material, counterbalancing and adequate observations. Do not recruit participants or collect sensitive data without appropriate review and consent.

Failure and transfer¶

An AI tutor produces fluent hints that reveal every intermediate step. Immediate accuracy rises while unaided delayed performance does not. Diagnose what was measured, redesign assistance, and predict which result would falsify the diagnosis.

Then transfer the protocol from a verbal task to a physical skill. Identify what cannot be assessed with a written explanation and where a qualified coach is needed.

Human and AI evidence¶

Internalize: task criteria, uncertainty, and the distinction between assisted performance and retained capability.
Reconstruct/evaluate: baseline design, alternative explanations and scoring.
Retrieve/collaborate: source lookup, feedback generation and schedule support.
Delegate with accountability: administrative reminders only after permissions and privacy checks. Never delegate the unaided test or claim ownership of AI work.

Use A0 for baseline and reconstruction, A1 for bounded questions, A2 for analysis, A4 for adversarial explanations and A5 for defense without the model. A3 is optional administration, not proof of learning.

Defense and moderation¶

Submit the preregistered criteria, task provenance, attempt log, delayed results and a two-page interpretation. Defend one rival explanation and one design revision. An adequate defense separates observation from causality. A strong defense proposes a discriminating next test and states what would reverse its conclusion. Polished charts, self-reported confidence and more study hours do not earn mastery credit.

Use the examiner calibration protocol. No efficacy or longitudinal claim follows from completing this exercise.