Technological Design and Durable Independent Learning

Strengths

• Pages 1, 7, 10 and 11 consistently disclose that there are no original participant data, completed trial, validated bank, registration or ethical approval. The abstract matches that status.
• H1 compares constrained adaptive practice with an active digital comparator. Print and early-solution arms address distinct questions without pretending to isolate a universal technology effect.
• Protected core blocks prevent review queues from displacing scheduled syllabus content. The paper correctly distinguishes scheduled exposure from completed exposure and effort.
• Cycle counts are explicitly descriptive and endogenous to errors and treatment. Neither a high count nor a cycle-outcome association is presented as efficiency or causal mediation.
• The planning arithmetic checks: approximately 174.42 complete outcomes per arm for d=0.30, a 0.2095-SD null-centered interval half-width at n=175, and item-proportion SE=0.0373 at n=180. Both chess examples passed independent legal-move checks.
• The recent-study summaries preserve important limits: the [Gallegos abstract](https://docs.iza.org/dp18513.pdf) supports the distinct intervention and exam claims; [Liu full-text passages](https://docs.iza.org/dp18989.pdf) support the exact-match sample restriction and nonsignificant overall academic-harm result.
• The [Contractor and Reyes version record and abstract](https://arxiv.org/abs/2607.08849v2) and [Southwick publisher abstract](https://www.psychologicalscience.org/journals/psychological-science/09567976261452568/) support the targeted descriptions, with observational and access limitations retained.

Weaknesses

• The narrative search is candidly purposive but has no database-by-query retrieval log or explicit appraisal matrix. Add these if readers are meant to assess selection balance; otherwise retain the present limited claim and avoid treating the synthesis as representative evidence.
• B4 says mapping and intended session are recorded when feedback qualifies, but for correct C responses the designated opportunity is the first related item actually selected later. Specify a pending designation state and deterministic later binding; otherwise two implementations can create different due-opportunity denominators.
• B1 describes approved mapping edges as same-level or harder, whereas B3 permits a one-level-lower review fallback using those edges. Explicitly permit and tag lower-level edges or remove that unreachable fallback. This affects implementation clarity, not the proposed H1 contrast.
• C5 should commit to the exact small-sample degrees-of-freedom combination after imputation, the coding of cohort effects in imputation, and the planned number and size of supervision cohorts. Recalculate power and demonstrate interval behavior with the final design before recruitment, as already required.
• The primary rubric deliberately permits explanation credit alongside a wrong decision. This is coherent as a two-component construct, but readers should see decision and explanation effects separately; a total-score benefit alone would not show improved move selection.
• External checking in this review was targeted. The Frontiers contextual reference could not be independently retrieved through the available web tool, and the other references were not all re-audited. Retrieval failure is not evidence that the citation is false.

Potential errors

• B4 needs an explicit collision example: a selected item can match an older correct-response source and a queued error-source review. State whether source ownership or oldest-source priority governs, and how the displaced designation is logged without double-counting.
• C5 allows a necessary repeat assessment and selects the visit closest to day 28. Specify whether repeat testing uses the same form and how repeat exposure is recorded; otherwise repetition could introduce practice effects despite the nominally novel-position endpoint.

Strong novelty candidates

• A protected common syllabus combined with a bounded, auditable adaptive review queue and delayed unaided testing is a potentially useful protocol configuration. Its priority is unestablished, and it is not presently an empirical discovery.

Unresolved material objections

None reported.

Strengths

• Pages 1 and 10 consistently distinguish a proposed study from original empirical evidence and explicitly permit findings adverse to the motivating hypothesis.
• Sections 3.3-3.5 distinguish randomized access from chosen AI use, the Liu exact-match sample from the full sample, and observational chess associations from causal effects.
• The B-versus-C contrast shares instrumentation and protects scheduled core content while acknowledging that completed exposure and difficulty can differ.
• Appendix C separates feasibility and assessment calibration, requires fresh calibration after form changes, and preserves pre-adjudication rater agreement.
• The primary delayed window is anchored to scheduled training completion rather than attendance; the protocol states an intention-to-treat analysis and explicit arm-specific missingness sensitivity grids.
• Recomputation gives approximately 174.42 complete outcomes per arm, a 0.2095-SD confidence half-width and a 0.03727 item-success standard error, consistent with the rounded text.

Weaknesses

• The selective synthesis supports motivation but establishes neither priority for the combined framework nor coverage of the adaptive-learning literature. Add a focused comparison with closely related chess or spaced adaptive practice protocols before advancing any stronger originality claim.
• Appendix B3 does not fully state when the fresh-item motif cursor advances after empty cells, fallback selections or review-only blocks. Publish explicit state transitions and deterministic replay cases in the promised frozen implementation.
• Appendix B4 says opportunity mapping and intended session are recorded when feedback qualifies, but for a correct C response defines the opportunity by the first related item actually selected later. Specify a pending state and rules for preemption by older sources so those records can be reproduced.
• The assessment-calibration schedule estimates form and visit effects at seven-day intervals without training. It does not establish measurement invariance after six weeks of instruction or by treatment; retain separate component and item-level diagnostics when interpreting the composite.
• The 880-person figure remains provisional. The number, sizes and arm composition of supervision cohorts and resulting power revision must appear in the eventual registered protocol before recruitment.
• The final analysis specification should name the chained-equation implementation, cohort handling and finite-sample pooling degrees of freedom when combining CR2 covariance estimates under multiple imputation; invoking Rubin's rules alone leaves these choices open.

Potential errors

• Appendix B's future-selected reapplication rule could yield inconsistent opportunity counts across implementations unless the pending and collision states are fixed. This affects descriptive H4 measurement rather than identification of randomized H1.
• A small number of stable supervision cohorts could leave CR2 inference imprecise despite 880 participants. The manuscript explicitly requires cohort-aware replanning, so the current independent-sample calculation must not be reused as final power.
• The first chess example has exactly one mating move, g6g7; the second is a legal checking fork and every legal Black reply permits f6h7. These local checks do not validate the absent assessment bank.

Strong novelty candidates

• A protected-core adaptation package that confines reviews and difficulty selection to supplementary time is a concrete candidate contribution, conditional on comparison with prior instructional designs.
• The one-to-one recorded-cycle scheme with explicit opportunity delivery and administrative censoring is a potentially useful measurement specification; its validity and incremental value remain untested.

Unresolved material objections

None reported.