Changes

New evidence: liu-2026 (PRELIMINARY working paper; instructor-randomized trial, 2,379 undergraduates at one US public university) - a course-grade harm signal, exact-match sample only (final grade -0.37 SD among sections of the same course; the full-sample primary estimate was not significant at p < 0.05). Grades were instructor-posted with unknown AI-access conditions and control students kept other GenAI tools, so it cannot test any unassisted-learning claim; linked consistent_but_does_not_test to a2 (university, unrestricted chat). No status or certainty changes.

"Learning gains from AI-assisted practice persist on delayed (one week or longer) retention tests taken without AI." moved from INSUFFICIENT_EVIDENCE to PRELIMINARY_SIGNAL (improving). The claim now has its first substantial gate-passing delayed evidence: a MODERATE multicentre trial with precise tool-free advantages at 7 days and ~16 weeks for one guardrailed, verification-heavy configuration. Against it stands kalam-2025's severely underpowered one-week null from an unguarded single-episode design — heterogeneity with a candidate explanation in recorded moderators (instructional design and exposure), not established contradiction. PRELIMINARY_SIGNAL rather than anything stronger because the supporting evidence is one non-randomized trial, the effect attaches to a bundled configuration, the instrument is author-developed without construct validation, and the subject sits outside the concentration areas.

A Bayesian meta-analysis in Educational Psychology Review (35 studies, 131 effect sizes) reports credible small-to-medium positive pooled effects in all four outcome categories it defines — including independent programming outcomes measured without the AI — with no credible moderators and the authors' own caution that the evidence base is too small and uneven to detect them. The article body is paywalled, so this record is assessed from the abstract, labeled, and not rated; its pool includes two studies already in this corpus.

Across six universities, 28 intact engineering classes assigned to a tightly bounded GPT-4o workflow (help only after an independent attempt, four-turn cap, mandatory verification) outscored matched active controls on supervised exams taken without AI, both one week and ~16 weeks after the module. Allocation was a deterministic balancing algorithm, not random, so this is a strong adjusted association for the whole configuration, not a causal effect of AI alone.

A six-week quasi-experiment in six classes at an all-girls Taiwanese senior high school (n=153) found classes using unrestricted ChatGPT during C++ practice scored lower than conventional-instruction classes on achievement, flow, and self-efficacy, adjusted for pretests. No claim status changed: the paper does not report whether AI was available during the posttest, so under this site's gating rule the achievement finding cannot test the claim it is linked to, and with only six classrooms analyzed at student level its precision is overstated.

"Secondary students with low prior knowledge gain less, or are harmed more, by AI assistance during practice than high-prior-knowledge peers." received its first assessed status: PRELIMINARY_SIGNAL (unclear). Two independent weak analyses point the claim's way (retention gains confined to high-prior learners; a positive treatment-by-baseline interaction), while the strongest trial's preregistered heterogeneity analysis — gated for stance because its assessment conditions are not reported — found no detectable ability moderation. An early, internally inconsistent signal.

"Learning gains from AI-assisted practice persist on delayed (one week or longer) retention tests taken without AI." received its first assessed status: INSUFFICIENT_EVIDENCE (stable). One underpowered null (N=33, ~43% power). The claim is effectively untested at meaningful precision.

"Learning gains from AI-assisted practice persist on delayed (one week or longer) retention tests taken without AI." received its first assessed status: INSUFFICIENT_EVIDENCE (stable). One study measured delayed unassisted retention: three concordant nulls with favorable trends at N=69, plus an unprespecified positive high-prior subgroup. Neither persistence nor decay is demonstrated.

"AI feedback on writing produces unassisted writing growth noninferior to human feedback — no worse than a stated margin." received its first assessed status: INSUFFICIENT_EVIDENCE (stable). No noninferiority evidence exists: the one head-to-head comparison is small, non-randomized (Serious ROBINS-I), and directionally favors the human tutor without significance. A nonsignificant difference does not establish the claim; as written it is untested.

"University students with low prior knowledge gain less, or are harmed more, by AI assistance during practice than high-prior-knowledge peers." received its first assessed status: PROXY_ONLY (unclear). The only relevant evidence is qualitative and process-tracing (struggling CS1 students' difficulties compounding, with an illusion of competence): consistent with differential harm, but entirely proxy measures with no unassisted outcome and no comparison arm.

"Guardrails prevent the unassisted-performance harm observed when secondary students practice with answer-giving chatbots." received its first assessed status: PRELIMINARY_SIGNAL (stable). One strong three-arm trial shows the vanilla-chat harm eliminated (not reversed) by teacher-designed guardrails — the claim's exact contrast, within a single study, site, subject, and product. Pass B read this as PROBABLE on the strength of the internal contrast; the single-study rule holds it at PRELIMINARY_SIGNAL until replicated.

"Guardrailed AI tutors (hints, no direct answers, pedagogical prompting) improve secondary students' performance on assessments taken without AI, compared with business-as-usual instruction." received its first assessed status: CONTRADICTORY (unclear). A well-powered trial shows no unassisted gain when guardrailed tutoring replaces equivalent practice time; a preliminary supervised after-school program shows +0.24 SD when the program adds instruction. Candidate explanation: what the comparator holds constant (time and materials). No design yet isolates it.

"Unrestricted use of a general-purpose LLM chatbot during practice improves university students' subsequent performance on assessments taken without AI." received its first assessed status: INSUFFICIENT_EVIDENCE (unclear). Evidence leans against the claim — a significant post-use quality deficit vs web search, and an underpowered retention null — but nothing adequately powered tests it directly, and the largest randomized offer cannot (partial AI at its exam). Pass B read the same evidence as NOT_SUPPORTED; the conservative status stands until something powered exists.

"Unrestricted use of a general-purpose LLM chatbot during practice improves secondary students' subsequent performance on assessments taken without AI." received its first assessed status: NOT_SUPPORTED (stable). The claim asserts improvement. The one direct, gate-passing test (preregistered cluster RCT, HIGH) found a 17% unassisted-exam deficit; nothing gate-passing supports the claim. A single setting and product generation keep this NOT_SUPPORTED rather than a harm claim of its own; replication would move it.