Independent research archiveOpen access · Versioned · Source-linked
Piccini Research Archive
Research notes with a falsifiable edge.
Bounded questions, current evidence, calibrated conclusions, and the conditions that would change them. These are independent research syntheses and exploratory empirical studies—not peer-reviewed publications.
45 papersEvidence current to Nov 2026Open web edition
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
Hot exam days reliably lower scores, especially in math and without air conditioning: a 90°F exam day can cost about 13% of a standard deviation, and heat effects are concentrated at extremes and near pass/fail cutoffs. But whether heat reduces cumulative learning is less settled. Cumulative-exposure studies find losses; long-run studies find adaptation and compensation. Two analyses of Brazil's ENEM reach opposite signs, a reminder that context and specification matter. Heat tends to widen gaps where cooling is unequal. The defensible policies are cooling, ventilation, exam scheduling, and sleep protection, not a single universal estimate of learning loss.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
Cash transfers are credited with large falls in child mortality. This brief separates the ingredients: money, conditions, and clinics. Money and scale have the clearest dose-response; conditionality has never been shown to save lives on its own; nutrition gains depend on complementary services. And the mortality evidence is far more Latin American than the global claim suggests.
Social protectionChild mortalityCash transfersConditionalityEvidence synthesis
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
In double-blind trials covering 123,940 participants, statins produced a 3 percent relative increase in muscle pain or weakness — about 11 extra events per 1,000 person-years in year one and no detectable excess afterward. Only roughly one in fifteen first-year reports was attributable to the drug. Three-arm n-of-1 trials explain the gap: symptom scores were 8.0 out of 100 with no tablet, 15.4 with placebo, and 16.3 with atorvastatin. The pill, not the molecule, carried most of the burden. Statins do cause a small real muscle excess, a modest increase in diabetes and liver enzymes, and rare myopathy or rhabdomyolysis. They do not appear to commonly cause clinically significant cognitive impairment. The policy problem is that labels and public discourse present reported effects as established ones, and discontinuation driven by that framing is associated with more heart attacks and cardiovascular deaths.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
Preschool raises school readiness by roughly 0.2 to 0.3 standard deviations for disadvantaged children, but those test-score gains usually fade within a few years. Long-run effects are real in some programs — Perry and Abecedarian produced large adult benefits, and Boston's public preschool raised on-time college enrollment by 8.3 percentage points — yet Tennessee's program produced negative estimates through sixth grade, and Quebec's universal expansion had lasting negative effects. The pattern is not random: intensive, high-quality programs serving very disadvantaged children with weak alternatives do best; low-quality expansions into already adequate care can do harm. Test scores alone mislead in both directions.
Early childhood educationPreschoolChild developmentEducation policyEvidence synthesis
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
Breastfed children score about 3–7 IQ points higher in observational studies, but most of that gap reflects who breastfeeds rather than what breast milk does. The PROBIT randomized trial produced a 5.9-point full-scale estimate at age 6.5 with a confidence interval that included zero; at 16 years the intention-to-treat effect was null, with small verbal and memory advantages after adjustment. Sibling comparisons find effects near zero. A 2025 Mendelian randomization study found no causal effect on intelligence, and the leading FADS2 interaction failed to replicate in a 12,000-person meta-analysis. The observational cohorts that survive adjustment suggest a modest, mostly verbal effect. The best estimate is 0–3 full-scale IQ points — real but small, and not a reason to pressure mothers.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
Alcohol is a Group 1 carcinogen causing cancers of the mouth, pharynx, larynx, oesophagus, colorectum, liver, and breast, and accounts for roughly 3–4% of global cancer cases and deaths. The dose-response is approximately linear with no demonstrated threshold, and a measurable share of cases arises below one drink per day. The newest Mendelian randomization study (about 1.5 million people, 20 cancers) supports head and neck, colorectal, and oesophageal sites but is null for breast cancer and overall cancer, while the observational breast literature and the acetaldehyde mechanism point the other way. The brief separates the population claim from the individual one: 'no safe level' means no threshold has been shown, not that light drinking is equally dangerous for everyone.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
The Family Health Strategy literature credits full coverage with roughly 5-20% lower infant and amenable mortality, concentrated in poor municipalities. But the cleanest natural experiment — the overnight exit of 8,332 Cuban doctors from about 2,800 municipalities in 2018 — cut chronic-care utilization without moving mortality, contradicting forecasts of tens of thousands of extra child deaths. The likely reason is that the programme's effect runs through team continuity and community health workers, not doctor headcount, so a doctor-only shock is the wrong test. Brazil's headline coverage figure also changed definition in 2021, making cross-break comparisons invalid.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
School phone bans reliably cut in-school use — from 61% to 13% of students in a national U.S. pouch study — but average test-score effects are small: near zero in the U.S., +0.06 SD in Rio, second-year gains in Florida, +0.07 SD in England, and zero in Sweden. New 2026 randomized evidence changes the question: mandatory collection raised grades in a 16,683-student Indian trial, while Chinese trials found teacher-guided phone use beat a ban, and unstructured use was worse than either. Well-being effects are weak and inconsistent. Phone-free instruction is defensible; a population mental-health fix is not.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
Household screen-reduction trials move time use and symptoms, not measured cognition: a Danish cluster RCT (+45.8 min/day activity, no EEG sleep change), a UK toddler RCT (sleep efficiency d 0.56, no effect on objective attention), a student smartphone RCT (symptom gains that reverted at six weeks) and a 2026 detox pilot (working-memory gain gone at two months). Observational executive-function associations are weak and bidirectional (81% of 58 longitudinal studies negative, but 9 run EF → screen use), and exposure measurement is poor (self-report vs logs; 14-18% under-reporting). School phone bans show near-zero average test-score effects. No causal effect on cognition is demonstrated.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
A log-scale regression coefficient reads as a percentage only approximately, and the error grows with the coefficient. A working paper on competition among Brazilian private higher-education providers reports IV coefficients of -3.33 and -3.10 on logged benefit and cost and calls them 'approximately 333%' against 'a 310% rise'. Read as ordinary percentages the same numbers are about 2,694% and 2,120%; read as natural-log percentages they are exactly 333 and 310. The paper's direction — benefits rise faster than costs — survives the natural-log reading (ratio 1.07) but not the exponential one (1.26). Three percentage conventions, one coefficient, magnitudes differing about eightfold: the slope cannot decide which is right.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
PISA 2025 (released September 2026) confirms the declines continued: OECD mathematics −9 and reading −14 points since 2022, science flat on average but with widened gaps; 2015–2025 losses reach −22 mathematics and −28 reading. PIRLS 2021 found reading down in 21 of 32 comparable countries versus 2016 (though flat since 2001); TIMSS 2023 found grade-8 mathematics down in 14 of 34 comparable countries versus 3 up, with the US bottom tail at record lows. Pandemic closures are the best-identified shock (Engzell 0.08 SD in the Netherlands; Jakubowski et al. closure dose-response across 72 countries) but cannot explain pre-pandemic weakening or post-reopening declines. School phone bans show no average achievement effect in the best quasi-experiments; OECD's own 2025 reading: digitalisation, screen time and falling reading enjoyment 'have coincided with' weaker literacy. The reading-engagement channel is the leading secular candidate; a large residual remains unexplained.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
INSS records map 40.6 million benefits to 36.5 million people; no source deduplicates across civil and military regimes. Half of RGPS recipients get exactly the floor, while 3.5% of survey pension recipients above the RGPS ceiling receive 19.4% of payments. Concentration coefficients: BPC −0.083, RGPS 0.433, RPPS 0.821 against a 0.535 Gini. BPC has about twice the static poverty exits per real of pensions. One federal cohort's lifetime returns rise from 4.2% to 5.5% with income. Seven questions remain unanswerable from public data.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
Non-employed residents per employed resident fell from 1.71 (1976) to 1.07 (2025), and demographic, employment, and NTA economic support all bottom out between 2017 and 2025 before rising. Meanwhile RGPS spending rose from 4.6% to 8.0% of GDP (1995–2024), mostly through more benefit records per resident, and all public pension systems cost about 12.4% of GDP in 2024. The official current-rules path reaches 16.76% of GDP for RGPS alone by 2100. Descriptive, not causal; no all-regime history exists.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
NTP (2024) and Taylor et al. ( JAMA Pediatr 2025) find inverse associations at higher fluoride exposure (above ~1.5 mg/L), with the water dose-response null below 1.5 mg/L in all studies. Prospective cohorts at policy-relevant levels — Dunedin, Swedish registry, Danish Odense, U.S. HS&B, and U.S. Wisconsin — show no IQ reduction at 0.7 mg/L; the best-identified ones show null or modest benefit. Modern evidence puts the caries benefit at small and uncertain. Verdict: the high-dose signal is real, the low-dose harm claim is not established.
Public healthEnvironmental healthChildren's cognitionEvidence synthesisCausal inference
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.1
A dozen quasi-experimental studies agree that dirty air on the day of a high-stakes test lowers scores by roughly 0.02–0.08 SD per 10 µg/m³ PM2.5, concentrated in math and consequential for later schooling and earnings; pollution also raises absences, discipline referrals, and teacher illness. The claim that years of exposure permanently lower cognition is contested: China CFPS and a PM2.5–IQ meta-analysis find harm, while a 17-year English cohort, a Dutch birth cohort, and PM2.5 mendelian randomization find little. The effect is small, real, regressive, and at least partly reversible.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
Meta-analyses agree on ~0.14–0.20 SD learning loss from COVID-19 closures, worse in math and for disadvantaged students. TIMSS 2023 shows achievement still 0.11 SD below trend across 71 systems. Recovery is mixed — full in Tamil Nadu, partial in São Paulo and the US, average-level catch-up in Germany — and tracks remediation effort (tutoring, extended time, phone-based targeted instruction), not reopening by itself.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
Plastic particles are in human blood, lungs, placentas, and plaque — that is established. That they cause cardiovascular disease is not: the human evidence amounts to a few hundred high-risk surgical patients, a flagship study with a wide 4.53 hazard ratio, an analytical-method audit calling the main measurement technique unsuitable, and zero exposure-reduction trials. A real hazard signal; not a demonstrated cause.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.1
Does childhood lead exposure cause adult crime? Yes, with converging causal evidence. Did lead abatement explain most of the 1990s crime drop? Probably not: the bias-corrected meta-analysis credits it with a minority share. Lead is one real causal lever among several, and legacy exposure from paint, soil, and water pipes still matters.
Public healthEnvironmental economicsCriminologyLead exposureEvidence synthesis
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.1
Does systematic phonics teach children to read? The strongest evidence says it reliably builds early decoding and spelling, most of all for beginners and struggling readers, but that the transfer to comprehension is small, early, and fragile, and that national phonics mandates have not produced the comprehension gains promised for them.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
Do smartphones and social media cause the adolescent mental-health crisis? The best causal studies disagree on size but not on direction: the average effect is small, the harm is concentrated in girls, heavy users, and early exposure, and the claim that social media explains the population trend outruns the evidence.
TechnologyAdolescent mental healthSocial mediaCausal inferenceEvidence synthesisPublic health
Italian municipal evidence supports a narrower claim than a universal cure: fiscal constraints can reduce pre-election capital-spending surges. Six cases inform a Brazilian design linking independent costing to procedure, correction and controlled exceptions. Includes a Portuguese edition.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
Iron is the world's most common deficiency and one of its most widely supplemented nutrients. The randomized evidence converges: fixing real deficiency buys small gains on cognitive tests, fixing nothing buys nothing, school achievement never budge, and giving iron to iron-sufficient infants may slightly harm their cognition for years.
An exploratory annotation study of Brazilian academic abstracts finds strongly asymmetric leftward labels among the minority assigned an ideological stance, alongside common use of data and substantial disciplinary variation. Includes a linked Portuguese edition and public methods supplement.
BrazilAcademic researchContent analysisIdeological orientationExploratory study
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
Meta-analyses of dozens of programmes agree that cash transfers raise school enrolment and attendance. The learning evidence is far thinner and often null — and can turn negative when enrolment growth crowds classrooms. Cash transfers are an access policy, not a learning policy, and Brazil's Bolsa Família fits the same pattern.
Social protectionEducationConditional cash transfersLearning outcomesEvidence synthesis
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
Randomized trials show stimulants fade as a symptom treatment within months, and the long MTA follow-up found no lasting symptom or academic benefit. Registries, in contrast, link medicated periods to 15-40% lower rates of crime, transport accidents, self-harm, and death. The reconciliation is that stimulants reduce discrete high-consequence events through momentary attention and impulse control without resetting the ADHD trajectory — and the registry benefits are real but inflated by confounding.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
PISA rankings, 'national IQ' datasets, and generational IQ-decline headlines all compare test scores — and all depend on measurement invariance, the property that a test measures the same ability the same way across groups and time. The evidence says comparability is conditional and local: PISA rankings are fairly robust within rich countries but not for dispersion or poorer countries; clinical batteries are invariant when properly adapted and not when they are not; and a 2019-2026 wave of invariant studies finds real generational gains alongside a weakening general factor. Treat every score comparison as a claim to be tested, not a given.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
Do gene-expression signatures of aging reproduce across independent studies? Mostly no — far worse than DNA-methylation signatures — and a large share of the reproducible whole-blood signal is a readout of shifting blood-cell composition rather than of cells aging. Signatures replicate better only when study-level heterogeneity, cell type, and a second molecular layer are modelled explicitly. The reproducible core is real, but smaller and more immune-dominated than the headlines.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
Retrieval practice is the most reliable learning technique psychology has found, but does it help with material you were never tested on? The evidence says the transfer is real but modest: strongest for near and inferential questions, weak or absent for far transfer, and much smaller than the headlines once feedback and exposure are equated. Testing builds durable, flexible knowledge only when retrieval succeeds, is repeated, gets feedback, and is applied to varied problems — and the field's own designs routinely overstate the effect by conflating direct and forward testing effects.
Learning scienceCognitive psychologyRetrieval practiceTransfer of learningEducationMeta-analysis
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.1
Minimum-wage law is written for the formal sector, but in most developing economies most workers are informal. The textbook prediction — a higher floor pushes workers into informality and lowers informal wages — is only partly borne out: Brazil's large increases raised informal wages and displaced few formal workers, Mexico's border doubling barely moved aggregate employment, and the losses that do appear concentrate on the least-skilled and best-enforced formal sectors.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
Diagnosed ADHD has climbed for decades while measured community prevalence has stayed flat. Three lines of evidence — a moving diagnostic threshold, the relative-age effect, and falling genetic liability in new cohorts — say most of the rise is a widening net, not a sicker generation. But overdiagnosis and undertreatment coexist, and the split runs along income and access.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
Bolsa Família was long found not to discourage work, and the real fight was about formality. But the program has been multiplied in value and rebuilt since 2019, and the newest causal tests disagree. This brief sorts the evidence by which version of the program each study actually measures.
Social protectionLabor supplyConditional cash transfersInformalityBrazilEvidence synthesis
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.1
A year of schooling reliably raises knowledge-loaded cognitive-test scores (1–5 IQ points), but designs that separate school days from aging find the gains concentrated in crystallized measures while fluid reasoning grows with age, not school. Whether the general factor of intelligence moves is contested — study-level, not school-level, questions matter — and a 2025 trial in India points to sustained-effort training as one real mechanism.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.2
Two completed randomized trials now support a large dengue reduction (~71–77%): Yogyakarta (wMel replacement) and Singapore (male wAlbB suppression), where protection was strongest during high transmission and faded in the inter-epidemic trough. Observational city programmes claim 63–97% on weaker designs, a 2026 meta-analysis pools nine studies to ~85%, and a 2026 review argues for a pause. Brazil's EVITA trial has now shown the replacement release established (median 71% introgression, 75% fewer wMel-negative mosquitoes) but has not reported clinical efficacy, and a new Rio analysis shows introgression can stall at 26%, especially in dense, low-income, high-dengue areas.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.1
Bolsa Família is credited with large reductions in child mortality, and a 2026 independent study of the 2012 reform strengthens the pre-2020 case. But the strongest evidence still stops at 2019 — before the pandemic suspension, Auxílio Brasil, and the 2023 rebuild. This brief separates what is measured from what is assumed.
Social protectionChild mortalityConditional cash transfersBrazilEvidence synthesis
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.1
Ding et al. 2025: 7,000 steps a day versus 2,000 gives 47% lower all-cause mortality (HR 0.53, 0.46–0.60; 14 studies), with inflections at 5,000–7,000 and GRADE-moderate certainty. A published correction touched only an appendix table; two 2025 letters attack reverse causation and wrist-versus-hip wear. RCTs add about 1,656 steps over 12 weeks; the Da Qing trial shows +1.44 years of life expectancy. Verdict: the gradient is real, the number is not.
Public healthPhysical activityEvidence synthesisMeta-analysisCausal inference
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
In 2025 a PNAS study of 283 British, Bolivian and Namibian children found that more than half of unschooled children could not repeat two words backwards at any age, and argued that executive-function tasks measure culturally specific skills alongside universal control. Two 2026 commentaries called the conclusion an overreach, and a reanalysis of the same data found EF still predicted math and reading inside schooled samples. The measurement-invariance record is mixed: partial scalar invariance in several large schooled samples, but a preregistered US study of about 11,400 fourth-graders could not establish even configural invariance between any pair of racial or ethnic groups. Treat an EF score from an unfamiliar educational background as uninterpretable until invariance is demonstrated; publish invariance tests by default; and treat any single low score as a statement about task familiarity, not about a child.
Cognitive scienceCross-cultural psychologyPsychometricsMeasurement invarianceExecutive function
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
Childhood self-control predicts health, wealth and offending for three decades in cohorts with 96% retention, but two replications by the original team reduced the famous delay-of-gratification effect to roughly 0.1 SD and then to null by age 26, and a GRADE-based review of 35 intervention studies found no converging efficacy evidence. The residual signal is substantially measurement and context: in Dunedin, childhood self-control correlates r = .45 with childhood intelligence, and preschoolers wait longer only when the experimenter's promises prove reliable. Treat self-control as a signal worth measuring and a poor standalone target; fund predictable environments over abstract inhibition drills, and require distal adult outcomes from any program claiming lifelong benefit.
Developmental psychologyExecutive functionMeasurement validityCausal inferenceChild development
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
Hearing loss appears in the 2024 Lancet commission's fourteen modifiable dementia risk factors, and cohort after cohort links it to later cognitive decline. Randomized evidence has now tested the obvious intervention twice and both primary endpoints came back null: ACHIEVE (977 adults aged 70–84, three years) found no difference in global cognitive decline, and CHOICE (703 adults with mild cognitive impairment in Shanghai, 24 months) found no significant reduction in conversion to dementia, despite a strong secondary signal on cognitive improvement. The trials did not sit idle on other outcomes. ACHIEVE found 27% fewer falls, one extra person in the social network after three years, and a slower rise in GFAP, a blood marker of neuroinflammation. Both found larger effects in higher-risk participants — a prespecified 48% reduction in decline among the cardiovascular-cohort subgroup in ACHIEVE, and a 62% reduction among its highest predicted-risk quartile. The observational record is large and consistent, including a Mendelian randomization odds ratio of 1.74 for all-cause dementia, but randomized evidence removes exactly the confounders that generate those numbers. The conclusion is deliberately narrow. Hearing loss should be treated for hearing, participation, safety, and quality of life, all of which are established. Its value as dementia prevention is plausible, concentrated in a high-risk subgroup, and unreplicated.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
The claim that cash transfers make people work less fails as a general statement. Across 115 studies of 72 programs in low- and middle-income countries, labor-force participation rises about 3 percentage points and earnings rise; in high-income settings, unconditional payments without a phase-out produce small or null participation effects and a few hours less work per week. The large negative effects come from means-testing: a Barcelona trial found 22% lower employment, roughly doubling under full withdrawal. The sign tracks which constraint binds. Subsistence, health and search constraints dominate in poor settings, where cash expands work; the net return to earning dominates where benefits are clawed back, where cash shrinks it. Labor supply is also the wrong summary statistic: in the largest U.S. trial, household income fell about $1,700 a year ex-transfer and partners missed promotions, while Brazil's Bolsa Familia expansion raised employment and cut mortality 14%.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
Controlled trials show large AI speedups, field studies show small ones, and one randomized trial found experienced practitioners slowed by 19%. The account that reconciles them is the expertise reversal effect plus a verification-cost term: assistance pays the novice's generation cost and charges the expert a review cost, so the effect changes sign across the user-task pair. A 2026 meta-analysis of 23 programming studies finds a moderate pooled effect (g = 0.33) with I² = 99% heterogeneity, where setting and not randomization is the only significant moderator. Deployment decisions therefore require task-level measurement stratified by task-specific experience, not occupation-level averages or self-report.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
Oversight of AI agents is feasible but structurally fragile. Error detection degrades with task length and sustained engagement. Intermediate checkpoints substantially outperform both confirm-at-end and confirm-every-step. Extended agent use erodes the vigilance and domain skill oversight depends on, making oversight quality a depletable resource rather than a fixed human input.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
AI reliably improves task output while degrading the accuracy of human confidence judgments. LLM input more than doubles overconfidence, and fluent explanations inflate the illusion of explanatory depth. The mechanism is fluency-driven false comprehension bypassing effortful self-assessment.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
Paternal age is a real but modest risk factor for autism and schizophrenia. De novo mutations mediate approximately 28% of the schizophrenia association. Mendelian randomization found no positive causal effect on children's test scores after genetic confounding control. The majority of the epidemiological association is attributable to inherited genetics and environmental confounding.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
Direct-answer LLM access can improve assisted work while reducing immediate unaided performance. Scaffolded tutors can produce genuine short-term gains. Durable retention and professional transfer remain under-tested, so capability claims require delayed, unaided holdouts rather than assisted completion metrics.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
Generative AI reliably improves many assisted outputs, yet its effect on the capability left behind has no stable sign. Randomized studies find both retained gains and performance-learning reversals. The most useful current model is whether assistance preserves or replaces the learner’s generation, retrieval, diagnosis, and evaluation.