Pedagogy Reference

---
name: pedagogy-reference
description: "Use when correcting an education misconception, looking up an effect size, working memory, spacing interval or achievement figure, finding the books and sources, or needing a quick-reference picker — plus the current state of AI in education and post-pandemic achievement data. Companion to the other pedagogy skills."
---

# Pedagogy and Learning Science: What's Live, Misconceptions, Numbers, and Books

> **Part 5 of 5** of the *Pedagogy and the Study of Learning and Teaching* reference (plugin `pedagogy-and-the-study-of-learning`), covering §25–§30. Sibling skills: `pedagogy-reliability-memory-cognitive-load-and-robust-effects` (§0–§7), `pedagogy-explicit-instruction-worked-examples-and-feedback` (§8–§13), `pedagogy-subject-instruction-curriculum-and-classroom-practice` (§14–§21), `pedagogy-reading-research-and-zombie-ideas` (§22–§24). Section numbers are shared across the set; a reference written as §N → `skill` points into that sibling skill.
>
> **Currency:** The cognitive science is stable and well-replicated. Two areas moved. See §25 for AI in education, and post-pandemic achievement data.

> **⚠️ This field has an unusually wide gap between what is well-established and what is
> widely believed and practised.** ⚠️ **Some findings here are among the most robust in
> all of psychology; others in the same textbooks failed replication entirely; and a third
> group were never research findings at all — they were marketed.**
>
> **Complements a psychology reference (§1's replication context) and a family/community
> reference (behaviour genetics, which constrains how much any intervention can do).**
>
> **⚠️ GOTCHA** boxes mark widely-taught ideas that the evidence contradicts.
>
> **⚠️ A disclosure about the political layer:** ⚠️ **education research is unusually
> ideologically loaded, because it touches how children are raised and how resources are
> allocated.** **⚠️ Where a genuine scholarly dispute exists (§8 → `pedagogy-explicit-instruction-worked-examples-and-feedback`, §14 → `pedagogy-subject-instruction-curriculum-and-classroom-practice`) I set out both
> positions; where the evidence is one-sided I say so and show the evidence.**
>
> **The three ideas that organize this document:**
> 1. **⚠️ Learning is a change in LONG-TERM MEMORY; performance during a lesson is not
>    learning** (§2 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`, §7 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`). **The two come apart routinely, and confusing them is the single
>    most consequential error in teaching.**
> 2. **⚠️ Working memory is tiny and prior knowledge is the great multiplier** (§2 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`, §3 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`).
>    **Almost every instructional design principle that works is managing that limit.**
> 3. **⚠️ Things that FEEL like good learning usually aren't** (§4 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`, §7 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`). **Fluency,
>    familiarity and ease are systematically misleading — for students AND teachers.**

---

## §25. What's Live — verified August 2026

### 25.1 ⚠️ AI in education: strong trial results, and a crucial condition attached
**⚠️ The evidence moved fast, and the headline and the caveat come from the same
literature.**

- **⚠️ The strongest single result**: **Kestin et al., published in *Scientific Reports*,
  ran an RCT with around 500 college students comparing a custom AI tutor against in-class
  ACTIVE LEARNING** (⚠️ **not against a bad lecture — against the current best practice**).
  ⚠️ **Students using the AI tutor "learn significantly more in less time," with reported
  gains over double, and reported feeling more engaged and motivated.** **⚠️ Notably, the
  tutor was designed around the same pedagogical principles as the in-class lessons.**
- **⚠️ Corroborating trials**: **an RCT in undergraduate macroeconomics found all treatment
  arms beat control, with ⚠️ combined AI tutoring PLUS group work producing the largest
  gains — suggesting complementarity rather than substitution.** ⚠️ **Other experimental
  work reports AI access improving immediate knowledge test scores by around 6.7
  percentage points on a 56.3% baseline.** **Further RCTs cover K-12 science
  argumentation and self-regulated learning support.**

> **⚠️ GOTCHA — the most important finding is the one that cuts the other way, and it
> should govern how anyone deploys this.** ⚠️ **Bastani et al. in *PNAS* (2025),
> "Generative AI without guardrails can harm learning," ran a high-school mathematics
> trial with three arms: a standard GPT-4 chat interface, a purpose-built tutor with
> teacher-designed guardrails, and no AI.**
> ⚠️ **During PRACTICE, the tutor arm performed 127% better and the plain-chat arm 48%
> better than control.** **⚠️ But on the subsequent unaided EXAM, the title tells you the
> result: unguarded access HARMED learning.**
> ⚠️ **This is §2 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`'s performance-versus-learning distinction in its sharpest possible form.**
> **An AI that supplies answers produces excellent performance while the support is
> present and worse learning once it's removed — precisely the pattern desirable
> difficulties predict** (§4 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`). **⚠️ The design decision — whether the tool answers or
> Socratically withholds — is not a detail; it is the whole intervention.**

**⚠️ How to read this literature carefully:**
⚠️ **Most positive trials are short, use researcher-designed tutors built on good
pedagogy, measure proximal outcomes, and are often delivered by the developers — every
inflation factor in §22 → `pedagogy-reading-research-and-zombie-ideas`.** ⚠️ **"AI tutoring works" is not established; "a well-designed
AI tutor built on sound pedagogical principles, in a short trial, beat a comparison
condition" is.** **⚠️ The related literature also raises "metacognitive laziness" and
over-reliance concerns** (§7 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`), **and a Stanford review of the K-12 evidence base exists
precisely because the field needed one.**
**⚠️ The practical implication I'd draw**: ⚠️ **the pedagogy is doing the work, not the
model.** **A tool that scaffolds, withholds answers, prompts retrieval and fades support is
implementing §4 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`, §9 → `pedagogy-explicit-instruction-worked-examples-and-feedback` and §12 → `pedagogy-explicit-instruction-worked-examples-and-feedback` at scale — which is genuinely valuable. A tool that answers
homework is implementing nothing.**

### 25.2 ⚠️ Achievement data: the decline predates the pandemic
**⚠️ The framing correction matters more than the numbers.**

- **⚠️ The scale**: **since 2019, reported average declines of about five points in 4th and
  8th grade reading; ⚠️ 4th grade maths three points below 2019; ⚠️ 8th grade maths nine
  points below pre-pandemic, described as roughly a full grade level.** ⚠️ **In reading,
  students in both grades score about where they were in the early 1990s.**
- **⚠️ The distributional finding is the serious one.** ⚠️ **Declines were driven
  disproportionately by LOWER-performing students.** **Reported: the highest-performing
  students gained ground in maths from 2022 to 2024 while those below the 25th percentile
  continued to decline; ⚠️ in 4th grade reading only the 90th percentile avoided a drop,
  while the 10th percentile fell four points.** **⚠️ A record share of 12th graders scored
  "below basic" in both subjects.**
- **⚠️ Recovery has been partial and uneven**: ⚠️ **average learning losses reported around
  0.12 SD in ELA and 0.17 SD in maths — against the observation that even effective,
  well-implemented interventions rarely produce more than 0.05–0.15 SD.** ⚠️ **That
  arithmetic is why recovery is hard: the hole is roughly the size of the best available
  shovel.** **⚠️ The Education Recovery Scorecard reported that the highest-income decile
  districts were nearly four times more likely to recover in both subjects than the lowest,
  and that chronic absenteeism slowed recovery.**

> **⚠️ GOTCHA — calling this "pandemic learning loss" misdiagnoses it, and Hanushek's 2026
> analysis makes the case directly.** ⚠️ **NAEP scores PEAKED AROUND 2013 and fell notably
> BEFORE 2019, continuing through and after the pandemic.** **⚠️ The gap between high and
> low achievers was already widening from 2013.**
> ⚠️ **So the pandemic accelerated and deepened a decline that was already a decade old —
> which means "recovery to 2019 levels" is the wrong target, and interventions designed
> purely as pandemic remediation are addressing a fraction of the problem.**

**⚠️ One genuine bright spot, and it's recent**: ⚠️ **NAEP long-term trend results reported
in June 2026 showed average reading and maths scores for 9-year-olds ROSE from 2022 to
2025** — **described by the acting NCES commissioner as an optimistic release.**
⚠️ **Treat one datapoint cautiously, and note the age-9 cohort had the least pandemic
schooling disruption.**
**⚠️ The policy response has moved toward §14 → `pedagogy-subject-instruction-curriculum-and-classroom-practice` and §15 → `pedagogy-subject-instruction-curriculum-and-classroom-practice`**: ⚠️ **states have legislated on
literacy — including banning three-cueing — and on maths access.** ⚠️ **Whether that
produces measured gains is genuinely open, and I'd want several more NAEP cycles before
concluding.**
**⚠️ A methodological caution on reading any of this**: ⚠️ **NAEP achievement LEVELS
("Basic," "Proficient") are standard-setting judgements, not natural categories** (§13 → `pedagogy-explicit-instruction-worked-examples-and-feedback`) —
**"below basic" percentages are alarming and are also partly an artefact of where the cut
score was placed.** **⚠️ The SCALE SCORE trends and the percentile breakdowns are the more
defensible evidence, and they tell the same story.**

---

## §26. Misconceptions

| Misconception | Correction |
|---|---|
| Match teaching to a student's learning style | ⚠️ **The meshing hypothesis fails on testing** (§23 → `pedagogy-reading-research-and-zombie-ideas`) |
| We remember 10% of what we read | ⚠️ **Fabricated. No traceable source** (§23 → `pedagogy-reading-research-and-zombie-ideas`) |
| A smooth successful lesson means learning happened | ⚠️ **Performance ≠ learning** (§2 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`) |
| Re-reading and highlighting are good study methods | ⚠️ **Among the least effective** (§7 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`) |
| If it feels easy, it's working | ⚠️ **Fluency is an illusion. Desirable difficulties** (§4 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`, §7 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`) |
| Testing just measures learning | ⚠️ **Retrieval CAUSES learning** (§4 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`) |
| Practise one type until you've got it | ⚠️ **Interleave. Blocking removes the selection step** (§4 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`) |
| Working memory holds about seven items | ⚠️ **Revised downward, ~4 for novel material** (§2 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`) |
| Discovery learning is more engaging so better | ⚠️ **Novices lack the schemas to search with** (§8 → `pedagogy-explicit-instruction-worked-examples-and-feedback`) |
| Explicit instruction is always best | ⚠️ **Expertise reversal — guidance must fade** (§3 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`, §8 → `pedagogy-explicit-instruction-worked-examples-and-feedback`) |
| Comprehension is a transferable skill | ⚠️ **It's largely domain knowledge** (§5 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`, §14 → `pedagogy-subject-instruction-curriculum-and-classroom-practice`) |
| Teach skills, not content | ⚠️ **Critical thinking is domain-specific** (§5 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`, §23 → `pedagogy-reading-research-and-zombie-ideas`) |
| Memorization is obsolete, just look it up | ⚠️ **You can't think with what isn't in memory** (§2 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`, §23 → `pedagogy-reading-research-and-zombie-ideas`) |
| Guessing from context is what good readers do | ⚠️ **It's what WEAK readers do** (§14 → `pedagogy-subject-instruction-curriculum-and-classroom-practice`) |
| Phonics solves reading | ⚠️ **Decoding × language comprehension. Two terms** (§14 → `pedagogy-subject-instruction-curriculum-and-classroom-practice`) |
| Procedural fluency opposes understanding | ⚠️ **They develop iteratively** (§15 → `pedagogy-subject-instruction-curriculum-and-classroom-practice`) |
| Growth mindset interventions reliably raise achievement | ⚠️ **Much smaller than sold** (§6 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`) |
| More feedback is better | ⚠️ **~a third of interventions made things worse** (§12 → `pedagogy-explicit-instruction-worked-examples-and-feedback`) |
| Grades plus comments give the best of both | ⚠️ **Students read the grade and ignore comments** (§12 → `pedagogy-explicit-instruction-worked-examples-and-feedback`) |
| Differentiate by giving different tasks | ⚠️ **Weak evidence, high cost, lowered expectations** (§18 → `pedagogy-subject-instruction-curriculum-and-classroom-practice`) |
| Lectures don't work | ⚠️ **Unstructured non-interactive ones don't** (§19 → `pedagogy-subject-instruction-curriculum-and-classroom-practice`) |
| Students know which teaching helps them most | ⚠️ **They rate active learning lower while learning more** (§19 → `pedagogy-subject-instruction-curriculum-and-classroom-practice`) |
| Hattie's effect sizes rank what works | ⚠️ **Substantively criticized. Hypotheses, not rankings** (§22 → `pedagogy-reading-research-and-zombie-ideas`) |
| Effect size 0.4 is the threshold for mattering | ⚠️ **Context-dependent; ~0.1 can matter** (§22 → `pedagogy-reading-research-and-zombie-ideas`) |
| Young people are digital natives | ⚠️ **Not supported** (§23 → `pedagogy-reading-research-and-zombie-ideas`) |
| It's pandemic learning loss | ⚠️ **Scores peaked ~2013 and fell before 2019** (§25.2) |
| AI tutors improve learning | ⚠️ **Well-designed ones did in trials; unguarded access HARMED it** (§25.1) |

---

## §27. Numbers

```
⚠️ Working memory        ⚠️ ~4 chunks for novel material (not 7±2)
⚠️ Wait time             ⚠️ ~3 seconds transforms answer quality
⚠️ Feedback              ⚠️ ~1/3 of interventions REDUCED performance
⚠️ Education effect size ⚠️ ~0.1 meaningful · 0.4 large · ⚠️ >1.0 suspect
⚠️ Learning loss         ⚠️ ~0.12 SD ELA · ~0.17 SD maths
⚠️ Best interventions    ⚠️ rarely exceed 0.05–0.15 SD
⚠️ NAEP since 2019       reading −5 pts (G4, G8) · G8 maths −9 pts
⚠️ NAEP peak             ⚠️ ~2013, declining before the pandemic
⚠️ Reading levels        ⚠️ roughly early-1990s equivalent
⚠️ Recovery inequality   top income decile ~4× more likely to recover
⚠️ AI tutor RCT          N≈500; ⚠️ reported >2× learning gains, less time
⚠️ Bastani PNAS          practice +127% (tutor) / +48% (plain chat);
                         ⚠️ unguarded access HARMED exam learning
⚠️ Age-9 long-term trend ⚠️ reading and maths ROSE 2022→2025
```

---

## §28. Books and Sources

| Author | Work | Why |
|---|---|---|
| **Willingham** | ***Why Don't Students Like School?*** | ⚠️ **The best single starting point. §2 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`, §5 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`** |
| **Brown, Roediger & McDaniel** | ***Make It Stick*** | ⚠️ **§4 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`, for a general audience** |
| **Sweller, Ayres & Kalyuga** | *Cognitive Load Theory* | ⚠️ **§3 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects` from the source** |
| **Rosenshine** | ***Principles of Instruction*** | ⚠️ **§8 → `pedagogy-explicit-instruction-worked-examples-and-feedback`. Free, ~10 pages, the best value in the field** |
| **Dunlosky et al.** | *Improving Students' Learning* (2013) | ⚠️ **§7 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`'s utility ratings** |
| **Soderstrom & Bjork** | *Learning versus Performance* | ⚠️ **§2 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`'s core distinction** |
| **Kirschner, Sweller & Clark** | *Why Minimal Guidance Doesn't Work* | ⚠️ **§8 → `pedagogy-explicit-instruction-worked-examples-and-feedback`, one side, stated well** |
| **Hirsch** | *Why Knowledge Matters* | ⚠️ **§16 → `pedagogy-subject-instruction-curriculum-and-classroom-practice`. Argumentative and important** |
| **Seidenberg** | ***Language at the Speed of Sight*** | ⚠️ **§14 → `pedagogy-subject-instruction-curriculum-and-classroom-practice`. The best book on reading science** |
| **Wiliam** | *Embedded Formative Assessment* | ⚠️ **§11–§13 → `pedagogy-explicit-instruction-worked-examples-and-feedback`, practical** |
| **Coe et al.** | *What Makes Great Teaching?* | §24 → `pedagogy-reading-research-and-zombie-ideas` |
| **Lemov** | *Teach Like a Champion* | ⚠️ **§17 → `pedagogy-subject-instruction-curriculum-and-classroom-practice`. Technique-level, atheoretical, useful** |
| **EEF Toolkit / What Works Clearinghouse** | — | ⚠️ **§22 → `pedagogy-reading-research-and-zombie-ideas`. Check claims here first** |
| **Kluger & DeNisi (1996)** | Feedback meta-analysis | ⚠️ **§12 → `pedagogy-explicit-instruction-worked-examples-and-feedback`'s uncomfortable finding** |

---

## §29. Quick Reference

### 29.1 Picker
| Question | Where |
|---|---|
| How should I study? | ⚠️ **Retrieval + spacing. Stop re-reading** (§4 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`, §7 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`) |
| They knew it yesterday | ⚠️ **Performance wasn't learning. Space and retrieve** (§2 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`, §4 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`) |
| Students are overwhelmed | ⚠️ **Cognitive load — cut extraneous, sequence intrinsic** (§3 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`) |
| My explanation isn't landing | ⚠️ **Worked examples; check for misconceptions** (§9 → `pedagogy-explicit-instruction-worked-examples-and-feedback`, §10 → `pedagogy-explicit-instruction-worked-examples-and-feedback`) |
| Should I lecture or let them explore? | ⚠️ **Depends on their prior knowledge** (§3 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`, §8 → `pedagogy-explicit-instruction-worked-examples-and-feedback`) |
| Nobody answers my questions | ⚠️ **Wait time, then all-student response** (§11 → `pedagogy-explicit-instruction-worked-examples-and-feedback`) |
| They ignore my written feedback | ⚠️ **Remove the grade; give time to act** (§12 → `pedagogy-explicit-instruction-worked-examples-and-feedback`) |
| Wide range of attainment in one room | ⚠️ **Vary support, not goals. Tutoring if available** (§18 → `pedagogy-subject-instruction-curriculum-and-classroom-practice`) |
| Child can't read | ⚠️ **Test decoding vs comprehension separately** (§14 → `pedagogy-subject-instruction-curriculum-and-classroom-practice`) |
| Comprehension is weak despite decoding | ⚠️ **Knowledge and vocabulary** (§5 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`, §14 → `pedagogy-subject-instruction-curriculum-and-classroom-practice`, §16 → `pedagogy-subject-instruction-curriculum-and-classroom-practice`) |
| Is this intervention worth it? | ⚠️ **§22 → `pedagogy-reading-research-and-zombie-ideas`'s questions before the effect size** |
| Someone cites a percentage pyramid | ⚠️ **It's fabricated** (§23 → `pedagogy-reading-research-and-zombie-ideas`) |
| Should we buy this edtech? | ⚠️ **What's the comparison, who funded, what outcome** (§21 → `pedagogy-subject-instruction-curriculum-and-classroom-practice`, §22 → `pedagogy-reading-research-and-zombie-ideas`) |
| Should students use AI? | ⚠️ **Depends entirely on whether it answers or scaffolds** (§25.1) |

### 29.2 Designing a sequence
- [ ] ⚠️ **Prerequisites identified and secured first** (§5 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`, §16 → `pedagogy-subject-instruction-curriculum-and-classroom-practice`)
- [ ] ⚠️ **Extraneous load stripped out of materials** (§3 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`)
- [ ] Worked examples before independent problems, faded (§9 → `pedagogy-explicit-instruction-worked-examples-and-feedback`)
- [ ] ⚠️ **Checking for understanding built in, not "any questions?"** (§11 → `pedagogy-explicit-instruction-worked-examples-and-feedback`)
- [ ] ⚠️ **Retrieval practice scheduled, spaced, cumulative** (§4 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`)
- [ ] Practice interleaved once types are individually secure (§4 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`)
- [ ] ⚠️ **Feedback students have TIME to act on** (§12 → `pedagogy-explicit-instruction-worked-examples-and-feedback`)
- [ ] ⚠️ **Assessment aligned to what you actually want learned** (§13 → `pedagogy-explicit-instruction-worked-examples-and-feedback`)
- [ ] ⚠️ **Success looks like retention weeks later, not a smooth lesson** (§2 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`)

---

## §30. Method

**§1–§24 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`, `pedagogy-explicit-instruction-worked-examples-and-feedback`, `pedagogy-subject-instruction-curriculum-and-classroom-practice`, `pedagogy-reading-research-and-zombie-ideas` rests on the cognitive science of learning and on instructional research** —
**memory architecture, cognitive load theory, the spacing and testing effects, the Simple
View of Reading, and the standard critical apparatus for reading education studies.**
⚠️ **The spacing effect dates to Ebbinghaus in the 1880s and has survived everything since;
none of §4 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects` needed verification.**

**Two searches were run in August 2026**, on **AI in education** and **achievement data** —
⚠️ **chosen because the first is where practice is changing fastest and the second is where
the public framing is most misleading.**

**Confidence.** **High** in §2 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`, §4 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects` and §7 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects`, which are the load-bearing sections and the
ones I'd most want read. ⚠️ **The performance-versus-learning distinction, the robust
effects, and the illusion of fluency together explain most of why students study badly and
why lessons that look good can teach little.**

**High** in §14 → `pedagogy-subject-instruction-curriculum-and-classroom-practice`'s reading science and §23 → `pedagogy-reading-research-and-zombie-ideas`'s zombie ideas — ⚠️ **learning styles in
particular has been tested directly and repeatedly and fails, and I've stated that without
hedging because the evidence genuinely is one-sided.**

**Moderate-to-high** in §8 → `pedagogy-explicit-instruction-worked-examples-and-feedback`, and I've flagged it as contested rather than settled.
⚠️ **My position — that explicit guidance is clearly superior for novices, that expertise
reversal means it must fade, and that the "which is better" framing is itself the error —
is a reading of the evidence, and constructivist scholars would characterize the debate
differently.** **I've given their strongest objection (that "minimal guidance" is a straw
man for well-scaffolded inquiry) rather than only the case I find persuasive.**
⚠️ **Similarly §6 → `pedagogy-reliability-memory-cognitive-load-and-robust-effects` on growth mindset and §22 → `pedagogy-reading-research-and-zombie-ideas` on Hattie are stated as substantive critiques
with named grounds, not dismissals.**

**High** on §25.1's individual study findings, **moderate** on what they collectively
mean. ⚠️ **The Kestin RCT in *Scientific Reports* and the Bastani *PNAS* trial are both
peer-reviewed and I've reported both — and the pairing is the point, because the second
contradicts the optimistic reading of the first.** ⚠️ **I want to be explicit that the
positive AI-tutoring literature has every inflation characteristic §22 → `pedagogy-reading-research-and-zombie-ideas` warns about: short
duration, developer-built tools, proximal outcomes, and researcher delivery.** **⚠️ My
conclusion that "the pedagogy is doing the work, not the model" is an inference from the
guardrails result, not a finding anyone has stated in those terms.**

**High** on §25.2's figures, which come from NAEP, NCES, Brookings, the Education Recovery
Scorecard and Hanushek's *Education Next* analysis. ⚠️ **The reframing — that the decline
began around 2013, well before the pandemic — is Hanushek's argument and I find it well
evidenced, but it is an interpretive claim about causes and it carries policy
implications, so I've attributed it.** ⚠️ **I've also flagged the standard-setting caveat
on NAEP achievement levels, because "below basic" figures are widely quoted without noting
that the cut score is a judgement.**