The evidence bar · Science of Learning Institute

Every practice on the site gets one of four grades. The grade is shown on every card, at every depth.

Only randomized trials in real schools earn the top gradeThe four grades shown on every card

Proven in classroomsRandomized trials in real schools, plus a lab reason it works
StrongMany controlled experiments; fewer classroom trials, same direction
PromisingSound theory, some experiments; classroom results thin or mixed
RetiredTested and failed, or never properly tested

The four grades

Grade	What it takes	Examples
Proven in classrooms	Randomized trials in real schools, with real teachers, plus a lab finding that explains why it works	Mixed (interleaved) math practice; low-stakes quizzing; systematic phonics; small-group tutoring; knowledge-building curriculum
Strong	Large body of controlled experiments and meta-analyses of experiments; classroom trials point the same way but are fewer	Spaced practice; worked examples in math; explicit teaching over unguided discovery
Promising	Sound theory and some experiments, but thin or mixed classroom evidence	Dual coding; self-explanation prompts; productive failure for older students
Retired	Tested and failed, or never properly tested	Learning styles; "7 plus or minus 2" working memory; Hattie rankings as proof on their own

House rules for evidence

Every number says what it compares. "Quizzing beat restudying" is a claim. "d = 0.5" alone is not.
Randomized trials and meta-analyses of experiments come first. Correlational studies only appear as "why it matters," never as "proof it works."
Hattie's effect sizes are never cited as evidence. They mix study types and are widely criticized. We go to the underlying studies instead.
Every principle page has a What we don't know yet box. The EEF's own review found classroom evidence is thinner than lab evidence, and that multi-strategy programs tested at scale often showed small effects (Perry, Chartered College(opens in a new tab)). Saying so out loud is what makes the rest believable.
Counter-evidence gets its own section, written fairly. If productive failure has a meta-analysis behind it, we show it and explain where it applies.
Every page shows when it was last reviewed, and every evidence file keeps a dated change log.

How studies are labeled

Label	Meaning
Randomized trial	Students, classes, or schools assigned by chance. The strongest single-study design.
Lottery study	Admission by lottery, which works like random assignment for those who applied.
Meta-analysis	A combined analysis of many studies; only as strong as the studies inside it.
Comparison study	Groups compared without random assignment. Useful, but differences may come from who chose what.
Lab study	Controlled experiment outside a regular classroom. Explains why; doesn't prove it works at scale.
Correlation	Two things move together. Shows a link, never cause. Used only for "why it matters."

How to read an effect size

An effect size of 0.5 means the average student in the trial group scored about as well as a student at the 69th percentile of the comparison group. School-based trials rarely find effects that large, so 0.2 to 0.3 in a well-run classroom trial is worth paying attention to.

Contested, shown fairly

Productive failure. A meta-analysis of 53 studies found problem-solving before instruction helped conceptual understanding, mostly for secondary and university students, less for primary (Futurity summary(opens in a new tab)). The site presents it as a tool for older students who already have the basics, not a replacement for explicit teaching.
Guided discovery. Alfieri's second analysis favored enhanced discovery, but critics note it includes adult and university studies (Ashman(opens in a new tab)). We show both.
Phonics effect size. The NRP's 0.41 and reanalyses' 0.24 are both positive. We show the range, not one number.

When a grade changes

A grade moves up only with a new randomized trial or meta-analysis of trials that meets the evidence bar.
A grade moves down when a strong study contradicts it, even if the old studies still stand.
Every change is logged with the date, the study, and one sentence on why.
Contested findings get a "contested" note rather than a quiet change.