Lumi is built on the Science of Reading. We are committed to rigorous evidence of efficacy — including any null or negative findings.
In partnership with our pilot schools. We will publish results here — including any null or negative findings. Transparency is the whole point.
Structured phonics, phonemic awareness, interleaved sight-word practice, progress-driven pacing. Lumi mirrors the scope-and-sequence of UFLI Foundations and CKLA — open programs with robust evidence behind them.
NWEA MAP, DIBELS — gold-standard literacy benchmarks.
Words correct per minute (once microphone ships in Phase 2).
Atoms reaching stable mastery per session-minute.
Percent of children who return for a session three or more days after first use.
Weekly self-rating of child's reading mood.
Short notes on what shipped recently and the design choices behind each release. The newest entries are at the top.
L17 ships fifteen 5-7 sentence nonfiction passages on cryptography, linguistics, logic, and economics, with multi-clue inference and evidence-based answer choices. L18 is the civic mirror — Roman republics, habeas corpus, the Constitution, vaccines, AI ethics — and tightens the focus on author's purpose. Both lean on stronger Latin and Greek roots (crypt, log, civ, demo, anthrop, philo). Lumi's reading band now reaches grade six.
L15 introduces 4-5 sentence passages spanning geography, ecology, and space, with multi-clue inference and academic vocabulary. L16 widens into cause/effect and sequence — food chains, glaciers, Edison's lab, internet packets. Greek and Latin roots (aqua, photo, mar, terr, struct, ject) become first-class. progress.ts already knew about L15-L18, so no schema changes were needed.
A voice-driven shape-naming game for ~6yo, gated to its own /games/shapes-club route so the play and math shells stay focused. Voice matching cascades from exact tokens through normalized Levenshtein and a tiny phonetic pass; after two voice misses, it silently flips into a three-option tap fallback. Audio reuses the existing ElevenLabs + Vercel Blob pipeline; SFX are WebAudio-synthesized so there are no asset files to ship.
M7 brings multi-digit addition and subtraction, area-as-array, perimeter counting, and missing-digit puzzles. M8 promotes facts to larger products (mult_fact_large), real division facts, factor questions, and skip-counting above ten. Each task UI was added alongside the new stimulus kinds, and the math registry's signature switch grew accordingly. Lumi's math band now covers K through grade four.
Seven composed rules pick the next task from the child's recent attempts: recover from misses, vary modality, refresh from the mastery interleave queue, and protect against streaks of the same atom. The new /api/child/pacing-context route hydrates the engine on session start; an in-session attempt ring buffer keeps it informed without a round-trip per task.
Comprehension becomes the first-class scope above grade three. L13 introduces Greek and Latin roots and a morphology tree task. L14 focuses on main idea, key details, and short inference across 4-sentence passages. Phonics retires; the TaskType union widened, ReadingLevelNum was extended to 18, and the play shell now switches task queues based on the level's category.