The Reading Alignment Index for AI Literacy Systems. A purpose-built evaluation instrument for scoring a single AI-generated output against the Science of Reading.
Every existing literacy evaluation framework — EdReports, What Works Clearinghouse, the Knowledge Matters review, the Reading League curriculum evaluation tool — was designed to evaluate a curriculum or a program. None of them were designed to evaluate a single lesson plan generated by a large language model in response to a teacher prompt. That is the unit of analysis RAILS exists to measure, because that is the unit of instruction AI tools are actually producing at scale inside classrooms today.
Six indicators. The decoding side of the Simple View of Reading.
Two indicators. The language half of Scarborough’s Rope.
One indicator. Strategies-as-ends versus knowledge-building.
Two indicators. Explicit instruction and encoding connected to phonics.
Two indicators. The most durable assessment practices the research has moved against.
Every indicator is tiered. Strong Consensus rests on converging meta-analyses. Emerging Consensus is reported with hedging. Five indicators are flagged Critical — the practices most likely to cause harm when dominant.
Binary scoring loses signal. An AI output that mentions three-cueing as one option among several is meaningfully different from one that recommends three-cueing as the only approach, and RAILS separates those cases explicitly.
MagicSchool AI, Khanmigo, Diffit, Curipod, Brisk Teaching. Products with system prompts, templates, and editorial decisions layered on top of base models.
ChatGPT (GPT-4o), Claude Sonnet, Gemini, Llama. General-purpose AI models responding to raw prompts with no education-specific product layer.
Reporting the two cohorts separately matters. A finding about Cohort A describes what a product delivers. A finding about Cohort B describes what a base model defaults to. Conflating them obscures where the evaluation gap actually lives.
If fewer than twenty percent of outputs across all tools have a median severity of 2 or higher on any Critical-tier indicator across three runs, the thesis that AI tools systematically reproduce debunked literacy practices is not supported by this data.
That threshold was set before data collection began. The study is falsifiable by design. The book does not survive the data failing it, and the instrument does not survive the thesis being wrong.