The first in a series introducing HMH Pulse, the learning science behind HMH Performance Suite®
Walk into any classroom on a Tuesday morning and you’ll see what every district leader already knows: tier 1 instruction is where the real work of teaching and learning happens. It’s also where any gap between data and decision-making is felt most acutely.
A teacher preparing a fourth-grade lesson on area and perimeter could, in principle, consult interim assessment results, formative checks from the most recent unit, the standards documentation for her state, the publisher’s resource library, and her own grade book. In practice, synthesizing those sources before class begins is unreasonable to expect of any educator. Decades of formative-assessment research point to the same conclusion: assessment evidence becomes useful when it is specific, timely, and clearly tied to next instructional moves. When it isn’t, the data sits unused, no matter how much of it there is.
The persistent gap between assessment and action isn’t a problem of educator effort. It’s a problem of representation and inference. Closing it requires something no standalone tool in a district’s stack can deliver alone: a structured representation of what students learn, a principled method for inferring what each student knows, and a way to translate those inferences into the specific decisions teachers face in the next planning window. That combination is what Performance Suite, powered by Pulse, was built to provide.
What Pulse is
Pulse is HMH’s learning sciences intelligence platform. It sits at the foundation of Performance Suite and serves as the underlying layer to provide customization for HMH curriculum solutions. It is built on three connected components, each grounded in a mature scientific tradition and implemented with peer-reviewed methods:
- An enterprise ontology, a structured map of 967 instructional skills, 28,958 standards from all 50 U.S. states, and the assessment items and instructional resources used to teach and measure them.
- A proficiency model, calibrated against millions of student responses, that infers what a student knows from the work the student has actually done.
- A cascading inference layer that turns those inferences into the specific recommendations and groupings educators see in the product.
Together, these components are what allow Pulse to take the irregular trail of work a student produces and turn it into something a teacher can act on the next day.
The ontology: A structured map of what students learn
Before Pulse can infer anything about a student, it has to know how the content fits together. That’s the job of the ontology.
Pulse’s ontology brings together three things every educator already works with: the standards adopted in their state, the skills those standards reference, and the instructional resources and assessment items used to teach and measure each skill. State standards come first: 28,958 standards from all 50 U.S. states across grades K–12 (15,642 in mathematics; 13,316 in English language arts), so similar standards across states can show overlap but remain distinguishable when a district needs them to be.
The skills come next: 554 skills in mathematics, organized into eleven clusters spanning grades K–8 (whole-number understanding and arithmetic, fractions, decimals, rational-number understanding and arithmetic, ratios and proportions, expressions and equations, measurement, two- and three-dimensional geometry, and data analysis), and 413 skills in ELA, with two reading-comprehension clusters in production today and additional clusters for vocabulary, grammar, and word study under active development. The instructional content follows: tens of thousands of assessment items from MAP Growth and HMH curriculum-embedded assessments, each tagged to the skills it measures.
The practical payoff is that Pulse can reason uniformly across HMH’s curriculum and assessment products, regardless of which state’s standards a school uses or which edition of a curriculum a district has adopted. We’ll go deeper on the ontology in the next post in this series. For now, the headline: this is what gives every Pulse recommendation a coherent map underneath it.
The proficiency model: Clear claims about what a student can do
A proficiency model is the research-grounded answer to two questions every educator already asks: what does it mean to know this skill, and what evidence would tell us a student has it?
In Pulse, skill-level claims are binary. A student is classified as either proficient at a skill or not yet proficient. Proficient means the student can engage reliably with problems that require the skill. What matters isn’t how many items a student answered correctly on a test; it’s which skills the student has actually mastered. That is the difference between a tool reporting “student got 6 out of 10 correct” and a tool that can say, with confidence, “this student is proficient at equal-groups multiplication and not yet proficient at multiplicative comparison.” A complete picture of a learner emerges from many of those skill-level claims taken together, alongside continuous proficiency estimates at the cluster and subject levels that support reporting and growth analysis over time.
Behind each of those claims is a peer-reviewed psychometric framework with more than two decades of methodological development: the diagnostic classification model, or DCM. DCMs are designed for exactly this task, classifying students into profiles defined by discrete skills, on the basis of item responses connected to those skills. Pulse uses the higher-order variant of the framework, an established extension that captures how skills relate to one another within a cluster, so evidence about one skill informs inferences about nearby related skills.
Calibrating a model of this size is not a routine application of off-the-shelf software. The challenge motivated two strands of methodological work by HMH’s learning sciences engineering team, both accepted for publication in flagship peer-reviewed psychometrics journals: a modular estimation framework in Multivariate Behavioral Research, and related work on high-dimensional diagnostic classification modes in the British Journal of Mathematical and Statistical Psychology. The approach running in production builds on the logic of both papers, with additional modifications developed to handle the full dimensionality at scale. Districts and partners who want methodological transparency have a citable reference.
One picture, built from many assessments
Most districts don’t run on a single assessment. They run on a portfolio: a universal screener in the fall, benchmark or interim assessments through the year, classroom formatives every week, and end-of-unit summative tests. Each was designed for a different purpose. Each produces its own report, in its own vocabulary.
Pulse takes a deliberate stance on how to make use of that portfolio: evidence flows skills-up, not scores-down. It builds understanding the way a mosaic comes together, one tile at a time, each placed precisely where it belongs. Every item response a student produces, whether on MAP Growth, a curriculum-embedded check, or a unit test, is a tile in the mosaic. Close up, each tile is just a tile. Step back, and the picture of the student and their learning appears, with every piece still traceable to its source. Other approaches try to blend finished paintings together and call the result a portrait. Pulse is purpose-built to render the picture from the tiles themselves.
Because the proficiency model describes the skill itself, not the assessment, item responses from any source can update the same coherent picture of student learning. Information propagates skill-to-skill within a cluster, cluster-to-cluster within a subject, and over time as new evidence arrives. The proficiency model is the constant. The assessments are inputs. The result is one trustworthy view of where each student stands, and what to do about it.
None of this is glamorous. All of it is what separates a system you can stake a curriculum decision on from a system that simply produces confident-sounding output.
How we know you can trust it
Trust in an instructional engine depends on how it's tested, not just how it performs. Pulse is evaluated under a protocol that mirrors real classroom use: the students used to test the model contributed nothing to building it. The model meets them as strangers, estimates their proficiency from their first responses, and then has to predict how they'll perform on items it hasn't seen — the same task it faces on the first day of school with any new student. Those predictions are scored against what students actually do. Strong results under this protocol mean strong results in production, not strong results in the lab where the model has already seen the answers.
A predictive model has to do two things well: rank students accurately (AUC) and produce probabilities that hold up in reality (Brier score). The numbers below show Pulse clearing both bars by a meaningful margin.
Across the eleven mathematics clusters and two ELA clusters currently in production, Pulse achieves:
- Predictive AUC between 0.74 and 0.79, well above the 0.70 threshold the field considers good for educational prediction.
- Calibration error stays below 0.20 (Brier score) in every cluster, meaning that when Pulse says a student has a 70% chance of getting an item right, that prediction holds up at roughly 70% in the actual data.
- The weighted percent-correct baseline that dominates the K–12 assessment platform market lands at 0.57 to 0.68 AUC across the same clusters, with substantially worse calibration.
The most important result is what happens at first contact. With as few as one item observed in the cluster being predicted (and for students whose prior evidence comes entirely from other clusters), Pulse is already producing strong predictions, above the threshold the field considers good for educational prediction (AUC between 0.72 and 0.77).
Pulse performs equitably regardless of whether there is small sample or an extensive assessment history for students. That cold-start strength is the practical payoff of the multidimensional model: evidence from anywhere in the ontology informs skill-level inferences in clusters where the student hasn’t yet been directly assessed. In other words, Pulse can make useful inferences about a student from day one because evidence about related skills helps it estimate skills it hasn't directly seen yet.
Validation does not end at launch. Models are continuously monitored, recalibrated as new evidence accumulates, and evaluated against the standards published by the American Educational Research Association, the American Psychological Association, and the National Council on Measurement in Education. Differential performance across student subgroups is treated as a first-class concern. None of this is glamorous. All of it is what separates a system you can stake a curriculum decision on from a system that simply produces confident-sounding output.
One principle is worth naming explicitly: educators decide; Pulse informs. Pulse does not place students into instructional tracks or make autonomous decisions about students. It surfaces probability estimates, recommendations, and group suggestions for educators to act on, modify, or override. Every product surface that uses Pulse is designed to keep the human in the loop.
Math and ELA aren’t the same, and that’s the point
Mathematics and literacy are fundamentally different learning experiences. Math builds in conceptual progressions; certain prerequisites need to be in place before later ideas click. Reading and writing develop along multiple dimensions at once: decoding, language comprehension, knowledge building, and writing skills all grow together, sometimes unevenly and in interaction, as texts and tasks increase in complexity.
Pulse’s ontology and inference logic are shaped to match. Math signals get read through a model that understands prerequisite relationships and the difference between procedural and conceptual mastery. ELA signals get read through a model built around the multidimensional nature of reading and writing development. Posts five and six in this series will go deeper into each.
Here’s the part that matters for a teacher in a classroom: the experience is the same. Whether the subject is math or ELA, what shows up is a clear, real-time view of what students know, what they need next, and what specific resources from their curriculum will move learning forward. The science differs because the disciplines differ. The promise to the educator does not.
What’s ahead in this series
This article is the first in a series. Over the next five posts, we’ll open up the engine and look inside.
- Post 2: What an ontology is, and why it’s the foundation of trustworthy recommendations.
- Post 3: How Pulse’s proficiency models turn assessment data into meaning teachers can act on.
- Post 4: How insights become specific, classroom-ready instructional recommendations.
- Post 5: The science behind mathematical understanding, and what makes math recommendations distinct.
- Post 6: The science behind reading and writing development, and what makes ELA recommendations distinct.
Our goal with Pulse and with this series is simple: to show district and school leaders, and the teachers and coaches who work alongside them, exactly how Performance Suite, powered by Pulse, turns assessment results into the kind of insight that gives teachers time back to teach, starting on day one and not stopping until day 180.
***
Turn data into timely, targeted instruction for every student with our free guide.
Related Reading
Aja Cormier, an honors English teacher at Ball High School in Galveston, Texas, confers with a student drafting a persuasive essay using the AI-powered Writable program.