Every algorithmic system carries an intellectual genealogy, whether a partner ever asks about it. For a scoring pipeline sitting inside a public-sector behavioral health contract, that genealogy is not incidental biography — it is evidence. Reviewers assessing a machine learning system for defensibility want to know two things that get conflated in casual descriptions of "the technology": who built the statistical judgment embedded in the model, and where the measurement philosophy behind the model's design choices actually comes from. This section answers both, tracing IMPACTER's scoring architecture back to a specific, checkable research lineage rather than treating the pipeline as a black box that arrived fully formed. That lineage runs in five distinguishable threads, and this section keeps them separate on purpose rather than blending them into one undifferentiated "the research supports us" claim: who architected the scoring pipeline and what causal-inference toolkit they bring to it; how that pipeline is kept honest in production between formal reviews; what the current construct-coding literature says about the specific architectural choice IMPACTER made; why the prompts ask students to reflect rather than rate themselves; and, stated as its own explicit boundary rather than folded quietly into the others, what a piece of mouse neuroscience is and is not doing in this citation set.
IMPACTER's Senior Scientific Advisor, Dr. Andrew Arnold, is the architect of the platform's scoring pipeline, per the organization's published Model Card. His public professional record runs through machine learning research and applied engineering at a scale relevant to a system like BHPM's: Principal Machine Learning Engineer at Shopify, former Chief Scientist for AWS CodeWhisperer, and a PhD in Machine Learning from Carnegie Mellon University. Before those industry roles, his academic work concentrated on a narrower and more technical problem than "machine learning" as a category suggests: recovering causal structure from time-series data. His 2007 paper with Yan Liu and Naoki Abe, presented at the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, introduced graphical Granger-causality methods for multivariate time series — formal techniques for asking, given several variables observed over time, which ones plausibly drive which. The paper's central contribution, the Lasso-Granger algorithm, uses Lasso-based variable selection to recover sparse causal graphs among many time-varying variables simultaneously, and the authors characterize the relative performance of a family of graphical Granger approaches from multiple angles, showing that the Lasso-based method gives consistent gains over the standard pairwise graphical Granger approach for that recovery task.
That paper does not stand on its own; it extends a specific, older methodological lineage rather than inventing the underlying idea of causal inference from scratch, and naming that lineage is part of what makes the credential checkable rather than decorative. The bivariate notion of Granger causality traces to Clive Granger's 1969 paper in Econometrica, which proposed testable, operational definitions of causality and feedback between two time series — showing that the cross-spectrum between them decomposes into components attributable to distinct causal arms, from which a measure of causal lag and causal strength could be constructed . Granger's original framework handled one pair of variables at a time. What Arnold, Liu, and Abe's 2007 paper adds is the move from that bivariate setting to a multivariate, graph-structured one: instead of asking whether variable A Granger-causes variable B in isolation, Lasso-Granger asks which sparse subset of many candidate variables plausibly drives each other variable, using variable-selection machinery to keep the resulting causal graph interpretable rather than a dense, uninformative tangle of pairwise arrows. This is a specific, load-bearing kind of expertise for a program like BHPM, and the fit is not decorative. Activities under this contract will eventually need to ask a harder question than whether an indicator moved. They will need to ask whether a given upstream lever — a training rollout, a change in screening cadence, a shift in service dosage — plausibly caused a downstream change in a student outcome trajectory, as opposed to merely preceding it on a calendar, across a system with many moving indicators rather than one clean pair of variables. That is precisely the class of problem graphical Granger methods were built to address, and it is the planned analytic foundation for the Year 2 upstream-lever and dosage analysis IMPACTER intends to run jointly with Stanford Medicine once the indicator set stabilizes. The KDD 2007 paper and the 1969 paper behind it are cited here for that reason: not as evidence about the behavioral health screener itself, but as the methodological throughline connecting the pipeline's architect, through a named and dated origin point in the causal-inference literature, to the specific statistical toolkit this contract will eventually need for causal, not merely correlational, questions.
The production expression of that same discipline shows up closer to the ground, in how IMPACTER keeps its scoring models honest between formal review cycles, and the mechanics of that process are more specific than a one-line description of "human review" suggests. The organization's internal technical note on the scoring pipeline's calibration layer describes a weekly, stratified, human-in-the-loop sampling process built around four independent stratification factors rather than a single random draw. Competency is sampled on its own axis so that no rubric-aligned construct goes more than one cycle without review, because agreement statistics behave differently across constructs of varying operational clarity and aggregating across them can mask construct-specific drift — the same lesson the construct-coding literature discussed below documents from a different angle. Predicted score band is a second axis, and it is deliberately not sampled uniformly: the clearest low- and high-evidence responses tend to be the most stable part of any rubric, while the middle bands are where neighboring levels are hardest to distinguish and where calibration errors are most likely to hide, so the sampling layer intentionally over-represents those middle bands rather than letting easy extreme cases dominate the review set and inflate agreement statistics. Course or prompt family is a third axis, protecting against prompt-level overfitting and ensuring newly authored content gets reviewed within one sampling cycle of release rather than waiting for a future audit to notice it. Item and response type — voice-elicited versus text-elicited, short-form versus long-form — is the fourth, because automatic speech recognition introduces a known source of input-level noise the technical note treats as something to represent explicitly in the validation distribution rather than average away. Responses selected this way surface inside a reviewer-facing interface where an expert either confirms or overrides the model's predicted score against rubric anchor language, and every re-scored item is written back with full provenance — the original prompt, the model version that produced the score, the reviewer's rating, and a timestamp — so that a later disagreement can be traced to a model-version change, a human re-rating, or genuine rubric ambiguity rather than treated as unexplained noise. Re-scored items then flow automatically into two places: a rolling recomputation of agreement statistics against the most recent human ratings, with movement beyond a pre-registered degradation threshold flagging a competency for targeted retraining or rubric re-examination, and the training partition for the next scheduled fine-tune. Model promotion itself stays gated the same way regardless of how the new training data was sourced: a candidate model is only promoted to production if its held-out agreement with human raters improves and clears the organization's standing thresholds, and once promoted it is frozen for deterministic inference . This is what human-in-the-loop review means in practice for this system: not a one-time human check performed before launch, but a standing sampling design that keeps expert raters continuously embedded in what the model is permitted to learn next, with the sampling logic itself doing real methodological work rather than functioning as a compliance gesture.
That same commitment to construct clarity over raw model scale connects IMPACTER's architecture to a second, independent research thread with its own institutional pedigree, and this is the thread where the design choice and the evidence for it line up most directly. Ober, Zhang, Zapata-Rivera, Schroeder, and Botelho's 2026 study in the Journal of Educational Data Mining compared multiple ChatGPT-4o configurations, varying both temperature and model size, against expert human coders on more than 10,000 conversational turns from 107 English learners . The coding task spanned several constructs drawn from a conversation-based math assessment with a pedagogical agent — persistence, self-efficacy, interest, perceived challenge, and prior knowledge — and the study's central finding was not that a better or larger model solves construct coding. It was that how precisely a construct is operationally defined governs coding reliability far more than which model, or what model size, is doing the coding. Self-efficacy, the most operationally well-specified construct in the study, showed the strongest alignment between human and machine coders; more theoretically diffuse constructs like persistence did not, and reliability for those diffuse constructs varied unpredictably depending on the temperature setting used, in a pattern the authors could not fully explain by construct alone. The paper documents genuine non-determinism and reproducibility problems when generative language models are asked to code the same psychological construct twice under different sampling conditions — the same instability, arrived at from a different research tradition, that Turpin, Michael, Perez, and Bowman documented in generative chain-of-thought reasoning more broadly, where models can produce fluent explanations that systematically misrepresent the actual basis for their own output . That finding sits directly behind an architectural decision IMPACTER made and has not revisited: its production scoring models are discriminative and ordinal, not generative. A fine-tuned, 184-million-parameter DeBERTa-v3-base encoder paired with a CORAL ordinal-regression head produces one deterministic score per response under a frozen, versioned model, evaluated across the eight competencies at an average quadratic-weighted kappa of 0.918, 77.3% exact-match agreement with human raters, and 98.0% adjacent agreement . It does not sample text on inference, and it does not produce a different judgment on a second pass over the same response. The instability Ober and colleagues document, and that Turpin and colleagues document in a related setting, is a property of asking a generative model to code a construct by sampling; it is not a property IMPACTER's scoring architecture inherits, because the architecture was built specifically to avoid that failure mode by never sampling in the first place. The organization's partner-facing guide to the screening approach describes the same design choice in plainer terms, and frames it against a broader pattern in validated assessment rather than as a bespoke IMPACTER invention: students respond to open-ended prompts in authentic voice, and a rubric-aligned scoring engine trained against expert human raters produces the developmental score that routes a student within a multi-tiered system . The guide's own framing is worth carrying forward here because it locates IMPACTER inside a familiar category rather than outside it: assessments that standardize the rubric rather than the prompt — AP exams, NAEP writing, ETS performance items, Smarter Balanced constructed-response components — accept that different respondents see different stimuli and place reliability in the scoring system instead, exactly the pattern IMPACTER's architecture instantiates for behavioral health screening.
Traced end to end rather than argued paragraph by paragraph, that causal-inference lineage forms a single visual line: a bivariate econometric framework in 1969, its multivariate extension at KDD in 2007, the production scoring pipeline that same discipline now leads, and the Year 2 causal analysis it is built to eventually support.
Research lineage — from bivariate econometric causality to IMPACTER’s production scoring pipeline and the planned Year 2 causal analysis
Bivariate Granger causality
Lasso-Granger: multivariate graphical Granger
IMPACTER Model Suite v6.0
Stanford Medicine upstream-lever & dosage analysis
Bivariate Granger causality
Granger · Econometrica
Proposes testable, operational definitions of causality and feedback between two time series, decomposing their cross-spectrum into causal-lag and causal-strength measures.
Lasso-Granger: multivariate graphical Granger
Arnold, Liu & Abe · ACM SIGKDD (KDD)
Extends Granger's bivariate framework to many variables at once, using Lasso-based variable selection to recover sparse causal graphs. First author Dr. Andrew Arnold is now IMPACTER's Senior Scientific Advisor.
IMPACTER Model Suite v6.0
Production scoring pipeline · Model Card
DeBERTa-v3-base (184M params) + CORAL ordinal-regression head, architected under the same causal-inference discipline. Avg QWK 0.918 · 77.3% exact-match · 98.0% adjacent agreement across eight competencies.
Stanford Medicine upstream-lever & dosage analysis
Joint with Stanford Medicine, once the indicator set stabilizes
The same multivariate causal-inference toolkit applied to ask whether an upstream lever — a training rollout, a screening-cadence change, a service-dosage shift — plausibly caused a downstream change in a student outcome trajectory.
Two further citations round out this lineage, and the framing on both is deliberately narrow, by design. Hirshberg, Colaianne, and colleagues' multi-university study of a for-credit undergraduate flourishing course found that students enrolled in a course pairing academic content with structured experiential and reflective practice showed improved flourishing and mental-health outcomes relative to comparison students who were not enrolled . That study did not originate the idea that structured reflection does real psychological work rather than merely delivering information; it extends a research line that runs back at least to Pennebaker and Beall's 1986 study establishing what later became known as the expressive-writing paradigm, in which students who wrote about a personally traumatic event on four consecutive days, rather than about a trivial topic, showed a short-term rise in physiological arousal and negative mood followed by measurably fewer health-center visits over the following six months . The specific finding in each of these two papers is different — one is about writing and subsequent physical-health utilization in a general undergraduate population, the other about a structured course and flourishing outcomes — but the throughline connecting them is the same premise Hirshberg's study tests directly and IMPACTER's prompt design assumes: that reflecting on one's own experience in one's own words is itself a mechanism, not simply a format for delivering a score to a waiting rubric. That premise belongs in this section as design rationale, not as validity evidence for a screener: it is part of why IMPACTER's prompts ask students to reflect in their own words rather than select from a fixed checklist, on the premise that structured reflection is itself a meaningful unit of measurement.
The second citation requires an explicit boundary, stated plainly rather than implied. Saab and colleagues' 2009 study in Neuron found that inducible overexpression of a calcium-sensor protein, NCS-1, in the dentate gyrus of adult mice selectively increased exploratory behavior alongside facilitated synaptic plasticity — including facilitated long-term potentiation — and faster spatial-memory acquisition, an effect reversible with a peptide that blocked the protein's binding to the dopamine D2 receptor . The manipulation was inducible and its effect was pharmacologically reversible, which is what let the authors attribute the behavioral change specifically to that protein interaction rather than to some earlier developmental artifact of the mouse's genetics — a methodologically careful study, on its own terms. But its own terms are mouse neurobiology. It appears in IMPACTER's research lineage as founding inspiration for taking seriously the biological basis of exploratory behavior as a construct worth measuring in young people at all — nothing more. It is not evidence that the screener is valid, it is not evidence about human behavioral health outcomes, and it should not be read as either. Mouse neuroscience does not validate a human skills measurement model, and this section states that distinction directly so that no reader has to infer it.
Set side by side, what these seven citations describe is an organization whose measurement choices trace back to identifiable, checkable sources rather than internal intuition alone: a causal-inference researcher, credentialed by a named and dated methodological lineage running from Granger's 1969 bivariate framework through his own 2007 multivariate extension, leading the scoring architecture; a documented weekly human-calibration process, stratified across four independent factors rather than a one-time validation event; a published construct-coding literature that independently supports the deterministic-over-generative design choice IMPACTER already made, and that IMPACTER's own architecture happens to sidestep by construction rather than by claim; a reflective-practice research base running from a 1986 expressive-writing study through a 2022 multi-university flourishing course behind the prompt design itself; and one explicitly bounded piece of mouse neuroscience that this section refuses to let drift into a claim it was never built to support. The organization's own plain-language guide to its machine learning model states the operating principle underneath all of it directly: human judgment, expressed through rubrics and expert raters, leads the system, and machine learning scales that judgment rather than replacing it . That is the claim this section exists to support with citations rather than assertion.
Bottom Line
IMPACTER's scoring architecture is a lineage, not a black box. A named causal-inference researcher, credentialed by a specific and checkable methodological line rather than a generic "machine learning" bio, leads it as Senior Scientific Advisor and pipeline architect. A documented weekly human-in-the-loop calibration process stratifies review across construct, score band, prompt, and item type. And the scoring design itself is discriminative, deterministic, and non-generative, a choice the current construct-coding literature (Ober et al., 2026) directly supports over generative alternatives. That lineage, not the model card's accuracy figures on their own, is what makes the pipeline trustworthy. One boundary holds regardless of how this framing gets reused, edited, or reorganized elsewhere: the Saab et al. (2009) citation is founding neurobiological inspiration only, never screener validity evidence, a distinction load-bearing for how a clinical or research reviewer reads the whole citation set.
