Four hypotheses about the structure of introductory CS.

Question 1 · Project P1

Do structural features of course design predict help-seeking suppression in LMS behavior?

Construct. Help-seeking suppression — the behavioral pattern in which students who are objectively stuck (long edit-compile-fail cycles, rising error rates, declining progress) do not post in the discussion forum, attend office hours, or reach out on Slack. Predicted by Seymour & Hunter to be structurally caused and by the Karabenick / Newman help-seeking literature to track feature-level design choices.

Operationalization. A behavioral index built from LMS and discussion-forum logs: stuck-ratio = (long idle stretches within an edit-fail cycle) / (total idle stretches). Course-design features coded from syllabus, assignment prompts, and office-hours listings using the Tool-2 rubric (validated against P2). DFW outcomes from institutional records.

Hypothesis H1. In a sample of at least 6 CC CS course-sections with ≥1200 enrolled students, course-level pedagogical debt score (composite of the four Tool-2 dimensions) will positively predict course-level help-seeking suppression rate, controlling for enrollment, instructor, term, and student prior-grade covariates, with β ≥ 0.2 and p < 0.05.

Disconfirming null. If the regression returns |β| < 0.1 or p ≥ 0.1 on a pre-registered specification, I accept that the structural-feature hypothesis is not supported at this level of measurement and report the negative result.

Methods: learning analytics · LMS log mining · regression · pre-registration

Question 2 · Project P4

Does MVC-distance in an instructor-annotated curriculum graph predict DFW rates and persistence?

Construct. MVC-distance — the gap between a real curriculum and its greedy Minimum Viable Curriculum on the typed-dependency graph of learning objectives (see Tool 3). The research claim is that curricula farther from the greedy MVC carry "redundant load" that falls disproportionately on students with the least academic slack.

Operationalization. Ten introductory CS course sequences at California community colleges, each annotated by two instructors blind to the other's annotations, against a fixed 60-item objectives vocabulary. MVC-distance = (number of course modules in current sequence) − (greedy-MVC module count on the annotated graph). Outcomes = section-level DFW rate and one-year persistence from institutional data.

Hypothesis H2. Across the 10 sequences, Spearman correlation between MVC-distance and DFW rate ≥ 0.45, and inter-annotator Cohen's κ on objective labels ≥ 0.65.

Disconfirming null. If |ρ| < 0.2, or if κ falls below 0.55 (indicating the typed-dependency grammar is not reliably coded even by experts), I publish the negative result and retire MVC-distance as a candidate metric.

Methods: curriculum graph annotation · inter-rater reliability · institutional data matching

Question 3 · Project P3

Do the Seymour & Hunter departure categories replicate at community colleges, and what is missed?

Construct. Structural departure — the claim that STEM attrition at 2-year institutions is driven, as it is at 4-year institutions, by teaching quality, weed-out culture, eroded belonging, and help-seeking suppression, rather than by aptitude. The CC-specific twist is what Seymour & Hunter could not measure: transfer-path fracture, financial-stop-out, caregiving labor, and the interaction of those with course structure.

Operationalization. Semi-structured interviews with 24–30 students who enrolled in an intro CS course at Foothill or De Anza in the 2024–2025 academic year and did not re-enroll in the subsequent CS course. Transcripts double-coded against the S&H taxonomy plus a pre-registered set of CC-specific candidate categories (transfer-stop, caregiving-load, advising-gap). IRB submission Spring 2026, fieldwork Summer 2026.

Hypothesis H3. ≥ 70% of coded departure episodes map onto at least one S&H category; additionally, ≥ 20% of episodes require at least one CC-specific category to be fully described. Inter-coder κ on the combined codebook ≥ 0.70.

Disconfirming null. If the S&H taxonomy accounts for <50% of episodes, or if the CC-specific extensions are not reliably coded, I revise the theoretical frame before publication rather than squeezing the data to fit.

Methods: semi-structured interviews · grounded theory · double coding · IRB-reviewed study

Question 4 · Project P2

Can rule-based text features reach expert-level reliability for detecting pedagogical debt in syllabi?

Construct. Pedagogical debt — a composite of motivational framing, scaffolding visibility, verification structure, and belonging signals, operationalized in Tool 2. The research claim is that the construct is measurable from syllabus text alone at reliability high enough to be useful to instructors.

Operationalization. A corpus of 120 introductory CS syllabi from California community colleges (IRB-cleared, public-domain where possible). Three trained annotators score each syllabus against a rubric derived from the Tool-2 rules. Rule-based Tool-2 scores compared against the mean human score per dimension.

Hypothesis H4. (a) Inter-annotator Cohen's κ ≥ 0.65 on each of the four dimensions. (b) Spearman correlation between the Tool-2 rule-based score and the mean human score ≥ 0.70 on each dimension. (c) Tool-2 scores, entered as a single composite, predict end-of-term student belonging scores (Walton & Cohen instrument) in a 6-section subsample with β ≥ 0.15, p < 0.05.

Disconfirming null. If any of the three reliability thresholds fail, the tool is not validated; I report which dimensions failed, what their rules looked like, and what the failure tells us about the measurability of the underlying construct.

Methods: annotation study · NLP rule evaluation · belonging-instrument administration