Enrollment Measures: DESR vs Classlist

CEDAR carries two independent enrollment counts that do not mean the same thing. This page is the full contract; AGENTS.md carries the short rule.

CEDAR carries two independent enrollment counts that do not mean the same thing. Comparing them — or comparing one measure across terms whose snapshots were taken at different points in the term — silently biases any demand/capacity analysis.

Measure	Column(s)	Source table	What it is
DESR enrolled	enrolled, available	cedar_sections (DESR snapshot)	Section headcount as of the file pull (as_of_date). Only one snapshot is retained per term (a newer DESR replaces the term’s rows — see R/data-parsers/parse-data.R).
Classlist registered	registered (from calc_cl_enrls, R/branches/enrl.R)	cedar_students	Distinct students in STATUS_REGISTERED (RE/RS/RR) — those still registered when the classlist was pulled.

Neither is a census freeze. The DESR CENSUS1/CENSUS2 fields are census dates, not counts (see DESR Input Schema). No census-frozen headcount exists anywhere in the raw data.

DESR snapshot timing varies by term (as_of_date vs term end), and this is the trap:

• Upcoming / current term — retained DESR was pulled during active registration, before census → enrolled = live, pre-drop count (≈ peak demand).
• Past terms — retained DESR was pulled after the term ended (backfills run hundreds–thousands of days post-end) → enrolled = final, post-drop count.

Empirically, DESR enrolled tracks the classlist registered (RE/RS/RR) bucket at whatever point the snapshot was taken (verified Fall 2024 and Fall 2026: per-course median ratio = 1.000). So a past term’s DESR enrolled ≈ classlist final headcount, while the current term’s DESR enrolled ≈ registered ≈ census (they coincide because no drops have happened yet).

How drops move each number

Reconstruct any lifecycle point from classlist status codes — all emitted by calc_cl_enrls (registered, dr_early, dr_late, dr_all):

Lifecycle point	Formula	Includes
First class day	first_day_enrl (calc_first_day_enrl())	registered rows dated on or before the section’s first class day, plus drops dated after it; NA where a group lacks status dates
Census / peak headcount	registered + dr_late	everyone present at census (only early drops DR removed)
Final / end-of-term headcount	registered	RE/RS/RR still enrolled
Total ever-registered	registered + dr_all	+ early drops

• Early drops (DR/DD, STATUS_DROP_EARLY) occur before the grade-consequence deadline → absent from both census and final counts (registration churn / melt). DD is treated like DR: a drop/delete with full tuition refund.
• Late drops (DG/DW, STATUS_DROP_LATE) occur after census → counted at census, gone from the final count. These are what make a course look less saturated at term end than it was at census.
• First day needs dates, not just statuses. first_day_enrl reads each row’s Registration Status Date (registration_date) against its section’s start_date. Coverage is all-or-nothing by pull: a class list exported without the status-date column gives NA for every course in that term until it is re-pulled. class_list_status_date_coverage(students) reports it per term (Admin > Data & Usage shows the undated terms). Banner keeps only the current status’s date, so a student who added after day one and later dropped counts as present. first_day_proxy (registered + dr_all, “Ever Registered Proxy”) is a different measure: it also counts students who dropped before classes began.
• One student, one bucket. calc_cl_enrls counts each student once per campus, college, term, and course. When a caller merges listings (a crosslist family counted under one code), a student can arrive with several rows, and the row kept follows STATUS_PRECEDENCE: registered, then late drop, then waitlisted, then early drop. Left to data order, students who dropped CS 375 and re-registered as MATH 375 in the same shared section were counted as drops, and Course Dynamics showed 58 and 59 for MATH 375 in Spring and Fall 2026 against DESR totals of 60 and 61.

Canonical census helpers (use these, don’t re-derive the formula): add_census_enrl(df) adds census_enrl = registered + dr_late. calc_census_enrl_baselines(df, target_terms, keys, prior_only) lives alongside it in R/branches/enrl.R. Regstats uses prior_only = TRUE: each course/term gets a mean, population SD, and count from strictly earlier matching terms. The default FALSE preserves the Waitlists all-history reference, excluding selected targets but potentially including later terms. Its count is reference terms, not necessarily prior terms. Both modes retain the full series for sparkline context. Neither reconstruction recovers a frozen census or peak-occupancy snapshot.

Regstats baseline contract (definition 3.0.0): group by course, delivery campus, college, season, and part of term. Enrollment, early/late-drop counts, and fill use the same prior-only population-SD policy via add_prior_history_stats() in R/trunk/history-stats.R. The target and later terms enter neither mean nor SD. Fewer than two prior observations or zero variation means unscored; show the coverage counts in baseline_info. Drop alerts remain count-selected operational volume screens; also expose early-drop rate over first_day_proxy and late-drop rate over census_enrl, with prior mean rates, so enrollment growth can be interpreted separately. Rates do not decide the flags. Saturation uses class-list census_enrl over DESR scheduled capacity, requires matched sources and usable capacity, and retains the enrollment-size floor. Its Full now badge is not an independent table-entry rule. Regstats and the Dept Dashboard that embeds it have versioned caches; bump both when these calculations change.

Fall 2024 magnitude (matched courses): 10,490 early drops (never in the DESR final) and 5,531 late drops. The late drops make census headcount ~5% higher than DESR enrolled (112,115 vs 107,808).

Consequence for saturation / capacity analysis

The Saturation report (R/features/regstats.R) computes fill_rate as class-list census_enrl / DESR capacity. This keeps the lifecycle numerator in one source; DESR enrollment remains visible only as independently timed snapshot context. The class-list numerator is reconstructed from final/current statuses rather than a frozen census roster, and capacity can have a different extract date. The result is not recovered peak occupancy or registration speed. Unmatched sources and unusable capacity are excluded and counted in baseline_info. The current fill threshold is user-controlled (90% by default), not a fixed institutional capacity standard.

To compare like-for-like occupancy, derive fill rate from classlist headcounts at a single explicit lifecycle point rather than the DESR enrolled snapshot. Census registered + dr_late is the right denominator for census occupancy and attrition reporting, but not for deciding whether a course was full: census is the 15th day of instruction, and three weeks of drops separate it from the first day (the term calendar is stated once, in AGENTS.md’s Enrollment Measures section). Enrollment projections (0.21.0) count capacity reached at first_day_enrl / scheduled_capacity >= 0.98, with first_day_enrl from calc_first_day_enrl(); over-cap overrides still count as reached. The ever-registered classlist_total is not an occupancy at all, because seats that turned over were never all held at once. Where the class lists carry no status dates there is no first-day count and the term is not judged. Reporting class-list registrations, first-day, census, and final enrollment together keeps those lifecycle questions separate.