Enrollment Projection Architecture

CEDAR’s enrollment projections are a reusable, non-UI computation and saved-data contract. The first pass projects only an explicit course scope, evaluates every method with rolling-origin aftcasts, and publishes one versioned bundle for Shiny and other CEDAR features to read.

The computation does not run in a module and does not read global data. This is deliberate: method development and artifact publication must remain testable in a vanilla R process without loading or restarting Shiny.

Production refresh

The morning refresh automatically checks projection freshness after successful data transformation. config/enrollment-projections.yml defaults to the next Spring after the settled enrollment edge, with that edge as the historical cutoff and critical_courses as the scope. It can pin explicit terms or disable automatic builds; the process never guesses an edge when data timing is missing.

The feature builder prepares the same inputs used by the analyses, normalizes row ordering, and compares their content hash, model-source hashes, configuration, and scope with the saved bundle. Missing, invalid, legacy-without-freshness, or changed bundles rebuild. Unchanged bundles skip fitting and are not rewritten. File timestamps and raw extract dates do not themselves cause rebuilds. The prepared inputs include upstream populations, target registrations, and scheduled capacity, so changes to those can legitimately refresh the saved output too. This is a data preparation check, not a zero-cost check, but it runs outside Shiny and does not refit or aftcast unchanged models.

An optional output/projections/rebuild-request.yml still forces one build with explicit terms and scope and is cleared on success. The following refresh returns to the configured policy; pin terms in the policy for a continuing override. Failures preserve the old artifact and are reported to the scheduler; the next successful data refresh retries automatically.

Deploy-time model gate

Shiny never fits a model, so a bundle saved by older code is not replaced just because new code shipped. A schema or scoring change would otherwise leave the Projections page on its empty state until the next morning refresh, or until someone remembered to rebuild by hand.

.github/workflows/deploy.yml closes that gap. After the health check passes it runs scripts/build-enrollment-projections.R --refresh-if-model-changed, which calls enrollment_projection_model_drift() to compare the saved bundle’s recorded schema version, model version, and per-file model source hashes against the deployed source. It loads no CEDAR table: a no-change deploy costs about a second, against roughly four minutes for a real rebuild, and most deploys touch no projection source at all.

The gate owns model staleness only. Data-driven staleness stays with the morning refresh, which still runs the full signature comparison including inputs_sha256 over the prepared inputs. Anything unreadable, invalid, or predating provenance tracking counts as drift — failing towards a rebuild is correct, because the alternative is serving a page built by unknown code.

It never fails the deploy. The app is already serving by that point; a failure leaves the previous bundle in place, is reported in the deploy log, and is retried by the next morning refresh. A concurrent morning refresh holds the build lock, so the deploy’s attempt exits with that message rather than racing it.

Because the gate reads the hashes of every file in enrollment_projection_model_source_files(), editing any of them — including the refresh feature itself — correctly triggers one rebuild on the next deploy.

Use scripts/restart-cedar.sh --update for the morning refresh plus app reload. The builder runs outside Shiny with a temporary writable output mount; normal Shiny startup continues to consume saved bundles only. Rebuild for a new target or cutoff, corrected historical inputs, or changed model/policy/scope. Retain deliberately published official vintages under output/projections/vintages/ with unique dated names and production backups; only *-latest.qs is replaceable. See the scripts guide for the policy, force-rebuild, immediate-check, and retry commands.

Grain and Scope

Every result is keyed by:

market_id + subject_course + term_type + target_term

college and department remain descriptive metadata, not projection keys. The named abq_ea_course_market is a deliberate exception to the normal campus grain: ABQ and EA are treated as substitutable deliveries because online enrollment is strongly governed by the seats offered. The market class list counts each student-course once across both campuses and total capacity is pooled.

The underlying delivery rows are not discarded. delivery_components remains keyed by:

market_id + campus + college + subject_course + part_term + target_term

Those rows describe where seats are allocated and the prior comparable census/capacity. They are evidence about delivery mix, not independent campus-demand forecasts. Branch campuses never enter this market.

A named course group is a monitoring scope, declared in the institution’s projection_groups.csv and projection_courses.csv. UNM’s critical_courses group contains the canonical Gen Ed list (include_gen_ed) plus the explicitly monitored FYEX and gateway courses. Its always-monitored core is FYEX 1010, 1030, and 1110; MATH 1215, 1220, and 1350; BIOL 1140; CHEM 1215 and 1215L; and ENGL 1110 and 1120. CHEM 1220 is not in the monitoring scope. The group is restricted to ABQ and EA. Course membership uses exact subject_course values; suffix variants are never included by prefix.

There is no implicit user scope because CEDAR has no login identity. A caller must supply courses, campuses, and a named market, directly or through a course group. Future department controls must resolve department codes through cedar_lookups$subject_lookup before calling the feature builder.

The inexpensive pressure screen runs before student-level methods. It can retain courses because they are always monitored, have a historical capacity shortfall, show repeated high class-list registration fill, or have growth beyond scheduled capacity. The always-monitored core proceeds to student-level methods and aftcasts on every build; the remainder of the broad Gen Ed scope must pass the pressure screen. An explicit ad hoc course request is also forced through for diagnosis. The saved screen includes the number of recent terms with usable capacity, the number at or above the fill threshold, every pressure flag, and a plain-language reason. Excluded rows therefore say, for example, 1 of 3 recent spring terms with class-list registrations at or above 90% of scheduled capacity and whether target capacity was available for the other checks. A pressure-only scope with no included courses publishes a valid zero-row bundle.

Comparable History Window

The model window begins at the institution’s projection_history_start_term (settings.csv; UNM: Spring 2022, 202210); a course may start later through projection_courses.csv. Earlier target enrollments, capacities, transitions, and aftcasts are excluded so pandemic-era behavior cannot silently influence accuracy or method selection. The immediately preceding source term may still support a structural transition into the first eligible target; for Spring 2022 that source is Fall 2021. It is a source population, not an older target outcome.

The requested as_of_term must be at or before cedar_data_edges(students)$last_enrolled_complete. The bundle builder fails closed when it cannot determine that edge or when the requested source term is still filling. This prevents an advance-registration snapshot from being treated as a completed upstream population.

Course-specific curriculum breaks can move the floor later. MATH 1215 begins with Fall 2025 (202580), when the current high-enrollment curriculum appears. Earlier small MATH 1215 offerings are not comparable, and MATH 1215X/Y/Z are not folded into the base course. For a Spring 2027 forecast this leaves Spring 2026 as the only same-season MATH 1215 observation. The prior-Spring method and its one-term first-day retention are permitted, but the row has accuracy Unrated and depth Thin until the current curriculum accumulates at least two aftcasts.

Both the general floor and course exceptions are saved in model_config. The validator rejects saved recent-history or aftcast rows that precede the effective course window.

Enrollment Target

The primary target is the first-day / ever-registered class-list proxy: every unique non-waitlisted student who appeared on the course class-list extract as registered, an early drop, a late drop, or another auditable status. Because CEDAR retains a final/current status rather than frozen roster snapshots, this is not a literal first-day count. It is the closest recoverable gross registration-demand measure:

classlist_total = registered + dr_early + dr_late + other_non_waitlist
first_day_enrl  = calc_first_day_enrl(): students holding a seat when classes began
census_enrl     = registered + dr_late
final_enrl      = registered

first_day_enrl is reconstructed from each row’s Registration Status Date against its section’s first class day (calc_first_day_enrl(..., pool_campuses = TRUE), each student-course once across the pooled market). It is the occupancy every seat judgment reads (0.21.0). It is NA, never a partial count, for a course-term whose class lists carry no status dates or whose rows were all pulled before a section began (late-starting part-of-term sections). class_list_status_date_coverage() reports which terms lack dates, and Admin > Data & Usage lists them; never name the terms in prose, because re-pulling a term’s class list fills its dates in.

Before market aggregation, repeated ABQ/EA rows are collapsed to one student-course-term. A currently registered record wins over a drop record; a late drop wins over an early drop. This status priority makes the equations reconcile while preventing a section or modality change from becoming two units of demand. An unfamiliar non-waitlist status is retained in other_non_waitlist for audit rather than dropped or assumed to occupy a census seat. Projection WAPE, signed bias, and method selection all use classlist_total. First-day, census, and final enrollment are saved with each aftcast for historical lifecycle comparison, but they are not alternative scoring targets.

Sections and seat gaps are planned for registrations, not first-day seats (0.23.0): projected registrations plus any pressure bump (below). Registrations run a median 8% above the first day, because a student who drops early frees a seat someone else takes, and the margin is kept on purpose: running short of seats costs more than an empty one. Over 207 past aftcasts with a known first day, sizing on the first day would have planned fewer seats than were filled on day one 61% of the time; sizing on registrations, 32%, with seats short (3,877) and spare (3,867) in balance. Most of those course-terms filled their seats, so the first day undercounts what they needed; see forecasting-lessons.md.

Pressure bump

A course near full on the first day (saturation_fill_threshold, 90%) in every one of its last pressure_bump_window (3) same-season offerings with a known first day, at least pressure_bump_min_terms (2) of them, is planned for more than it registered last time:

unserved          = max(pressure_bump_share x spine, students left waiting)
pressure_bump     = max(0, unserved - floor already in the published projection)
planned_registrations = projected_classlist_total + pressure_bump
recommended_sections  = ceiling(planned_registrations / reference_section_size)

pressure_bump_share is 0.10. The bump is withheld when the course’s feeder courses’ enrollment fell pressure_bump_feeder_decline (5%) or more in the term just finished; when that term is not observed yet (the two-step horizon), the check cannot run, the bump stands, and the note says to recheck. A term whose first day is unknown is skipped, never read as full or not full.

The reason is that a near-full course’s registrations follow its seats: given 10% or more seats they filled them (89 course-terms), held flat they stayed flat (139), cut they fell by the cut (51). The 10% is an allowance, not a measurement, and every row says which. projection_pressure_bump() writes pressure_judged_terms, pressure_full_terms, pressure_bump_applied, pressure_bump, and pressure_bump_note (why it applies or why not); add_projection_recommendations() adds planned_registrations and pressure_bump_sections (the sections it adds). The forecast, WAPE and the accuracy axes never include it, because a full course’s registrations are capped and would score any allowance as over-projection. validate_pressure_bump_rows() rejects a bundle whose plan is not projection plus bump or whose sections are not sized from the plan. The page states the total above the table (projection_pressure_bump_summary()).

Reliability, seat consistency, and where error comes from

The accuracy axis is shown to readers as Reliability (High within 10% WAPE, Moderate to 20%, Low beyond; it was Close / Fair / Poor before 0.23.0). The column keeps its name; only the values and the label changed.

annotate_projection_fold_conditions() tags every aftcast row with fold_seat_change (the fold term’s scheduled seats over the course’s previous same-season offering, minus one) and fold_prior_near_full (whether that previous offering was near full on the first day; NA when its first day or seats are unknown). Two summaries read them from the published rule’s aftcasts (policy_selected):

• add_projection_seat_consistency() gives each published row seat_consistency: High when the seats held within seat_consistency_change (10%) in every one of the course’s reliability terms, Moderate when they moved in up to seat_consistency_moderate_max_moves (1), Low beyond, Unrated when no term can be compared; with seat_consistency_reason, seat_consistency_terms and seat_consistency_moves. Low reliability with High seat consistency is the forecast’s own miss, flagged “Off 20%+ with steady seats”; with moving seats it is “Off 20%+ as seats moved”.
• summarize_projection_error_sources() attributes each aftcast to one source – seats changed 10%+, seat or first-day data missing, seats held while near full, seats held with room – and sums the absolute error by source. The page renders it in Where forecast error comes from (projection_error_sources_summary()); the text preview prints it.

On the Spring and Fall 2027 bundles, 77% and 83% of the published rule’s aftcast error came from terms whose seats changed, and no Low-reliability course had High seat consistency (the breakdown is in forecasting-lessons.md).

Draft schedule comparison

build_enrollment_projection_draft_comparison(view) sets each published row’s planned_registrations beside the target term’s schedule as the bundle captured it (scheduled_sections, scheduled_capacity, and per-campus delivery_components), dated by the bundle’s target_schedule_as_of (the latest DESR as_of_date for the target term; NA when none is loaded). Each course is one of: short of day-one seats (draft seats below projected_first_day_equivalent), covering day one but short of the plan, covering the plan, more than a section beyond it, or not in the draft. It is arithmetic over saved rows: the bundle’s refresh signature includes the target-term sections, so a changed draft rebuilds the bundle at the next refresh, and Shiny never refits to compare.

Expected first-day seats are still shown beside the plan, as context. For each aftcast or current projection, CEDAR estimates a leakage-safe same-season retention rate from earlier terms only, skipping any term whose first day cannot be reconstructed:

first_day_retention_rate = sum(first_day_enrl) / sum(classlist_total)
projected_first_day_equivalent = projected_classlist_total
                              x first_day_retention_rate

The default uses up to four prior same-season terms with a known first day and requires two. This converts gross registration demand into expected first-day occupancy; it does not change the forecast target, determine whether registration was constrained, or set the sections recommended. The target term’s class-list snapshot is kept outside training and supplies target_classlist_total_to_date, target_registered_now, and drop counts to date. DESR supplies the schedule and capacity only; its total_enrl is not a projection input or live-demand check.

Waitlists are not a method input. Class lists only began retaining waitlist rows in 2026 (Summer 2026: 543; Fall 2026: 3,615); a past term’s extract keeps only students still waiting when it closed. There is no history yet to fit or test a waitlist method against, so it cannot provide a defensible demand signal.

Method Roles

Model 0.20.0 has two candidates and one idea: seats needed is last year’s number, corrected for what can actually be seen.

Method	Role	What it is
Prior same-season	The spine, and the default	Last comparable season’s class-list demand
Prior same-season + floor + Fall deltas	Challenger	The spine, raised to a floor when that term was saturated, plus two deltas read from the term just finished

Reader-facing method descriptions live in CEDAR_ENROLLMENT_PROJECTION_METHOD_GUIDE beside the registry in R/lists/enrollment_projection_groups.R (platform; the monitoring scopes it reads are the institution’s files); the page generates its accordion from that catalog and a registry test keeps the two in step.

Why this shape

Two measurements on the 0.19.0 Spring 2027 bundle decided it (the full audit is in Forecasting Lessons):

• The error lives where the schedule moved. Over 238 rolling aftcasts, terms whose capacity held within 10% of the prior Spring aftcast at 6.1% WAPE; terms with 10% or more seats added at 23% (enrollment followed the seats, bias -22%); terms with seats cut at 39% (bias +38%). Seventy-seven percent of all error sat in the moved-capacity terms. A same-season baseline cannot see a schedule that has not been written, but it can see when the previous one was full.
• A Spring roster is 86% continuing students, 7% new entrants, 5% repeaters after a Fall DFW, 2% returners. Last Spring’s count already carries the steady state of all four. Replacing the continuing block with population times take-rate gained 0.6 points; the components only add information when a Fall driver changed from the year before.

So the published number is the spine, moved only where something measurable moved: the prior term was clipped by its seats, the course’s own Fall enrollment changed, or its feeders’ Fall enrollment changed.

The adjusted candidate

seats needed = spine + floor uplift + repeat delta + feeder delta

spine         prior same-season class-list total
floor uplift  if first-day fill of that term >= 90%:
              students still waiting at close who are on no campus's
              class list in the market; else 0
              (unknown first day: saturation cannot be judged, uplift 0)
repeat delta  (course enrollment in the preceding term
               - same a year earlier) x repeat rate
feeder delta  (distinct enrollment across the feeder set in the preceding term
               - same a year earlier) x feeder rate

The rates are learned only from pairs (preceding term, following same-season term) that precede the target, over the last four such pairs and at least two. The repeat rate is the share of students in the course in the preceding term who were in it again the following same-season term. The feeder set is the up-to-eight courses whose preceding-term students most often took this course next (at least three transitions over at least two pairs), and the feeder rate is one pooled share across that set so a student in two feeders counts once. Both deltas read enrollments, not grades, because grades are not posted when the schedule is built; the rates embed the historical pass and repeat behaviour, and the same terms can be re-read once grades land.

The floor is in the spine’s units (0.22.0). The spine is the ever-registered total, which already counts everyone who held a seat at any point, early drops included. A student still waiting at close never held one, so demand was at least the spine plus those students, and the uplift is exactly their number. Until 0.22.0 the floor was first-day enrollment plus the waitlist, compared with the spine. First-day enrollment leaves out the early drops the spine counts, so the floor rose above the spine only when the waitlist outnumbered the term’s early-drop churn: on Fall 2027, 245 waiting students in 36 saturated courses lifted the projection by 77, and 24 of those courses not at all (ISSUES.md I27). validate_spine_rows() now rejects any adjusted row whose uplift is not its waiting count when saturated and zero otherwise.

The waiting count starts from the canonical class-list true-waitlist rule (R/branches/waitlist-demand.R), which removes anyone registered in the same delivery group. The market then removes anyone with a non-waitlist row on either campus, because the class-list total already counts them: a student holding an EA seat while waiting at ABQ survives the per-campus rule but is not waiting in the pooled market (6 of 331 on Fall 2026). Class lists retain waitlist rows only from Summer 2026, so a saturated earlier term says “no waitlist rows retained” and the spine is itself the floor. That leaves the floor almost no aftcast record: on the 0.22.0 bundles a handful of stray Spring 2023 waitlist rows lift two Spring 2024 aftcasts by two students each, and nothing else in either season’s history reaches it. The saturation test is enrollment_capacity_metrics() from R/branches/enrl.R applied to first-day enrollment. Regstats’ Saturation report still reads census through capacity_saturation_metrics(), the same arithmetic on a different lifecycle point; the two pages each say which they use.

Beside a floor the bundle reports the market’s expansion response: among same-season pairs before the target whose prior term was saturated and whose capacity then grew 10% or more, the median share of the added seats that filled (93% on the Spring 2027 build, 44 pairs). It is context for a reader deciding whether to add seats, not a multiplier, because the question is how many seats are needed rather than what a hypothetical expansion would fill.

A regime check flags any course whose own enrollment in the preceding term moved 25% or more from the year before. It does not change the number; it tells a reader to look (FDMA 2110 rebounded from 106 to 355 in Fall 2026 and the 0.18.0 engine published 106).

What a Fall target gets

The floor needs only the prior same-season term and applies to either season. The deltas read the term immediately before the target; two steps past the settled edge that term is unobserved, so a Fall bundle built from a Fall edge carries the spine and the floor and says “Preceding term … is not observed yet” for the deltas. Which season holds which horizon alternates through the year:

Settled edge	One step out (deltas available)	Two steps out (spine and floor only)
a Fall term	next Spring	next Fall
a Spring term	next Fall	next Spring

Both current targets are pre-schedule. Spring 2027 and Fall 2027 carry target_schedule_available = FALSE for every row, so capacity-derived comparisons are unavailable rather than zero and the recommendation reads “Plan N section(s)”, or “Plan at least N section(s); prior term was saturated” when the floor applies.

Capacity Censoring

A full course records enrollment up to the seats the institution made available; it does not reveal how many additional students would have enrolled. Ordinary forecast error therefore rewards a naive method for reproducing a historical cap and can punish a structural estimate for projecting above an unknown ceiling.

CEDAR retains four accuracy views:

Metric	Meaning
wape	Ordinary error against actual class-list total; used when the history is not mostly capacity-reached
first_day_equivalent_wape	Error between projected first-day load and actual first-day enrollment, on terms where it is known
capacity_censored_wape	One-sided class-list error that does not penalize overprojection after registration reached capacity
uncensored_wape	Class-list WAPE only on terms whose registrations remained below capacity

Historical capacity is marked reached when first-day occupancy filled 98% of the seats or more:

first_day_enrl / scheduled_capacity >= 0.98    (calc_first_day_enrl())

98%, not 100% (CEDAR_FULL_FILL, R/lists/seat_fill.R, shared with every page that calls a course full). Capacity is pooled over every section, and a course with a few seats open across many section times is full to the student choosing one: CHEM 1215 filled 682 of 690 seats on the first day of Spring 2026 (98.8%). At 98% a 50-seat course with one empty seat is full. On the 0.21.0 bundles the line adds 12 Spring and 17 Fall course-terms of recent history to the reached set.

It is not census either (0.21.0). Census is the 15th day of instruction, three weeks after the first day (the term calendar in AGENTS.md), and the drops in between empty seats that were full on day one: on 598 ABQ+EA course-terms with a known first day, median fill was 98.9% on the first day and 95.6% at census, 46% of course-terms were full on the first day against 15% at census, and 199 were full on day one but under capacity at census. An unknown first day is unknown occupancy: capacity_usable is FALSE and the term is neither reached nor unreached, so it leaves uncensored_wape rather than counting as an unconstrained term.

It is deliberately not classlist_total >= scheduled_capacity. The ever-registered total counts seat churn: every early drop freed a seat that someone else took, so it exceeds capacity whenever a course turns over. On the Spring 2027 audit it marked 72% of course-terms as capacity reached, while census marked 17% and final enrollment 6%, and in 87 of those 284 “reached” terms fewer than 85% of seats were still filled at the end. Every downstream reading of the flag (the Capacity-bounded label, the capacity caveat, the demand signal, the pressure screen’s chronic-fill test) depends on it measuring real occupancy. For a target term still filling, which has no first day yet, the students registered right now are its occupancy. First-day enrollment above the nominal cap still counts as reached; overrides do not turn a bounded observation into an unbounded one.

For a reached-capacity term, the observed class-list total is treated as a lower bound on demand. The one-sided scored projection is:

if capacity_reached and projected_classlist_total > actual_classlist_total:
    capacity_censored_classlist_projection = actual_classlist_total
else:
    capacity_censored_classlist_projection = projected_classlist_total

This preserves the full penalty for underprojection while declining to call an above-observed estimate wrong when registration was bounded. A 0% cap-censored error means that overprojection cannot be established; it does not prove that latent demand was predicted exactly. Raw WAPE, cap-censored WAPE, uncensored WAPE, bias, and exact aftcast terms remain visible together. User-facing tables do not render that technical zero as 0.0%. They label the term Capacity-bounded, and an all-bounded accuracy window appears as Capacity-bounded (n/n). Mixed windows report the bounded count and identify the numerical cap-censored WAPE as a minimum rather than an observed error.

The class list records everyone who registered but has no timestamps from which to reconstruct peak concurrent occupancy. classlist_total >= capacity is therefore CEDAR’s operational historical ceiling signal, not a claim that the exact minute of constraint is known. Retained registration snapshots or usable waitlists would be required to measure blocked demand directly.

Retired candidates

The registry once held nine candidates: three same-season baselines, a student-level feeder-transition estimator, two Spring population methods, and 50/50 blends of each upstream signal with prior same-season. 0.19.0 retired the undamped trend (+11.7% bias), the feeder estimator (+35% bias) and its blend, and the population blend (a duplicate at correlation 1.000). 0.20.0 retired the seasonal median, the raw and blended preceding-term ratio, both Spring population methods and the anchored cohort blend: every one sat within 0.3 WAPE points of the spine or lost to it under rolling selection. The feeder idea survives as a delta, not a level estimator. The Spring cohort-cell preparation went with the population methods; cohort_composition stays because the growth scenario depends on it. The figures are in Forecasting Lessons.

Aftcasting and Selection

Backtests are rolling-origin evaluations. For each historical target, every candidate receives only terms and student populations that precede that target. The current target and future rows never enter training.

Accuracy is method- and course-market-specific. Every candidate and published row stores target_term, target_term_label, n_backtests, the exact comma- separated backtest_terms, their readable labels, and backtest_term_range. Any UI showing WAPE must display this context at the row level; a percentage without its evaluated terms is not an interpretable accuracy claim.

Selection is rolling, and the published accuracy is out of sample

Prior same-season is the default for every course. A challenger replaces it for a target only when, on at least selection_min_backtests (3) earlier aftcast terms the two share, the challenger’s WAPE was lower by at least selection_challenger_margin (two points). Three, not two: replayed on the 0.19.0 Spring 2027 bundle with two shared folds, 26 challengers were live on the 2026 fold and scored 11.4% where the default scored 8.2% on the same courses; with three, the rule’s realised record equals the default’s.

The same rule is replayed at every historical fold with only the folds before it, and the accuracy published for a course (n_backtests, wape, the Depth and Accuracy axes, the prediction interval) is the realised error of the method the rule would have picked at each fold. The saved backtests carry policy_selected, and the recent-history table shows, for each past term, the method that would have been live then. The published row’s selection_common_n, selection_default_wape and selection_challenger_wape record the evidence that made the current pick; a default pick has none.

This replaced in-sample selection on 2026-10-16. The 0.18.0 engine chose the minimum-WAPE method on the folds all eligible methods shared, which the seasonal-trend candidate shrank to two terms for 80 of 83 courses, and then reported the winner’s WAPE on those same folds. Measured on the Spring 2027 bundle: the published picks scored 13.7% in-sample, the same rule replayed rolling scored 16.5%, and prior same-season alone scored 15.9%. A rule that cannot beat the default out of sample does not get to replace it, and a selection_wape that is lower than the rule could have delivered is not a number a reader should see. tests/testthat/test-enrollment-projections.R guards this with a fixture where the in-sample winner’s fit is 2% and the honest record is 8%.

Sparse rows (no prior same-season at all) fall back to the first applicable observed baseline and say so.

Three axes, not one label

A single confidence label was retired on 2026-09-07. It fused three independent questions into one word and then reported the worst of them, so a reader could not tell which had failed. Every row in the production scope came back Low or None, including courses whose enrollment had barely moved in five years and which aftcast to within 2%.

Axis	Question	Values	Needs a model?
stability	Is this course predictable from its own history at all?	Stable / Moderate / Volatile / Unrated	No
depth	How much comparable aftcast evidence stands behind the selected method?	Deep / Moderate / Thin / None	No
accuracy, shown as Reliability	When we did aftcast it, how close were we?	High / Moderate / Low / Unrated	Yes
seat_consistency	Did the seats hold steady in those same aftcast terms?	High / Moderate / Low / Unrated	No

stability is measured with no model, from the coefficient of variation of the course’s own comparable class-list series over demand_history_window terms: Stable at or below 10%, Moderate at or below 25%, Volatile above, and Unrated below stability_min_terms (default 3) comparable terms. Median and maximum year-over-year change are carried beside it, because a single structural break and steady churn produce similar CVs and call for different responses.

That independence is the point: stability is publishable for a course the aftcast machinery cannot rate at all, which on a build where the upstream population has not settled is most of the scope. It is also predictive rather than merely descriptive — on the Spring 2027 build its CV correlates r = 0.69 with realised aftcast error, and no course rated Stable was forecast Poorly (13 Close, 2 Fair, 0 Poor, in the labels of the time).

A Volatile rating is a finding, not a failure. A course whose enrollment swings 50% year to year is one where something is moving, and the swing is what deserves attention.

depth and accuracy read the rolling-selection record (n_backtests, wape on the published row): how many folds the rule published and how far those published aftcasts missed. For a course where the default was never beaten that is prior same-season’s own record; where a challenger took over it is the mixed record a reader would actually have lived through. Since 0.19.0 selection_wape and selection_n_backtests are the same figures, so a row carries one accuracy number. (The retired single label failed in part by reading a common-fold count that topped out at 2 against thresholds written for a per-method count that reached 4, so Medium and High were unreachable.)

depth is Deep at depth_deep_min_aftcasts (4) folds, Moderate at depth_moderate_min_aftcasts (3), Thin at 1-2, None at 0. accuracy is High at or below accuracy_close_max_wape (10%), Moderate at or below accuracy_fair_max_wape (20%), Low above, and Unrated below accuracy_min_aftcasts (2) folds. Term-to-term error variation is reported in the reason text rather than collapsing the rating, so a close-on-average but inconsistent method is visible as such.

The axes are allowed to disagree, and usually do. Deep evidence spans High, Moderate, and Low reliability; a Stable course may have Thin depth. The validator checks each axis against its own vocabulary separately, because a shared check would have to accept every value on every axis and would catch nothing.

Qualifications are reported beside, rather than folded invisibly into, these axes. A mostly capacity-reached history says that fit is to observed class-list enrollment and cannot validate latent demand. A regime check says the term just finished moved sharply. Material disagreement between the two candidates is also named. The summary table compresses this to a short phrase; the course spotlight preserves the full interpretation. Prediction intervals use the selected method’s historical 80th-percentile absolute error. These are empirical planning intervals, not formal probabilistic confidence intervals.

Signed error

WAPE remains unsigned. It answers how far projections missed in aggregate and cannot honestly carry a plus or minus sign:

WAPE = sum(abs(projected - actual)) / sum(actual)

The companion weighted_bias records direction using the same denominator. Each method row also stores the standard deviation of its term-level percentage errors, over- and underprediction rates, direction_consistency, and a readable signed_error_history. A row-level interface should show WAPE and bias together, followed by the exact aftcast terms.

Bias correction was retired in 0.20.0. The rolling multiplicative calibration needed four training aftcasts plus two validation trials, and five comparable Springs never supplied them: zero rows were ever corrected. The floor and deltas correct the spine in the two places the audit found it wrong, which is a narrower and more defensible adjustment than a learned multiplier.

Code Ownership

Layer	File	Responsibility
List	R/lists/enrollment_projection_groups.R	Method registry and schema/model versions; reads the institution’s course groups from projection_groups.csv / projection_courses.csv and the window start from settings.csv
Branch	R/branches/enrollment-projections.R	Inputs, the spine and adjusted candidates, pressure, aftcasts, rolling selection, recommendation, persistence
Cone	R/cones/enrollment-projections.R	One question: projection and section need for the screened scope
Feature	R/features/enrollment-projections.R	Explicit-cutoff orchestration, Spring-only publisher guard, artifact loader, filtered reusable view payload, content fingerprints, model-source provenance, and canonical text preview
Module	R/modules/enrollment-projections.R	Read-only Registration page over the saved feature payload
Script	scripts/build-enrollment-projections.R	Clean-process artifact publication

Future modules and Course Dynamics must call the feature/cone or read the saved bundle. They must not reproduce a model, backtest, pressure calculation, or recommendation pipeline.

Saved Bundle

Schema 22 (model 0.23.0) sizes recommended_sections, seat_gap, projected_over_capacity, the capacity notes, and the recommendation from planned_registrations (registrations plus the pressure bump) instead of expected first-day seats; projected_first_day_equivalent and its range columns are unchanged and shown as context. Published rows gain pressure_judged_terms, pressure_full_terms, pressure_bump_applied, pressure_bump, pressure_bump_note, planned_registrations, and pressure_bump_sections, and seat_consistency, seat_consistency_reason, seat_consistency_terms, seat_consistency_moves; candidates gain feeder_change_rate and feeder_check, published as adjusted_feeder_change_rate and adjusted_feeder_check; aftcasts gain fold_seat_change and fold_prior_near_full; the bundle gains target_schedule_as_of. The accuracy values become High / Moderate / Low.

Model 0.22.0 keeps schema 21. Its floor_uplift is the spine-unit floor, and waitlisted_prior counts only students on no campus’s class list in the market (ISSUES.md I27); validate_spine_rows() enforces that the two agree.

Schema 21 (model 0.21.0) judges and sizes seats on the first day of classes. The market enrollment history gains first_day_enrl (and its prepare_projection_market_enrollment_history() takes sections for the start dates); aftcasts and recent_history carry actual_first_day beside actual_census; recent_history fills become first_day_fill / first_day_capacity_gap; and every census_retention_* and *_census_equivalent* column is renamed first_day_retention_* / *_first_day_equivalent*, because sections are now sized for day-one occupancy. The per-campus delivery_components keep their census columns, which are labelled as census and only audit the market rollup.

Schema 20 (model 0.20.0) has two candidates and no calibration columns. Each candidate row carries the model’s audit fields (spine, spine_term, saturated_prior, prior_first_day_fill, waitlisted_prior, floor_uplift, floor_basis, preceding_term, preceding_term_change, regime_alarm, delta_repeat, repeat_rate, delta_feeder, feeder_rate, feeder_courses, delta_basis), and validate_spine_rows() rejects an adjusted row whose parts do not sum to its projection. Every published row carries the adjusted candidate’s evidence as adjusted_* columns whichever candidate the rule selected, plus the expansion response and the floor and regime notes, so the detail can show them for every course. The inputs gain waitlist_history; student_inputs$course_enrollments returns for the deltas. Schema 19 (model 0.19.0) added policy_selected on backtests, the per-fold aftcast_method_id / aftcast_method_label on recent_history, the rolling selection evidence columns on projections, a preceding_term_history input, and renames the capacity columns to capacity_fill / capacity_gap (backtests) and census_fill / census_capacity_gap (recent history) because they then measured census occupancy. It drops selection_uses_uncensored and every feeder_* column. Published rows size recommended_sections and seat_gap from the occupancy equivalent, not the class-list headline: ever-registered runs about 13% above census at the median, and sizing from it had recommended 596 sections against 505 scheduled for the same courses on a demand estimate only 5% above their capacity. Schema 16 and later store the named market, exact course and campus scopes, campus/part-term delivery components, metadata, source fingerprints, the full effective model configuration, the pressure screen, published projections, every current candidate, every aftcast row, method performance, and a normalized four-term recent_history audit table. Each audit row stores the actual ever-registered total, first-day enrollment, census enrollment, final/last-day enrollment, sections, capacity, first-day fill, the aftcast the rolling rule would have published that term and which candidate it was, signed error, the capacity-censored flag and score, prior same-season enrollment/capacity changes, and a typed potential-miss explanation. Schema 16 also attaches the preceding term’s all-campus university headcount, pooled projection-market headcount, first-semester freshman count, and course DFW count/rate. DFW values are capped at the derived graded edge. The post-hoc repeat count joins canonical DFW students to same-course enrollment in the immediately following non-summer term; it is explicitly diagnostic and never enters a candidate formula. It is a recreated current-method aftcast, not a claim that CEDAR published that method in the historical term. Saving the effective configuration records defaults as well as caller overrides, including history floors and selection thresholds. Candidate rows mark the selected method and retain method role, applicability, evidence size, class-list WAPE, first-day-equivalent, cap-censored, and uncensored WAPE, signed bias and error history, and the spine / floor / delta audit fields listed above. The validator rejects an adjusted row whose components do not reconcile or whose spine term does not precede its target.

The validator rejects cutoff/target inversions, rows outside the saved course, campus, or market scope, duplicate market rows, duplicate delivery components, component totals that do not reconcile to market capacity, target-term mismatches, and disagreement between a published method and its selected candidate. Publication writes a temporary file and atomically renames it into place.

model_version identifies the calculation contract; it changes when a formula, method-selection rule, calibration rule, or threshold changes. schema_version identifies the saved-file shape and can change without changing the estimates. Model 0.17.0 excludes AUD from DFW and repeat-demand signals even when the registration status is DG/DW. Rebuild older bundles before using them with this model; validation rejects their older model version. The saved schema is unchanged. Every schema-16 bundle also stores the Git commit when available, whether any model source file differed from that commit, SHA-256 hashes, and an embedded normalized copy of the source files that define the model. Validation recomputes the hashes before a bundle can be read. This makes a dirty development artifact inspectable, while official production artifacts should still be built from a clean commit.

Routine development and UI recomputation replace enrollment-projections-TARGET-latest.qs; they do not create permanent forecast vintages. Retain an official vintage only through a deliberate --output publication action when its numbers are used for a scheduling decision. A materially revised decision can create another labeled vintage. The eventual outcome record should reference that official forecast; an automatic archive of every code run is explicitly out of scope.

Docker mounts the repository’s gitignored output/ directory read-only at the same path inside the app. The one writable mount belongs to the cedar-projections Compose service, which publishes bundles; a --volume on a docker compose run against the app service does not override the app’s read-only mount, so the publisher runs as its own service. Publish the latest bundle on the host before starting or restarting CEDAR; the UI does not write artifacts. An absent or invalid bundle produces an explicit empty state instead of a model run.

format_enrollment_projection_preview() renders the validated bundle as a stable Markdown/text table containing the current projections and up to four same-season evidence rows per course. It is the fast development view and the table-contract reference for the Shiny module. The formatter performs no model computation, and the module consumes the bundle’s typed projections and recent_history tables rather than parsing the text output.

The course detail also consumes typed backtests rows to plot every applicable method against the three enrollment-lifecycle measures. The feature layer filters these rows to the course’s term type and builds the plot; the Shiny module does not recompute methods or reshape model data.

The enrollment-movement accordion is built from the typed diagnostic columns on recent_history. Student-term population signals are deduplicated before counting. Course outcomes flow through classify_enrollment_outcomes(), so the default passes only A+ through C and CR; every other recorded non-audit outcome plus a non-audit late drop is DFW, while early drops never are. AUD is excluded even under late-drop status; a blank/NA late-drop grade still counts as a withdrawal. The feature layer builds the schedule/upstream/DFW narrative and presentation table; the module only renders that payload. Correlation is descriptive: schedule changes can be a response to demand, so the UI never labels it a causal capacity effect.

Registration > Projections lists saved bundles with find_enrollment_projection_bundles(), loads the reader’s chosen target term with load_enrollment_projection_bundle(), and filters it through build_enrollment_projection_view(). It never runs a candidate method, aftcast, calibration, or pressure screen. server.R owns the path and the file read; the module receives a discovery reactive and a loader function. Course Dynamics must reuse that same loader/view boundary when projections are added there.

Discovery is season-aware because CEDAR publishes one bundle per season. A Fall 2027 and a Spring 2027 projection are both current and neither supersedes the other, so a loader must name a season (load_latest_enrollment_projection_bundle(dir, "fall")) or an exact target term. The earlier “highest saved target term” rule would have made the first published Fall bundle silently hide Spring from every reader.

Population Growth Scenarios

A dean asks “if health professions grows 10% a year for five years, how many students and sections will the critical courses need?” That question is answered by build_enrollment_projection_scenario() (R/features/enrollment-projection-scenario.R) as arithmetic over rows the bundle already carries. Nothing fits, aftcasts, calibrates, or selects a model in a user session — the same rule that governs the projection page, and also what makes the scenario fast enough to explore.

What the bundle saves for it

cohort_composition (schema 18) records, per published course, the baseline roster counted by each student’s same-term major_code and student_classification. The baseline term is the most recent comparable same-season term at or before the cutoff — the same comparability rule the projection methods use, so the composition describes the term the projection is anchored to. Students with no program row that term are kept with major_code = NA: they really are in the seats, and dropping them would stop the composition summing to the course total.

Per-term Major is one of the fields the reliability contract marks safe, which is why the composition can be attributed to a past term at all.

The arithmetic, and the assumption it rests on

pop_cohort   = baseline students whose major_code is in the population
year 1       = the published projection, unchanged
year k       = year 1 + pop_cohort * ((1 + g)^(k-1) - 1)
sections_k   = ceiling(year k / reference_section_size)

The assumption, stated on the page rather than implied: the population’s headcount grows at the given rate and its students keep taking each course at today’s rate; everyone else in the course is held flat. A course whose baseline roster is 13% health-professions students therefore moves 1.3% when that population grows 10%. The share is displayed per course, because a reader who cannot see it will assume the whole course grew.

Section sizing goes through projection_sections_for_demand() / projection_additional_sections(), shared with the bundle builder, so a scenario at 0% growth cannot disagree with the projection it is anchored on.

Honesty rules

• Year 1 is the published projection at any growth rate. Growth applies to the years after the target term.
• Later years are labeled Scenario, carry no accuracy axes, and never borrow year 1’s. No aftcast has been run against a hypothetical.
• A course whose population cohort is below min_population_cohort (default 10) is listed with a reason rather than dropped or shown as zero — a thin population must read as thin evidence, not as an absence of demand.
• A course with no published projection to grow from is listed the same way.

Named populations

Groups live in CEDAR_POPULATION_GROUPS (institution/<id>/population_groups.csv, read by R/lists/population-presets.R) and are shared with Pathways, so “health professions” means one thing across CEDAR. They declare program names; population_group_major_codes() resolves those to Banner codes through two mechanisms, because neither is complete:

Mechanism	Catches	Misses
Name match	A pre-major carrying its major’s name (FRAD / RADS), including codes Banner added after the list was written	A pre-major whose name has drifted
leads_to (programs.csv)	A drifted pre-major whose row leads to the right major	A pre-major with no leads_to

Both once failed on the same row for FMDL (“Medical Laboratory Science” against MEDL’s “Medical Laboratory Sciences”, with no canon entry). Its programs.csv row now leads it to MEDL, and the registry still lists both spellings and population_group_audit() reports near_miss names. Run that audit when adding a group. ISSUES.md I7 documents the live mapping defect this work uncovered.

Measured on real data (Fall 2026 baseline, 139-course scope): the clinical group resolves to 27 codes and holds 4,358 of 27,982 seats — 15.6% — across 1,576 of 11,440 students. A hand-listed code set omitting FRAD, FDEH, FEMS and FMDL would have missed 1,281 of those seats.

Real-data audit of the saved composition

The check the designed fixtures cannot make: cohort_composition keys on cedar_students$major_code while a population group resolves against cedar_programs$major_code. If those vocabularies disagreed, every cohort would silently be zero and the scenario would report a flat line as a finding. They agree:

	Spring 2027	Fall 2027
Composition rows / courses	6,028 / 83	7,016 / 90
Students counted	17,467	23,451
Baseline terms used	202310–202610 (Springs)	202380–202680 (Falls)
Rows with major_code = NA	0	0
Composition rows matching the clinical group	718	774
Students matched	2,767	3,471

Each course takes its own most recent comparable term as its baseline, which is why four baseline terms appear rather than one, and every one is in the target’s season.

Fall 2027 at 10% annual growth, always-monitored scope — 11 courses scaled, none excluded. The spread is the reason the share column exists:

Course	Population in course	Share	Fall 2027	Fall 2031
BIOL 1140	475	68.5%	684	904
MATH 1350	290	25.0%	1,161	1,296
ENGL 1110	412	18.6%	2,219	2,410
CHEM 1215	113	12.8%	915	967

The same 10% assumption moves BIOL 1140 by 32% and CHEM 1215 by 5.7%, because the population is 68.5% of one course and 12.8% of the other. A reader shown only the demand column would reasonably assume both grew 10%.

Development and Release

Use scripts/cedar-repl.R for repeated real-data work, then re-source only the changed files. Publish in a clean process with:

Rscript --vanilla scripts/build-enrollment-projections.R \
  --target-term 202710 \
  --as-of-term 202660 \
  --group critical_courses

Inspect the working artifact without starting Shiny:

source("scripts/cedar-repl.R")
bundle <- read_enrollment_projection_bundle(
  "output/projections/enrollment-projections-202710-latest.qs"
)
print_enrollment_projection_preview(
  bundle,
  courses = c("MATH 1215", "CHEM 1215")
)

# List the saved model files and hashes, then inspect exact saved source.
enrollment_projection_model_source(bundle)
cat(enrollment_projection_model_source(
  bundle, "R/branches/enrollment-projections.R"
))

# This path form remains available after CEDAR has moved to a newer schema.
cat(enrollment_projection_model_source(
  "output/projections/official-spring-2027.qs",
  "R/branches/enrollment-projections.R"
))

Focused fixture and architecture tests live in tests/testthat/test-enrollment-projections.R and tests/testthat/test-architecture.R. The projection browser suite exercises the saved-bundle table, filters, and row evidence through ./run-tests.sh --e2e enrollment-projections; the release gate remains ./run-tests.sh --all.

Model research findings, failed assumptions, and current real-data benchmarks are maintained separately in Forecasting Lessons. That file records evidence; the live backlog remains in ROADMAP.md.

Known Limits and Next Methods

• Structural estimates are signals, not identified latent-demand counts.
• The class list has no registration timestamps, so capacity reached is an operational signal rather than a recovered peak-occupancy observation.
• Method selection can be unstable with only two or three aftcasts; the rolling rule limits the damage (a challenger needs two shared folds and a two-point margin) but pooled or hierarchical shrinkage should be evaluated before expanding the scope.
• A robust ensemble of the observed baselines may improve class-list demand fit. A damped trend and a pooled growth factor were tested on the 0.21.0 aftcasts and both did worse than prior same-season (Forecasting Lessons).
• About two-thirds of the aftcast error moves with the schedule, one registrant per seat. A projection conditioned on the submitted draft schedule is the largest open improvement; it needs that schedule archived.
• A censored count model is the principled next latent-demand candidate once the capacity contract is reliable.
• Registration-pace models require retained historical snapshots; the current one-snapshot-per-term data cannot support them.