Right-Edge Policy

Why the right edge of the data is not one number, which edge each kind of analysis uses, and the measured failures that established the rule. AGENTS.md carries the rule; this page carries the reasoning.

Roadblocks observation grain: get_stopout() selects the first eligible outcome per student/course/delivery campus after scope, population-window, and right-edge filtering. Agreeing first-term records collapse; conflicting first-term outcomes exclude the comparison and are reported in observation_info. Counts, rates, DFW context, and chi-square tests use the same selected observations. Later enrollments remain available to the separate full-history return lookup. This Roadblocks-specific rule does not change CEDAR’s all-attempt course DFW rates; get_dfw_rates() is a separate ever-DFW measure and must not supply Roadblocks context.

Never bound an analysis with max(term), a hardcoded term, or arithmetic on cedar_current_term. Use the edges computed once at startup by cedar_report_edges() (R/branches/data-edges.R). This is the same class of rule as the campus and DFW policies.

Why an edge is not one number

CEDAR’s local data runs behind the registrar, and a term does not arrive all at once — registration appears months early, grades land weeks late. At any moment the tail looks like this:

term	rows	graded
Fall 2025	128k	91%
Spring 2026	119k	7%
Summer 2026	15k	0%
Fall 2026	85k	0%

So “the last term in the data” is at least two different terms depending on what you are asking.

Which edge

Edge	Use it for
last_graded (cedar_graded_through)	anything reading a grade: DFW, pass rates, grade distributions, stop-out after a DFW, course outcomes
last_enrolled_complete (cedar_report_end_term)	settled enrollment reporting and the hard observation edge for every longitudinal analysis
last_enrolled (cedar_report_latest_term)	descriptive current-registration charts — headcount, course enrollment, attempted SCH — only with every term after last_graded labelled in progress (cedar_in_progress_terms(), mark_plotly_terms()). Never a trend comparison, a longitudinal denominator, or a grade-based rate
last_degree	completions

cedar_report_end_term is no longer hand-maintained, and has no config fallback. cedar_report_edges() derives all four edges at startup, for global.R and the cache-warming scripts alike, so the app and the caches it reads agree. When the derived value replaced the hand-set one, it matched what config had been set to by hand, which was the check that it works.

Settled enrollment, and why not by size

A term counts as settled when the newest pull covering it happened at least min_days_after_start days (default 14, clearing add/drop) after the term began. Term start is approximated from the term code — Fall mid-August, Spring mid-January, Summer mid-June.

Comparing a term’s row count against the same season in prior years separates the cases just as cleanly today (Fall 2026 sits at 71% of a typical fall). Do not use it. A genuine enrollment decline is indistinguishable from an unfinished term under a size rule, so it would quietly truncate reports in exactly the year a chair most needs to see the drop. Pull timing cannot make that mistake — it describes when the data was captured, not how many students exist. There is a test for this.

Why this is a silent-wrongness rule

Bounding a grade rate by the enrollment edge does not error — it returns a plausible number. Measured before this policy existed: the DFW rate for Spring 2026 read 0.3% against 6–8% for every other term, because prepare_course_attempts() capped at cedar_report_end_term (the enrollment edge) and the denominator counted all 104,431 attempts while the numerator saw only the 7% of grades that had posted. Every DFW surface in the app showed the newest term as a dramatic improvement.

Why config arithmetic is not the answer

cedar_report_end_term <- subtract_term(cedar_current_term) encodes a guess that exactly one term is in flight. It fails both ways:

• Stale. Grades land after the term ends; nothing advances until somebody edits config.
• Overshoot. A config set ahead of the data nominates a term with no grades at all, and every student in it reads as having no outcome.

cedar_data_edges() reads the data, so it is right on whatever snapshot is loaded and moves on its own. cedar_report_end_term survives as the enrollment edge and as a cache-key component; it is not a grade boundary.

In practice

• Grade-dependent code uses cedar_graded_through; enrollment reporting uses cedar_report_end_term. Neither falls back to config: if the data cannot determine the settled or graded edge, cedar_report_edges() stops startup, because some surfaces (Pathways’ population window) read a NULL bound as no bound at all.
• Nothing caps the edges at cedar_current_term. A term registering ahead of the current one reaches the descriptive charts, labelled in progress, the day it lands; the settled rule alone keeps it out of trend comparisons. A cap here used to hide each next term’s registrations from Dept Trends until someone edited config.
• Startup prints all four edges.
• The threshold (min_graded_share, default 0.5) is not finely balanced — finished terms sit at 83–91% and in-flight terms at 0–7%, so anything from ~0.2 to ~0.8 picks the same edge.
• An edge that cannot be determined is NULL, never a guess. Fail closed.
• Say which edge you used. cedar_edge_note() produces the sentence. A capped view that does not explain itself reads as a stale pipeline; “Spring 2026 is enrolled but not yet graded” reads as a data state and tells the user when it will move.

Grade outcomes and longitudinal follow-up require two separate right edges

Do not cap an entire page merely because it contains a longitudinal panel. Current registration is valid descriptive data and may appear in Course Dynamics Overview and other explicitly current-enrollment summaries. The hard right edge belongs to computations that need comparable history or a later observation: retention, next-term persistence, course flows, sequence effects, and downstream success. Those analyses stop at last_enrolled_complete. If they also read a grade, use cedar_longitudinal_edge(edges, grade_dependent = TRUE), which is the earlier of last_enrolled_complete and last_graded. Never let a partial current term enter a longitudinal cohort, lookup, denominator, cache key, or outcome.

This distinction is deliberate: on the August 2026 snapshot, Fall 2026 registration is useful in the Course Dynamics Overview, but longitudinal sampling ends at Summer 2026, the previous complete term. A page-wide filter would throw away good current information; no filter would turn advance registration into apparently observed follow-up.

A row with no posted grade is unknown, never a failure or a non-pass. Filter grade-dependent event rows to last_graded before selecting an attempt or classifying its outcome, then pass them through classify_enrollment_outcomes(). Do not use a catch-all branch such as TRUE ~ "failed": blank grades, incomplete work, audits, NR/NC, and other unclassifiable records must be excluded from the grade-rate denominator and reported as unobserved.

An A→B analysis also needs an opportunity edge for the A cohort. A student who took A too recently to have the declared follow-up interval before the longitudinal observation edge has not failed to continue; the record is right-censored. For a one-regular-term opportunity window:

1. Compute the first possible follow-up with add_next_term_col(..., summer = FALSE).
2. Exclude A rows whose follow-up term is after last_enrolled_complete from the continuation denominator.
3. Report the number excluded and show the data window in the page methodology.
4. Keep this separate from the B outcome edge: a graded B must be at or before the earlier of last_enrolled_complete and last_graded.

On the August 2026 snapshot, Fall 2026 enrollment is already present but has no grades, so grade sampling ends at Summer 2026. This is data-derived, not a hardcoded “current term minus one” rule; on another pull the named terms will move. The Course Dynamics failure that established this rule was measurable: 33 Fall 2026 ENGL 1120 registrations after Oravetz’s FYEX 1030 sections were classified as failures. Capping at the graded edge changed that group’s apparent ENGL 1120 pass rate from 66.7% (98/147) to 86.0% (98/114 observed outcomes). The corrected opportunity window excludes 123 of Oravetz’s 418 first-attributed FYEX students because their next regular term falls after the graded edge (90 of them are Fall 2026 records with no later term in the data at all). Both errors produced plausible instructor comparisons rather than an exception, which is why every longitudinal page must display its outcome edge, opportunity rule, and exclusion counts.

For prerequisite/order questions, distinguish strictly earlier (term_y < term_x) from same-term (term_y == term_x) completion. Concurrent courses must never be described as having been passed “before” the focal course. When a single downstream course was already passed in a strictly earlier term, exclude that student from a progression-to-that-course denominator and surface the count. For a multi-course rollup, show prior completions as context but do not infer that passing one member makes the student ineligible for every course in the set. The worked FYEX 1030 audit illustrates why: among 439 distinct students ever taught by Oravetz, 38 had passed ENGL 1110 or 1120 in a strictly earlier term, 80 passed one in the same term, and 110 were in the student-level union (8 did both). Reporting the same-term rows as “passed before,” or adding the two counts without deduplicating, materially overstates the reverse-order population.