Orbuculum Fragility Score — Methodology

Orbuculum runs two separate scoring models with the same four-pillar architecture (Cash Position, Pricing Power, Debt Burden, Liquidity) and the same 0–100 composite scale, but different inputs and different thresholds. Public and private higher-ed finances are structurally non-comparable — the accounting standards differ (FASB vs. GASB), state appropriations create a structural asymmetry on the operating-margin line, and public-university debt is fragmented across off-balance-sheet entities that don't appear in IPEDS at all. We never blend them: no national score, no cross-sector peer matching, no mixed leaderboards.

Jump to: Private nonprofit colleges · Public universities · Debt per student

A handful of state-related schools (Penn State, University of Pittsburgh, Temple, Lincoln) are CONTROL=1 (public) but file under FASB on the F2 form — a legacy of their hybrid public-private status. We score them on the private path because that is how their financials are reported.

Inflation adjustment. All dollar comparisons that span time inside the scoring engine are real (CPI-U, base year 2024). This includes the pricing-power pillar's net-tuition trajectory, the debt-acceleration penalty's 5-year real-debt comparison, and the tuition-collapse override on the liquidity pillar — for both the private and public models, using the same CPI-U series. The headline "15-yr debt growth" stat on the homepage is the lone exception and is currently shown nominal (with a sublabel that says so); converting it to real is on the Phase 2 list. See Section 4 — Inflation Adjustment for the formula and rationale.

Private nonprofit colleges

Score method: mvp_v3. Covers ~1,650 U.S. private nonprofit four-year institutions. Calibrated against five canonical schools spanning the fragility spectrum, then validated by predictive lookback against publicly-announced closures.

1. Cohort and Scope

Calibration cohort (5 schools):

Sector: Privates only. Public institutions use the separate model below. The two models share architecture but never share thresholds — the sector firewall is enforced in code (separate scoring scripts, separate score_method strings, no cross-method peer matching).

Time range: Fiscal years 2004-05 through 2022-23 (19 years). Fiscal year is identified by fiscal_year_end = calendar year of FY end (e.g., 2023 = FY 2022-23).

2. Data Sources

Primary: IPEDS (NCES Data Center)

• Finance F2 component (FASB privates) — https://nces.ed.gov/ipeds/datacenter/data/F[YY1][YY2]_F2.zip per fiscal year.
• Instructional Activity (EFIA) — https://nces.ed.gov/ipeds/datacenter/data/EFIA[YYYY].zip per academic year.
• Header (HD) — institutional metadata.

Secondary

• CPI-U annual averages from BLS series CUUR0000SA0 — used for inflation adjustment to base year 2024. Hardcoded in ingestion script for reproducibility; updated annually.

Why IPEDS only. Coverage and uniformity. IPEDS is mandatory federal reporting for any institution receiving Title IV aid — comprehensive across all U.S. higher ed. ProPublica 990s, audited statements, NACUBO endowment data, and bond prospectuses are richer for individual schools but inconsistent at scale. Pragmatic choice: IPEDS-only as primary source, label proxies honestly, layer richer data only as targeted enrichment for flagged schools.

3. IPEDS Variable Map (F2 / FASB)

Core variables (stable across full 2005–2023 window):

Bonus variables (added in newer years):

Enrollment variables (EFIA):

Total FTE = EFTEUG + (EFTEGD or 0).

Schema-drift handling. The Finance F2 schema added new variables in 2008, 2016, and 2020. The data model uses long-format facts tables (finance_facts rows = unitid × year × variable_code × value), so new variables become new rows, never new columns. Year-aware variable selection in scoring picks the best-available variable per year, falling back gracefully when a "modern" variable is missing.

4. Inflation Adjustment

All dollar comparisons are real (inflation-adjusted) using CPI-U, base year = 2024.

real_value = nominal_value × (CPI_2024 / CPI_year)

CPI-U was chosen over HEPI (Higher Education Price Index) because:

• BLS-published, freely redistributable, universally recognized.
• Marginal accuracy gain of HEPI (1-2pp/yr higher) doesn't justify the licensing/familiarity tradeoff for MVP.
• HEPI may be revisited later as a methodology toggle.

5. The Four Pillars (Privates)

Each pillar produces a 0–25 score. Higher score = more fragile.

Pillar 1: Operating Margin

Raw metric: Operating margin

Robustness. The legacy adjusted margin is clamped to [−25%, +25%]. Severe endowment-loss years (e.g., FY 2008-09 financial crisis) can produce mathematical extremes that pollute multi-year rolling averages; clamping at ±25% means a really bad year shows as max distress without dragging 5 subsequent years' rolling averages into permanent distress territory. When the adjusted denominator becomes degenerate (negative or smaller than 50% of expenses), the formula falls back to expense-denominated: (F2B04 − ΔEndowment) / F2B02.

Scoring (Fitch-style bond-agency bands):

Trajectory smoothing — "worst-of" rule. Score is max(score(current_year), score(rolling avg)). The window length is 3 years for FY 2019-20 onwards (F2I-based margin is already clean), and 5 years pre-FY 2019-20 (the F2B-derived proxy needs more smoothing because non-endowment, non-operating items — pension actuarial gains/losses, restricted-gift volatility, capital grants — can move year-over-year and aren't separable from operations in the legacy data). This catches both acute distress and chronic distress; a single positive year does not erase a prior multi-year pattern.

Persistence penalty: +3 points if 3+ consecutive negative-margin years (capped at 25).

Pillar 2: Pricing Power (Market Strength)

Raw metrics:

• real_net_tuition_per_fte = F2D01 × cpi_factor / total_FTE
• discount_rate = (F2C05 + F2C06) / (F2D01 + F2C05 + F2C06) — funded + unfunded institutional grants over gross tuition. This is the all-student, institution-wide tuition discount rate. NACUBO and individual schools typically publish the first-time-in-college (FTIC) freshman discount rate, which runs 5-10 points higher because recruitment merit aid is front-loaded onto incoming classes and often step-downs in subsequent years. The all-student rate is the right input for revenue-stress modeling (it's what actually shows up in F2D01); the FTIC rate is the right input for thinking about future-year recruitment economics.

Trajectory measured over 5 years.

Scoring:

Discount-rate add-on: +5 points if all-student discount rate > 55% (heavy aid as a substitute for pricing power). Threshold is on the institution-wide rate; for context, FTIC rates above ~63% typically correspond to all-student rates above 55%.

Pillar 3: Debt Burden (Leverage on Eroding Base)

Raw metric: Viability ratio = expendable net assets / plant-related debt. CFI-aligned.

Scoring (CFI-aligned bands):

Debt-acceleration penalty (covenant-risk proxy):

Penalties combine, capped at +10. Final pillar score capped at 25.

Why this matters. A school whose viability ratio looks "fine" only because they recently took on new debt and parked it in reserves is structurally at high covenant risk. Hampshire College in FY16 doubled long-term debt in one year (102% jump) — the viability ratio fell from 1.74 to 1.07 (still inside an 8-point band under simpler scoring), but the actual fragility was severe. The acceleration penalty catches this.

Pillar 4: Liquidity (Survival Runway)

Raw metric: Days cash on hand = expendable net assets / (total expenses / 365)

Scoring (CFI-aligned bands):

Tuition-collapse override. If real net tuition per FTE has dropped > 30% over 5 years, minimum liquidity score = 16 regardless of days cash.

Why this matters. A school that cut deeply and received emergency relief gifts can show very high days cash even when structurally insolvent — "deathbed cash." Hampshire in FY 2022-23 had 298 days cash (would normally read 0/healthy), but their real net tuition revenue had collapsed 77% from FY 2017-18 to FY 2022-23. The override prevents misreading reserve cash from a relief grant as operational health.

6. Composite Score (Privates)

composite_pre_adjustment = Pillar1 + Pillar2 + Pillar3 + Pillar4 (range 0–100)

Compounding-fragility bonus. If 3 or more pillars worsened over a 3-year window (i.e., pillar_score(year T) > pillar_score(year T-3) for at least 3 of 4 pillars), add +10. This captures the user's debt thesis: schools where multiple pillars co-move toward distress are more fragile than schools where any single pillar deteriorates in isolation. Pillar co-movement is signal, not noise.

Endowment Backstop Adjustment. Most endowments are not a usable cash reserve (donor-restricted). However, at sufficient magnitude, the accessible portion (heuristic: 10% of total endowment) becomes a meaningful operating-deficit absorber. The adjustment captures this dynamically:

Inputs:

• endowment = F2H02 (end-of-year endowment value)
• available_endowment = endowment × 0.10
• cum_deficit_5y = absolute value of summed operating deficits over trailing 5 years (only deficit-years counted; positive years do not replenish — restored capacity from gifts is structurally different from durable operations)
• stress_deficit = worst single-year operating loss in trailing 5 years; if no recent deficits, 5% × current operating revenue as hypothetical stress
• burn_ratio = cum_deficit_5y / available_endowment
• remaining_buffer = max(0, available_endowment − cum_deficit_5y)
• coverage_years = remaining_buffer / stress_deficit

Adjustment table:

Relief by coverage years (when in relief zone):

When burn_ratio is 0.5–1.0, the relief is halved (partial-burn dampening).

Why this works. The model now distinguishes three structural states:

1. Endowment-anchored health (Harvard): Operating tightness is fine because reserves dwarf any plausible deficit.
2. Endowment-exhausted distress (The New School): Operating deficits have already consumed the accessible cushion. They had a backstop. They used it. It's gone.
3. No backstop at all (Hampshire, Birmingham-Southern, Wells): Either insufficient endowment from the start, or burned past it with active deficits continuing.

Final composite. composite_final = clamp(0, 100, composite_pre_adjustment + compounding_bonus + endowment_backstop)

Buckets:

7. Calibration Targets (Validation)

The 5-school cohort was chosen to validate the model produces correct ordering and spacing:

• Harvard ≪ The New School < Hampshire < (Birmingham-Southern, Wells)
• Closed cohort should reach High Fragility (>60) at least 3 years before public closure announcement.
• Harvard should never exceed Low Fragility (in clean post-FY2020 years).

v3 actual FY 2022-23 results:

Closed-cohort predictive lookback:

8. Score Versioning (Privates)

The score_method column in pillar_scores and fragility_scores distinguishes algorithm versions. Multiple versions live side-by-side; we never overwrite history when iterating.

• mvp_v1 — initial absolute thresholds (CFI-style soft bands).
• mvp_v2 — added tighter viability bands and debt-acceleration penalty.
• mvp_v3 — current. Added Fitch operating-margin bands, "worst-of" trajectory smoothing, tuition-collapse override, endowment backstop adjustment.

Data-quality guard: F2I05 anomaly sanity check

A small number of elite privates filed F2I05 (expendable net assets) anomalously through the FASB ASU 2016-14 transition window — Brown reported $5M in FY 2021-22 (down from $5.35B the prior year, with operations and endowment unchanged); Claremont McKenna reported $3.4M (then negative) starting FY 2019-20 against $336M F2A04 and a $1.2B endowment. Without a guard, this drove three pillars (debt burden, liquidity, and cash position via smoothed worst-of) to max scores, producing false-positive composites of 80–100 for operationally-healthy schools.

The scorer detects this pattern using three complementary triggers (any one fires):

1. Negative F2I05 with healthy fundamentals. F2I05 < 0 while F2A04 > 0 and F2H02 > 0 — mathematically nonsensical for an operating school.
2. YoY collapse with F2A04 stable. F2I05 drop > 90% year-over-year while F2A04 stays within ±10% — a filing break, not a real change.
3. First-year cliff far below normal envelope. F2I05 < 5% of F2A04 when F2A04 > $50M and F2H02 > $50M — legitimate F2I05/F2A04 ratios are 20–50% for schools with substantial PP&E and quasi-endowment, so a sub-5% ratio with healthy underlying fundamentals indicates the F2I05 line wasn't filled in correctly.

When any trigger fires, the scorer substitutes F2A04 for F2I05 in the viability ratio and days-cash computations. The cash position pillar is unaffected (it uses F2I01/F2I02, a separate operating-margin block). Verified: Brown FY 2022-23 corrected from 82 to 30 (low); CMC FY 2022-23 from 100 to 75 (high — still elevated, but now driven by real plant-debt acceleration and thin operating margin rather than the filing artifact).

9. Known Limitations (Privates)

1. Pre-FY 2019-20 operating margin uses a noisier proxy (F2B04 / F2B01) which includes investment returns. Visible mostly for schools with large investment portfolios (e.g., Harvard 2008-2013 scores show financial-crisis volatility that's not really "fragility"). Tuition-dependent schools are largely unaffected. Could be smoothed in future with a non-tuition-dependent filter.
2. Covenant violations are not directly visible. The score detects financial conditions correlated with covenant risk (negative DSC, debt acceleration, low liquidity) but cannot read specific covenant terms — those live in bond prospectuses (EMMA filings), Phase 2.
3. Discount rate from F2C05 + F2C06 is a proxy. True institutional discount rate uses F2C10 (total discounts and allowances) where available (FY 2015-16+) — to be migrated in future.
4. Five-school cohort is small for floor calibration. Harvard is the cohort floor at 13; truly stronger schools (e.g., Williams, Pomona — small endowed liberal arts) would likely score 1–5. Demonstrating the floor requires expanding the cohort.
5. Pre-2008 viability uses a low-confidence proxy (F2A04 / (F2A03 × 0.6)). Marked separately as low-confidence in the data model; should not be used for trajectory analysis pre-2008.
6. One-time accounting events distort single years. Mitigated via 3-year smoothing in cash position pillar; not yet applied to other pillars.

Public universities

Score method: mvp_v3_public_dso. Covers ~800 U.S. public four-year institutions plus the four PA state-related FASB filers scored on the private path. The model shares architecture with the private model — four pillars, 0–100 composite, same buckets, same compounding bonus, same endowment backstop framework — but every input and every band differs because GASB and FASB are not interchangeable.

Why a separate model

Different accounting standard. Privates file under FASB (Financial Accounting Standards Board); publics file under GASB (Governmental Accounting Standards Board). FASB tracks "net assets" and distinguishes restricted from unrestricted; GASB tracks "net position" and distinguishes invested-in-capital-assets from restricted from unrestricted. In IPEDS, this shows up as two separate finance components: privates report on F2, publics on F1A. The variable codes are entirely different. F2I05 ("expendable net assets") is conceptually similar to F1N05, but the numbers come from different sources, are computed differently, and use different category boundaries. The scoring code branches on which form an institution filed.

State appropriations create a structural asymmetry. Under GASB, state appropriations are classified as non-operating revenue, sitting below the operating-margin line in the income statement. Under FASB at private institutions, donor-funded analogs typically flow through unrestricted operating revenue — they're already in the operating margin. A public university running a 0% operating margin is often perfectly healthy because state appropriations cover its operating shortfall. A private college running a 0% operating margin has nothing else to fall back on. The same number signals very different fragility states.

Off-balance-sheet debt is far larger for publics. Public-university debt is fragmented across affiliated entities (university foundations, athletic associations, real estate corporations, hospital operating companies, state bond authorities, system-level boards of regents) that are functionally part of the institution but file separately and do not appear on the institution's IPEDS F1A reporting. For a fragility model that uses only IPEDS-reported debt, leverage is systematically and severely understated — by 2-10× for major systems like UC, UT, Wisconsin, Florida.

Cohort and Scope

Sector: Public four-year institutions (CONTROL=1, ICLEVEL=1) that file F1A. Approximately 800 institutions across community-college-adjacent regional campuses, mid-tier regional publics, R1 flagships, and state-system campuses.

Calibration: No closed-school cohort; major publics rarely close. The model is calibrated against bond rating tier expectations (Moody's Aa-equivalent → Aa-band scoring outcome) and validated by spot-check against well-known stress patterns (UPR/Puerto Rico, Illinois publics under the budget-impasse years, post-Bolling cuts at NC publics).

Time range: Same as privates — FY 2004-05 through FY 2022-23.

Data Sources (Publics)

Primary IPEDS:

• Finance F1A component (GASB publics) — https://nces.ed.gov/ipeds/datacenter/data/F[YY1][YY2]_F1A.zip.
• Instructional Activity (EFIA) — same as privates.
• Header (HD) — same as privates.

LLM-enriched: Off-balance-sheet (DSO) debt. A separate enrichment layer queries Claude Sonnet 4.6 with web search, asking for total off-balance-sheet long-term debt held by entities affiliated with each public institution. Required primary-source citations: audited financial statements, EMMA bond filings, state bond authority annual reports. Output is per-entity ("FSU Athletic Association: $326M, citing flbog.edu/Project-Summary-1.pdf"), rolled up into two totals per institution: DSO_DEBT_CAMPUS (debt of campus-specific entities) and DSO_DEBT_SYSTEM (debt of system-level entities, reported only at the system flagship to avoid double-counting). The full citation index is at /sources.

Inflation Adjustment (Publics)

Same as the private model — all dollar quantities that span time are CPI-U adjusted to base year 2024 before scoring. The pricing-power pillar's real_net_tuition_per_fte, the debt-acceleration penalty's 5-year real-debt comparison, and the tuition-collapse override all run on real dollars. Snapshot quantities (FY 2022-23 leverage, FY 2022-23 days cash, etc.) are nominal at the as-of date because no time comparison is involved. See Section 4 under the private model for the formula and the CPI-U-vs-HEPI tradeoff.

IPEDS Variable Map (F1A / GASB)

The Four Pillars (Publics)

Pillar 1: Operating Margin (Moody's-aligned bands)

Raw metric: (F1N01 + F1B11) / (F1N02 + F1B11) for FY 2019-20 onwards — operating gain/loss plus state operating appropriations, divided by operating revenue plus state operating appropriations. Pre-FY 2020 falls back to a (F1D03 − ΔEndowment) / (F1D01 − ΔEndowment) proxy with the same endowment-volatility strip used on the private side.

The "include state appropriations" detail matters: under GASB they sit in the non-operating revenue block, but Moody's higher-ed scorecard treats them as part of the operating-revenue base when calibrating its band thresholds. Using a strict operating-only ratio (F1N01 / F1N02 alone) produces -50% to -200% readings for state-funded publics that the bands aren't designed for — a category mismatch the model previously had and now corrects.

Scoring (Moody's-aligned bands — roughly 2-3× lower than private bands):

A public running 4% margin scores 5 (workable); a private running the same margin scores 22 (disaster). Both readings are correct for their respective sectors — privates have no state-appropriation cushion to fall back on, publics often do.

State-appropriation-share modifier. When F1B11 / F1N02 > 0.30 (state appropriations exceed 30% of operating revenue), the Cash Position score is capped at 5 (workable) regardless of how strong the headline margin looks. A margin propped up by heavy state-appropriation reliance can reverse fast in a state fiscal downturn — Moody's applies a similar qualitative state-support adjustment in committee before assigning a final rating.

Trajectory smoothing. Same "worst-of (current, rolling avg)" rule as privates: 3-year window for FY 2019-20 onwards (clean F1D-based margin), 5-year pre-2020 (legacy proxy needs more smoothing).

Pillar 2: Pricing Power

Raw metric: real_net_tuition_per_fte = F1B01 × cpi_factor / total_FTE. Trajectory measured over 5 years with the same band structure as privates.

No discount-rate add-on for publics. GASB doesn't break out institutional grants the same way FASB does (no F1 analog of F2C05 + F2C06).

Caveat — weaker signal for publics. Public tuition rates are often set by the state legislature or system board of regents, sometimes capped by statute. A flagship public's "pricing power" is partly determined by political conditions in Sacramento or Tallahassee or Madison, not by what the market will bear. We still compute the same metric but a flat reading might mean the legislature froze tuition rather than that the market rejected the value proposition.

Pillar 3: Debt Burden (with DSO debt)

Raw metric: Viability ratio = expendable net position / (long-term debt + DSO debt).

viability = F1N05 / (F1A10 + DSO_DEBT_CAMPUS + DSO_DEBT_SYSTEM)

The DSO additions are the central correction over the prior mvp_v3_public method. Without them, leverage was understated by 2-10× for major public-university systems. Florida statute requires its publics to use Direct Support Organizations for almost all bonded debt; Texas A&M's debt is held by the system Board of Regents (RFS Bonds + PUF Bonds = $4.76B as of FY 2023-24, across 12 campuses); Wisconsin's debt is held at the state level via GO Bonds for UW System Academic Facilities ($6.62B over 13 campuses); UC and CSU debt is held by the system Regents/Trustees, separately from each campus.

DSO_DEBT_SYSTEM is reported only at the flagship UNITID of each system to avoid double-counting across sibling campuses (a regex-driven dedup map identifies flagships for ~50 systems including UC, CSU, UT, TAMUS, PASSHE, CUNY, UW System, OSU, Penn State).

Scoring bands: Same CFI-aligned table as privates (0 at ≥3.0 viability ratio, 25 at <0.5 or negative). Same debt-acceleration penalty applies, including DSO debt in the year-over-year comparison.

Pillar 4: Liquidity

Raw metric: Days cash on hand = F1N05 / (F1C191 / 365) (FY 2019-20+) or F1A17 / (F1C191 / 365) pre-2020.

Same band structure as privates (0 at ≥180 days, 25 at <30 or negative). Same tuition-collapse override (real net tuition drop > 30% over 5 years caps minimum liquidity score at 16).

Composite Score (Publics)

Same composition as privates:

composite_final = clamp(0, 100, P1 + P2 + P3 + P4 + compounding_bonus + endowment_backstop)

The compounding bonus (+10 if 3+ pillars worsened over a 3-year window) applies identically. The endowment backstop runs on F1H02 instead of F2H02 but uses the same burn-ratio logic. A caveat for publics: many public-university foundations hold the endowment off the institution's balance sheet, so F1H02 is sometimes incomplete. A future enhancement supplements with foundation-reported endowment values via the same LLM-enrichment pattern.

Same buckets: 0–30 low, 31–60 moderate, 61–100 high.

Score Versioning (Publics)

• mvp_v3_public — initial public model: GASB variables, Moody's-aligned bands, state-appropriation-share modifier. Used IPEDS-reported debt only — systematically understated leverage.
• mvp_v3_public_dso — current. Two corrections over the prior method:
  1. Adds LLM-enriched DSO debt to the Debt Burden pillar. For systems like UC and Wisconsin, the viability ratio dropped from "comfortable" to "moderate-to-tight" once off-balance-sheet debt was incorporated.
  2. Switches the Cash Position numerator/denominator to include state operating appropriations. Earlier versions used F1N01 / F1N02 (excludes appropriations entirely), which produced extreme negative readings outside the Moody's-aligned bands' calibration range. Re-scoring on this fix shifted the public bucket distribution from heavily skewed-high to roughly 21% low / 46% moderate / 32% high.

Public-finance idiosyncrasies

Public-university financial reporting carries several structural quirks that the model has to handle carefully — straight ports of the private methodology produce category errors otherwise.

State-held debt is not on the campus's balance sheet. Florida statute requires its publics to use Direct Support Organizations and the Board of Governors bond pool for almost all bonded debt; California's UC and CSU debt is held by the system Regents/Trustees, not the individual campus; Texas A&M's debt is held centrally at the System Board of Regents. For these schools, IPEDS F1A10 shows $0 every year — not because there's no debt, but because the campus filing legitimately doesn't carry it. Our LLM-enriched DSO debt layer (see /sources) tries to capture this, but coverage is incomplete; a public flagged as $0 debt with no DSO citations is best read as "leverage indeterminate from IPEDS alone," not "no leverage."

GASB 68 (FY 2014-15+) and GASB 75 (FY 2017-18+) put state pension and OPEB underfunding on the campus balance sheet. Public universities participating in state retirement systems must book their proportionate share of the system's unfunded actuarial liability — even though the campus didn't take on the obligation directly and can't be sued for it. For Florida publics with thin actual debt, the FRS pension allocation alone can flip total net position (F1A17) deeply negative. UWF in FY 2023 carries F1M02 + F1M06 = $130.7M in pension+OPEB allocations against F1A17 = -$69.5M; net of the GASB allocation, net position is +$61M positive. Reading F1A17 as institutional leverage in this regime measures FRS underfunding, not the campus.

The model handles these by skipping the legacy viability proxy (F1A17 / F1A13) when F1A10 = 0, no DSO citations, and pension+OPEB explains a meaningful share of total liabilities. In that case viability_ratio is left null and the Debt Burden pillar defaults to a neutral score of 12 with an indeterminate-leverage flag rather than reading pension underfunding as debt distress.

Fiscal-year alignment is clean. State legislatures, state appropriations, and university operations all run July 1 – June 30 in the great majority of states. So a school with an end-of-year "loss" can't be hand-waved away as "appropriations come in tomorrow" — under GASB, the year's appropriations are recognized as revenue when appropriated, which is at the start of the fiscal year. UWF's recurring operating shortfall is real, even if the model now reads its leverage correctly.

Tuition is partly political. State legislatures or system boards set tuition rates, sometimes capped by statute. Florida froze in-state tuition in 2014 and has held it nominally flat since. The pricing-power pillar still flags real (CPI-adjusted) tuition decline as fragility — which is technically correct as a financial signal but doesn't distinguish "market rejected the value proposition" (the meaning for a private) from "Tallahassee said no." This is documented as a known weaker-signal caveat for publics; future work could attach a state-policy adjustment.

Some systems file balance-sheet data centrally, not per campus. The University of California is the canonical case: each UC campus IPEDS filing carries revenue (F1B), expense (F1C), endowment (F1H), and pension/OPEB (F1M) detail, but the entire balance-sheet block (F1A06–F1A18) and the operating-margin summary blocks (F1D, F1N01/F1N02/F1N05) are reported at the UC Office of the President system level instead of per campus. Without those blocks, the standard formulas can't compute the modern operating margin, the viability ratio, or days cash on hand for a UC campus. Three of four pillars would default to score 12 (neutral middle) for the entire decade — visible as a "baseline 36" composite that doesn't move except when pricing power swings.

For these schools, the model uses an all-in operating-margin fallback when both the modern F1N block and the legacy F1D block are missing but top-line aggregates are present:

operating_margin_all_in = (F1B27 − F1C191) / F1B27

That is, total revenues (including state appropriations, investment returns, gifts, and capital appropriations) less total operating expenses, divided by total revenues. This reads more generously than Moody's strict "operating margin" — it lets investment income offset operating shortfalls, where Moody's framework would not — but the trade-off is accepted because the alternative is a default-12 neutral score that hides real signal. For UC Berkeley FY 2022-23 the fallback gives (3.52B − 3.61B) / 3.52B = -2.6%, which lands in the Ba "disaster" band per the Moody's-aligned thresholds. Real, defensible, and based on actual reported numbers rather than a placeholder.

The Debt Burden and Liquidity pillars stay indeterminate (default 12) for these schools — there is no per-campus expendable-net-position figure to compute either ratio from. Affected campuses carry an explicit "data-quality note" callout on their school detail page so users know which pillars are real signal and which are defaults.

Known Limitations (Publics)

1. Pension and OPEB liabilities (F1M02, F1M06) are ingested but not yet incorporated into scoring. For Illinois, New Jersey, and Connecticut publics especially, these liabilities can dwarf operating budgets. Phase 2.
2. Foundation-held endowment is invisible. The endowment backstop runs on F1H02 only. Where a public foundation holds significant restricted assets off the institution's balance sheet, the backstop is understated.
3. No closed-school calibration cohort. Major publics rarely close, so the model is calibrated against rating-agency tier expectations rather than predictive lookback against announced closures.
4. DSO debt enrichment is point-in-time. Citations are dated and reflect the audit year stated in the source — not necessarily the same fiscal year as IPEDS reporting. Treated as the most recent debt totals available; refreshed annually with the IPEDS update cycle.
5. State-related FASB filers (Penn State, Pitt, Temple, Lincoln) score on the private model. They are CONTROL=1 in IPEDS but file F2 (FASB), so the public path doesn't apply. This is methodologically defensible (their financials are FASB) but means the "public" leaderboard excludes them.
6. Pricing-power signal is weaker than privates. Tuition decisions are partly political. A flat reading does not necessarily mean market rejection — it may mean a legislative tuition freeze.

Debt per student (cross-sector)

Total long-term debt divided by full-time-equivalent enrollment — surfaced on every school page and as three leaderboards at /debt: aggregating systems, per-public school, per-private school. Cross-sector by design: same denominator concept (FTE), but debt composition and the right unit of aggregation differ by how the school carries it.

What's in "total debt"

For both sectors, the numerator combines what IPEDS reports on the institution's books with what the LLM-enrichment layer found off them:

• Privates (FASB / F2): F2I06 plant debt (FY 2019-20+) or legacy F2A03A. Privates rarely have material DSO debt — most of their bonded debt is on-balance-sheet.
• Publics (GASB / F1A): F1A10 long-term debt + DSO_DEBT_CAMPUS_ACADEMIC (campus-specific off-balance-sheet bonds) + FTE-weighted share of DSO_DEBT_SYSTEM_ACADEMIC (the system-pool bonds, allocated across siblings — see disaggregation below).

The _ACADEMIC suffix matters: those variants exclude medical-center / hospital bonds from the DSO totals. See "Medical and hospital debt excluded" below.

What's in the FTE denominator

Sum of three IPEDS enrollment measures, all from the 12-month EFIA survey for the matching academic year:

• EFTEUG — estimated undergraduate FTE (credit-hours based)
• EFTEGD — estimated graduate FTE
• FTEDPP — reported doctor's professional-practice FTE (medical residents, law students, pharmacy students, etc.)

Including FTEDPP matters for medical/law/pharmacy-heavy institutions; without it, a school like UCSF (purely graduate/professional) would have an artificially small denominator. FTEDPP is only available FY 2016-17 onwards — earlier years use UG + grad only.

Disaggregating system bond pools

The biggest methodology problem: most large public university systems hold their bonded debt at a single legal entity (the system Regents, a state bond authority, etc.) that funds operations across many campuses. Examples:

• University of California Regents — $42B in system bonds funds 9 academic campuses + 5 medical centers
• Texas A&M System — Revenue Financing System + PUF Bonds, ~$5B across 12 campuses
• University of Wisconsin System — state GO Bonds for Academic Facilities, ~$5.4B across 13 campuses
• California State University Trustees — system revenue bonds across 23 campuses
• Pennsylvania State System of Higher Education (PASSHE) — system bonds across 10–14 schools post-consolidation
• Similar patterns at CUNY, SUNY, UMass, Indiana, Ohio State, Penn State, Minnesota, Missouri, Hawaii, etc.

Our prior dedupe pass attributed the entire pool to a single "flagship" UNITID (the sibling with highest FTE) to avoid double-counting in aggregate stats. But for per-student debt, putting $40B of UC system bonds onto UC Berkeley's row alone produces a wildly misleading $900K-per-student number — the bonds actually fund 290,000+ students across 9-10 campuses.

The fix is FTE-weighted allocation across all sibling campuses:

this_school's_share = total_system_pool × (this_school's_FTE / sum_of_sibling_FTEs)

The system pool is computed as the SUM of DSO_DEBT_SYSTEM_ACADEMIC across all siblings, not just the flagship — this catches cases where the LLM enrichment labeled some bonds as "Berkeley campus share of UC Regents bonds" rather than as a UC Regents row, leaving them at non-flagship UNITIDs.

28 system families are currently curated, covering 178 sibling UNITIDs:

Families are auto-derived from dso_citations by matching each citation's entity_name against the dedupe regex patterns; a system is included once at least 2 sibling UNITIDs have citations matching its pattern. Re-running scripts/generate_system_families.py refreshes the list as enrichment coverage grows.

Medical and hospital debt excluded

A separate pre-aggregation step (scripts/compute_dso_academic.py) writes DSO_DEBT_SYSTEM_ACADEMIC and DSO_DEBT_CAMPUS_ACADEMIC rows that exclude bonds funding hospital or medical-center operations. These bonds are serviced primarily by patient-care revenue, not student tuition, so including them in a per-student-debt comparison would systematically over-rank schools with academic medical centers (UCLA, UCSF, UMich, UNC, Penn State, etc.).

A DSO citation is treated as medical/hospital when its entity_name matches any of: medical center, medical centers, hospital, patient care, health system, health sciences center, MCPRB, medical pooled, medical center pooled, clinical revenue, health enterprise.

In the current data, 5 citations totaling $15.88B were excluded as medical. The biggest single one is the UC Medical Center Pooled Revenue Bonds (MCPRB) at $13B; without exclusion, every UC campus's allocated share would carry ~$3B of medical debt that has nothing to do with educating their undergrads.

Auxiliary debt (housing, parking, dining, student unions) is kept in academic — these directly serve students even if they're separately funded by user fees. This could be revisited in a future iteration.

Netting against IPEDS-reported debt

The LLM enrichment sometimes returned debt that's already on the IPEDS balance sheet (e.g. New College of Florida's $19M Certificates of Participation matches its F1A10 of $19M exactly). Without correction, the per-student calculation would add those amounts twice — once as IPEDS plant debt, once as DSO campus debt.

The fix nets DSO campus totals against IPEDS plant debt at aggregation time:

off_balance_sheet_campus = max(0, LLM_campus_total − IPEDS_plant_debt)

Interpretation: if the LLM total is ≤ IPEDS, the LLM was capturing on-balance-sheet debt that's already reported, so the genuine off-balance-sheet component is zero. If the LLM total exceeds IPEDS, the excess represents genuinely off-balance-sheet debt that IPEDS doesn't capture.

Applied to the current data: 76 schools were netted, removing $11.74B of double-counted on-balance-sheet debt from the DSO totals. Examples: NCF (LLM $19M ≈ IPEDS $19M → off-balance = $0), UConn (LLM $2.03B ≈ IPEDS $2.17B → off-balance = $0), UF (LLM $650M ≤ IPEDS $757M → off-balance = $0).

System pool debt (DSO_DEBT_SYSTEM_ACADEMIC) is not netted — by design, system bonds are off the individual campus's balance sheet, so they don't overlap with F1A10.

Three leaderboards

The /debt page splits rankings into three views because systems and individual schools aren't directly comparable:

1. Aggregating systems — one row per system family. Used for systems where the centralized bond pool exceeds the sum of sibling campuses' own reported debt, i.e. the system actually pools debt centrally (UC, UW System, TAMU System, CSU, OSU, UMass Building Authority, UT System, Missouri, Minn, PASSHE, Illinois, Indiana, MSCBA, Hawaii, Maine). For these, the meaningful unit is the system: per-FTE share = (pool + summed sibling own debt) / total sibling FTE.

2. Per-public school — one row per public university whose debt sits on the campus's own books rather than aggregating at a system level. Includes standalone publics and members of non-aggregating families (NJEFA, SUNY, CUNY, USG, FL SUS as currently captured, etc.). Schools in aggregating families are intentionally absent here — they're represented in the system view, so a row in both would double-bill the same debt.

3. Per-private school — one row per private nonprofit. Privates don't have system structures, so per-school is the natural unit.

A family is classified as "aggregating" when system_pool > sum(sibling_own_debt). With the current data, 15 of 28 curated families qualify. The other 13 have campus-held debt and are routed to the per-school view.

Schools that have a non-zero DSO_DEBT_SYSTEM_ACADEMIC row but no curated family at all are excluded from all three lists — their pool would be wrongly attributed to a single campus. As enrichment coverage grows, more schools become eligible to appear.

Schools with zero debt across all sources (genuine zero or genuine data gap) are also dropped.

Specialty and graduate-only institutions are excluded from the per-school leaderboards via two complementary filters:

1. UGOFFER = 2 (graduate-only) — schools that don't offer undergraduate education. Catches institutions whose Carnegie classification is missing or non-obvious: Albert Einstein College of Medicine, Rockefeller University, Western Michigan Homer Stryker School of Medicine, theological seminaries (Princeton Theological Seminary, Jewish Theological Seminary, etc.). 235 privates and 15 publics filtered.
2. Carnegie Basic 2021 codes 25, 26, 31 — specialty undergrad-offering institutions whose F1A10 carries hospital or specialized infrastructure that distorts per-FTE comparisons:
   • 25 Medical Schools & Centers (UCSF, OHSU, UMass Chan Medical)
   • 26 Other Health Professions Schools (Charles R Drew, Clarkson College of Nursing)
   • 31 Law Schools (UC Hastings, California Western School of Law)

The DSO medical-debt exclusion above only filters off-balance-sheet bonds; these two filters handle the on-balance-sheet equivalent for whole-institution specialty schools.

Note: Carnegie 24 (Faith-Related), 27 (Engineering/Tech — Rose-Hulman, Webb), 28-29 (Business — Babson), 30 (Arts/Music/Design — Juilliard, San Francisco Conservatory), 32 (Other Special Focus), 33 (Tribal) stay in — these include legitimate undergraduate-degree-granting institutions.

3. FTE minimum of 200 — schools with fewer than 200 full-time-equivalent students are excluded. At small enrollments, per-FTE ratios produce noise rather than signal (e.g. a school with 9 FTE and $2M in debt reads $222K/FTE, but that's not directly comparable to a school with 9,000 FTE at the same per-FTE figure). Trims tiny seminaries, micro-institutions, and other edge cases. Doesn't materially affect the top of either list.

Known limitations

1. DSO snapshot, not history. DSO citation values reflect the audit year stated in the source (typically FY 2023-24). They're applied only to the latest year of debt-per-student. Historical trajectory on each school page uses IPEDS-reported plant debt only — for UC schools (no per-campus balance-sheet) that trajectory is empty.
2. Schools missing DSO citations show as low-debt even when they have significant system debt. Examples in this category include some FL SUS schools that weren't fully enriched (FAMU reads $0 currently). The model trusts what the LLM enrichment captured.
3. Family coverage is incomplete. Even within a curated family, schools without any matching DSO citation aren't included in the sibling list. UC Riverside and UCSF, for instance, are missing from the auto-derived UC family in the current snapshot — their FTE doesn't enter the allocation denominator, slightly over-allocating to the other 8 siblings.
4. Medical-debt regex is heuristic. A more rigorous approach would re-enrich citations specifically to ask "is this revenue serviced by tuition or by patient/auxiliary revenue?" — that's a larger LLM pass deferred for now.