Open-source early-warning model
Compass: support, not a label
Research and build
Compass is an open-source early warning model built by MANGO students. It suggests check-ins, explains every flag, and is audited for fairness in public.
Prototype: tested on synthetic and public data
Wisconsin tried this first
- Wrong nearly 3 out of 4 timeswhen Wisconsin's system predicted a student would not graduate on time.The Markup, 2023
- 42 percentage pointshigher false alarm rate for Black students than for white students.Wisconsin DPI internal analysis, via The Markup
- No improvementin graduation rates for students it labeled "high risk."Perdomo et al., FAccT 2025
So we built the version we wish existed.
Signals in. Reasons out
School records go in. A plain-language reason to check in comes out. Here is how that compares with Wisconsin's system.
| Issue | Wisconsin DEWS | Compass |
|---|---|---|
| Inputs | Over 40 features, including race and free or reduced-price lunch status | Attendance, behavior, and course performance only |
| Output | Low, moderate, or high risk label | "Check-in suggested" with the specific reasons listed |
| Explainability | Ensemble model, hard to inspect | Interpretable model you can check by hand |
| Fairness checks | Internal analysis, made public through a records request | Built-in audit of error rates by group, published with every release |
| Data location | State system | Runs inside the division. MANGO never sees student records. |
| Code | Not public | Open source |
Meet a synthetic student
Move the sliders or pick a preset. Compass runs right here in your browser, with the same published weights as the open-source code.
Synthetic student
All students here are synthetic. Compass never sees real student data on this site.
Compass output
Check-in suggested
- Missed 18% of school days this grading periodAttendance outreach
What moved the result
Compared with a typical synthetic student. Bars to the right push toward a check-in. The biggest push here: days absent this grading period.
- Days absent
- Suspensions or serious incidents
- Core course failures
- GPA
- Credits behind on-track pace
- Trend: GPA change since last period
Core course failures has a weight of zero in this version: GPA already carries that signal in the synthetic data. The model card explains why.
Same students, two models
A false alarm is a student flagged for a check-in who would have graduated on time anyway. We compare Compass with a baseline built like Wisconsin's, which does use group and income.
False alarm rate by group. Largest gap between groups: 8.0 points for the DEWS-style baseline, 6.1 points for Compass.
- Ranking quality (AUC)
- Baseline0.775Compass0.785
- Precision at capacity
- Baseline43.0%Compass46.2%
- Largest false alarm gap
- Baseline8.0 pointsCompass6.1 points
AUDIT FAILED: Compass's false alarm rate ranges from 7.7% (Group A) to 13.7% (Group C), a gap of 6.1 percentage points (limit 5). Compass does not meet our own limit here yet, so it stays a prototype.
Synthetic data with built-in structural inequality. See the model card for methods.
Removing race isn't enough
Compass never sees a student's group, but other data can carry the same inequality: some groups get more recorded suspensions for the same behavior, and check-ins pile up where school resources are thin. Researchers found the same pattern in Wisconsin (Perdomo et al., 2025). That is a question about counselors and funding, not about any one student.
Methods and full results in the model cardWith and without suspensions
Recorded suspensions vs. true incidents
Per student, on average. In our synthetic schools, Group C gets 1.8x as many recorded suspensions per real incident as Group A.
| Group | True incidents | Recorded suspensions |
|---|---|---|
| Group A | 0.23 | 0.22 |
| Group B | 0.26 | 0.34 |
| Group C | 0.30 | 0.51 |
Compass false alarm rate by group, with the suspensions input. Largest gap: 6.1 points.
Here, dropping suspensions barely moved the gap (6.1 points with it, 6.1 points without). The rest of the gap comes from unequal school resources, which shape attendance and grades too.
Check-ins by school
School information alone ranked students less well (AUC 0.66 vs. 0.78 for Compass).
Each dot is a synthetic school. In the lowest-funded third, Compass suggests a check-in for 24.2% of students on average; in the best-funded third, 5.9%. Correlation between funding and check-in share: -0.95.
Show data table
| School | Funding index | Students per counselor | Check-in share |
|---|---|---|---|
| S24 | 0.61 | 591 | 38.4% |
| S08 | 0.74 | 502 | 27.8% |
| S01 | 0.74 | 484 | 23.4% |
| S27 | 0.76 | 500 | 23.0% |
| S20 | 0.78 | 576 | 30.5% |
| S19 | 0.82 | 545 | 26.2% |
| S10 | 0.84 | 496 | 21.1% |
| S30 | 0.84 | 460 | 19.5% |
| S18 | 0.87 | 441 | 18.0% |
| S05 | 0.88 | 434 | 14.0% |
| S26 | 0.89 | 453 | 16.9% |
| S04 | 0.90 | 480 | 16.9% |
| S22 | 0.94 | 453 | 14.1% |
| S29 | 0.95 | 463 | 16.3% |
| S25 | 0.96 | 416 | 15.0% |
| S17 | 0.96 | 481 | 17.9% |
| S03 | 0.97 | 441 | 16.8% |
| S15 | 1.01 | 420 | 13.0% |
| S21 | 1.02 | 442 | 13.3% |
| S28 | 1.08 | 387 | 8.2% |
| S13 | 1.11 | 346 | 8.1% |
| S02 | 1.11 | 399 | 9.2% |
| S09 | 1.12 | 374 | 8.4% |
| S23 | 1.13 | 341 | 8.5% |
| S14 | 1.20 | 365 | 6.2% |
| S06 | 1.23 | 334 | 6.6% |
| S16 | 1.26 | 324 | 4.1% |
| S07 | 1.32 | 267 | 3.4% |
| S12 | 1.34 | 260 | 2.4% |
| S11 | 1.37 | 266 | 2.0% |
What Compass will never do
- 01
No demographic inputs.
Compass never scores students using race, ethnicity, sex, income, disability, English learner status, or home address.
- 02
Every flag explains itself.
Each result shows, in plain language, which attendance, behavior, or course indicators triggered it.
- 03
Fairness is measured, not assumed.
Compass reports error rates for every student group and publishes the results.
- 04
Support, not labels.
Compass suggests check-ins. It never labels a student and is never used for discipline or placement.
- 05
Student data stays in the school.
Compass runs inside a school division. MANGO never receives student records.
- 06
Open source.
The code, documentation, and evaluation results are public.
Prototype first. Proof later
- Phase 1In progress
Synthetic data
Built: data generator, model, audit, and this demo. Not yet passing our fairness limit.
- Phase 2Started
Public research data
First check on the UCI Student Performance data (649 students, two schools in Portugal). Small and not from Virginia, so a sanity check, not proof.
- Phase 3Planned
Validation with a Virginia school division
Under a formal data agreement consistent with FERPA, a division runs Compass on its own de-identified records and shares only aggregate results. We publish them, including any failures.