SOTA Overfitters

What it is EquiAlgo diagnoses regional bias in student-funding decisions and produces a transparent corrected allocation, a reproducible audit and a governance plan. The institution and data are synthetic. Historical committee awards describe past behavior, not who deserves funding.

What we deliver The public repository contains the four required root files: predictions.csv, audit_rapport.ipynb, model_corrige.py and presentation.pdf. An English explanation with vector diagrams connects the diagnosis, policy, fitted score, allocation, evaluation and oversight. The five-minute presentation maps to the official rubric.

The final file contains 4,000 original IDs in their supplied order, with 1,600 awards (40%). The five-region selection-rate range is 0.4851 percentage point; Centre/Remote difference is 0.2279 point. These observable rates do not establish hidden-reference equal opportunity.

Diagnosis and policy Historical committee award rates are 48.37% for Centre and 27.30% for remote regions. Removing region and postal code leaves strong proxies: a retrospective region probe reaches 98.65%, versus 60.05% for the majority class. With only permitted score features it still reaches 77.25%. We also audit R-score bands, study programmes, first-generation status and intersections. These are observational findings, not causal proof.

Our declared policy lets academic achievement help and prevents lower income, more work or first-generation status from lowering a score at otherwise equal inputs. Distance, postal code and programme do not enter the score. We choose observable five-region selection parity because equalizing historical committee positives could preserve historical exclusion. The choice and its costs are explicit and appealable.

The method actually used A regularized logistic regression fits historical committee labels using nonnegative slopes and regional nuisance offsets. At inference it averages the five regional predictions equally. The fitted R and work-hour slopes are positive; income-need and first-generation slopes reach zero. We do not claim a financial-need bonus that the fit did not learn.

Five bounded model families were compared across three 60/20/20 historical splits. Calibration selected the monotone logistic family and the strictest feasible parity setting costing at most 1% of unconstrained allocated score in each calibration. The selected 0.5-point bound is an internal policy, not an official sponsor threshold.

An exact integer allocator maximizes quantized score while jointly satisfying the 40% award count and regional rate range. Its Pareto sweep uses the same selected scoring method and applies the real allocation at every point. There is no later budget repair that breaks the constraint.

Evidence and limits The corrected implementation passes 46 tests, including an independent exhaustive allocation oracle, small-group infeasibility, all-zero score sweeps, ties, malformed inputs, new applicant batches and deterministic refitting. The executed notebook has six code cells with no errors. A fresh reproduction from the distributable package is recorded with exact hashes.

Retrospective historical agreement is 84.75%, but the committee-label TPR gap remains 17.961 points and FPR gap 13.664 points. Parity is not equal opportunity. Assumed academic/need sensitivities demonstrate that different normative merit policies can disagree strongly with our output.

All historical data had previously been explored; new split seeds do not make it unseen. The final evaluation features have no supplied reference labels. Hidden-reference EO, reference utility, portal accuracy/F1 and jury points are unmeasured for this release. No full-mark or new-applicant generalization claim is made.

Governance and contribution Before real use, the institution must approve the policy, appoint accountable owners and collect independently adjudicated outcomes. Batch checks, monthly fairness monitoring, proposed pause triggers, quarterly policy reviews and a meaningful appeal route are specified. An appeal can correct inputs or contest group constraints; a human reversal must transparently reconcile the whole budget.

Earlier evaluation-feedback reconstruction raised an integrity ambiguity. At the user's direction, this release instead excludes all reconstructed labels, anchors, inferred prevalence and earlier score feedback from training, calibration and selection. The old work remains historical locally; it does not validate this model. No organizer approval of the earlier method is implied.

The sponsor schema and earlier contract checks are reused, and source/pre-existing work is disclosed. OpenAI Codex assisted implementation, testing and explanation. Kenneth Chen and Ning Ye remain responsible for review and submission. Raw participant data and earlier inferred records are excluded from GitHub; the new predictions are our own output on synthetic IDs, explicitly required by the public-repository deliverable. HxBuddy and Devpost actions remain manual.

Built With

• jupyter;
• matplotlib;
• nbclient;
• numpy;
• openai
• pandas;
• pymupdf
• pytest;
• python;
• reportlab;
• scikit-learn;
• scipy;