MAGD-Fraud

Learning AI Reliance Risk for Assurance-Guided Human–AI Routing in Financial Fraud Review

MAGD-Fraud is a research framework for studying when continued reliance on an AI prediction is justified and when human intervention is warranted. Financial fraud review is used as the experimental domain, but the primary research problem is broader: AI assurance for Human–AI decision routing.

Rather than treating model confidence as sufficient evidence of reliability, MAGD-Fraud constructs a case-level AI-reliance risk representation from multiple complementary assurance signals and uses this evidence to support AI retention, expert review, or multi-expert escalation.

Research scope. MAGD-Fraud is an offline research prototype evaluated on the FiFAR benchmark with synthetic experts. It is not a production fraud-detection or financial decision system.

---

Research Motivation

High-stakes AI systems need to answer more than what does the model predict? They must also ask is there sufficient evidence to rely on that prediction for this case?

MAGD-Fraud therefore separates:

Case
  ↓
Base AI prediction
  ↓
Assurance evidence
  ↓
Learned AI-reliance risk
  ↓
Human–AI routing
  ├── Retain AI
  ├── Single-expert review
  └── Multi-expert escalation
  ↓
Final decision + audit evidence

The central contribution is an explicit multi-evidence assurance layer between AI prediction and routing.

---

Final Assurance Representation

The submitted paper uses:

• calibration risk;
• representation-space distance uncertainty;
• local neighbourhood reliability;
• confidence-gated wrong-confident evidence;
• pre-specified interaction terms.

The final 7-term assurance representation is:

1. normalized distance uncertainty;
2. normalized calibration risk;
3. normalized neighbourhood risk;
4. revised wrong-confident risk;
5. calibration × distance;
6. calibration × neighbourhood;
7. confidence × calibration.

A fixed L2-regularized logistic model is used with:

Component	Setting
Regularization C	1.0
Class weighting	balanced
Random seed	42
Low-risk threshold	0.4063
High-risk threshold	0.4410
Escalation panel size	5

The output is treated as a ranking score for AI-reliance risk, not a calibrated probability of AI error.

---

Assurance Signals

Signal	What it measures	Assurance role
Calibration risk	Whether confidence matches historical correctness	Detects poorly calibrated regions
Distance uncertainty	Distance from familiar training structure	Measures case unfamiliarity
Neighbourhood reliability	Historical AI error among similar cases	Measures local reliability
Wrong-confident risk	High confidence gated by independent unreliability evidence	Targets dangerous confident errors
AI confidence	Model certainty	Used as context/gating evidence

A key methodological finding is that simply adding assurance signals is not sufficient. The developmental additive formulation produced near-random failure discrimination. The final model improves this by handling signal scale, orientation, redundancy, and conditional relationships.

---

Evaluation Protocol

MAGD-Fraud is evaluated on the FiFAR financial fraud review benchmark.

Split	Cases	Fraud prevalence
Train	278,627	1.02%
Validation / development	119,323	1.18%
Benchmark test	96,843	1.47%

The assurance model is developed with five-fold out-of-fold evaluation. Fold-local preprocessing and signal transformations are fitted only on the corresponding inner-development data.

The benchmark test split had previously been inspected during diagnosis of the earlier additive formulation. Therefore, the paper treats the five-fold out-of-fold development evaluation as the primary model-development evidence and the frozen benchmark-test result as supporting post-development evidence.

---

Main Results

Assurance discrimination

Evaluation	AUROC	PR-AUC
Calibration risk alone	0.6466	0.0748
MAGD-Fraud, five-fold OOF development	0.6837	0.1269
MAGD-Fraud, frozen benchmark test	0.6912	0.1460

AI-error concentration

Risk fraction reviewed	AI errors captured	AI-error rate	Enrichment
Top 1%	21.45%	31.51%	21.46×
Top 5%	42.55%	12.49%	8.51×
Top 10%	46.13%	6.77%	4.61×
Top 20%	48.24%	3.54%	2.41×

The highest-risk 5% of cases contain 42.55% of all AI errors.

---

Routing Results

Method	Precision	Recall	F1	Cost	AI coverage	Expert	Escalation	Reviewer workload
AI-only	0.5288	0.0385	0.0718	1375.79	100.00%	0.00%	0.00%	0.00%
L2D-Standard	0.6986	0.2451	0.3629	1086.61	96.25%	3.75%	0.00%	3.75%
MAGD-Additive	0.7608	0.1359	0.2305	1237.48	80.43%	0.00%	19.57%	97.83%
MAGD-Additive-Tuned	0.4966	0.0518	0.0938	1358.28	98.05%	1.95%	0.00%	1.95%
MAGD-Fraud	0.9414	0.2927	0.4466	1011.48	84.84%	0.00%	15.16%	75.81%

MAGD-Fraud achieves the strongest F1 and lowest cost-sensitive classification loss among the evaluated methods, but at substantially higher reviewer workload.

The final benchmark policy effectively behaves as an AI-versus-panel router: 84.84% of cases remain with AI, 0% use single-expert review, and 15.16% are escalated to a five-expert panel.

---

Reliance-Oriented Evaluation

Method	Wrong-confident avoidance	Correct rejection	Overreliance
AI-only	0.0000	0.0000	0.0147
L2D-Standard	0.1470	0.0062	0.0085
MAGD-Additive	0.2179	0.0030	0.0117
MAGD-Additive-Tuned	0.0156	0.0004	0.0143
MAGD-Fraud	0.2200	0.0070	0.0077

These metrics distinguish predictive performance from the narrower question of whether the system appropriately avoids unsafe AI reliance.

---

Ablation Findings

Assurance representation	Development AUROC
Calibration risk alone	0.6466
+ normalized distance and neighbourhood evidence	0.6729
+ revised wrong-confident evidence	0.6826
+ pre-specified interactions	0.6837

Most of the gain comes from properly representing complementary assurance evidence; interaction terms provide a smaller additional improvement.

---

Statistical Robustness

Comparison	ΔF1	ΔCost	McNemar
MAGD-Fraud vs AI-only	+0.3748	−364.31	p < 0.001
MAGD-Fraud vs L2D-Standard	+0.0837	−75.13	p < 0.001
MAGD-Fraud vs MAGD-Additive	+0.2160	−226.00	p < 0.001

Paired bootstrap confidence intervals for F1 and cost exclude zero in the favourable direction for these comparisons.

---

Repository Structure

MAGD-Fraud/
├── src/
├── scripts/
├── tests/
├── docs/
├── figures/
├── data/
│   └── outputs/
├── config.yaml
├── requirements.txt
├── ARCHITECTURE.md
├── WORKFLOW.md
└── README.md

Large raw benchmark files, processed datasets, embeddings, per-case decision logs, and archived experiment runs are intentionally excluded from Git.

---

Installation

git clone https://github.com/mahend72/MAGD-Fraud.git
cd MAGD-Fraud

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Windows:

.venv\Scripts\activate

---

Dataset

The FiFAR benchmark is not redistributed through this repository.

FiFAR is publicly available from:

https://doi.org/10.6084/m9.figshare.28351172

After downloading the benchmark, place the files according to the paths expected by config.yaml and WORKFLOW.md.

---

Reproducing the Experiments

Full experiment:

python scripts/run_full_magd_experiment.py --config config.yaml

Run tests:

pytest -q

Key output locations:

data/outputs/final_metrics/
data/outputs/paper_tables/
data/outputs/magd_policy/
data/outputs/audit_pack/
data/outputs/final_reproducible_run/

---

Baselines

The repository includes:

• AI-only;
• numerical-confidence threshold;
• distance-based threshold;
• best-expert and random-expert references;
• independent Learning-to-Defer (L2D-Standard);
• oracle upper-bound analysis;
• developmental additive MAGD variants.

The repository does not claim state-of-the-art performance over every recent deferral architecture.

---

Reproducibility and Scientific Assurance

The pipeline includes:

• frozen configuration and random seeds;
• train/development/test separation;
• fold-local preprocessing;
• validation-derived thresholds;
• no test-label use in deployable routing;
• independent L2D baseline features;
• canonical manifests and configuration snapshots;
• deterministic experiment checks;
• paired statistical comparisons;
• case-level audit evidence;
• scientific guardrail tests.

---

Continuous Integration and Research Assurance

CI is treated as a research-quality gate, not as a production deployment pipeline.

Push / Pull Request
        ↓
Environment setup
        ↓
Dependency installation
        ↓
Syntax/import checks
        ↓
Unit and integration tests
        ↓
Scientific invariant tests
        ↓
Pass / fail

The full FiFAR experiment is not run in standard CI because the benchmark data and large generated artifacts are intentionally excluded from the repository.

No production Continuous Deployment claim is made.

---

Auditability

MAGD-Fraud preserves an evidence chain:

AI prediction
→ assurance evidence
→ AI-reliance score
→ routing decision
→ final decision

The audit layer records assurance signals, routing metadata, selected route, decision rationale, and policy configuration. This demonstrates computational traceability, not validated usability by real fraud auditors.

---

Limitations

• FiFAR uses synthetic benchmark experts, not real fraud investigators.
• Evaluation is offline and benchmark-based.
• The study uses one fraud-review domain.
• The final policy does not exercise single-expert routing on the benchmark test split.
• The assurance score is a ranking score, not a calibrated probability of AI failure.
• Reviewer workload is high because escalation uses a five-expert panel.
• Operational cost parameters require institution-specific validation.
• The benchmark test split had been inspected during earlier development diagnosis.
• Independent temporal and external validation remain necessary.
• Audit completeness does not establish real-world governance usability.

---

Responsible Use

MAGD-Fraud is a research prototype for AI assurance and Human–AI routing. It should not be used as a production financial decision system without independent validation, monitoring, governance, human oversight, workload validation, and domain-specific risk assessment.

---

Code and Data Availability

Code:
https://github.com/mahend72/MAGD-Fraud

FiFAR benchmark:
https://doi.org/10.6084/m9.figshare.28351172

---

Citation

The MAGD-Fraud manuscript is currently under submission.

@misc{magdfraud2026,
  title  = {MAGD-Fraud: Learning AI Reliance Risk for Assurance-Guided Human--AI Routing in Financial Fraud Review},
  author = {Kumar, Mahender},
  year   = {2026},
  note   = {Research software and manuscript under submission},
  url    = {https://github.com/mahend72/MAGD-Fraud}
}

Update this entry after publication with the final venue and DOI.

---

Author

Mahender Kumar
GitHub: https://github.com/mahend72