Research Programme Overview
The TOTALFLOW Compliance Intelligence Engine is a research platform that evaluates whether multimodal large language models can reliably apply organisational expense policy to real-world receipts. Each experiment is a self-contained page — open the one you want to run.
E1: OCR Extraction Baseline
Establish a baseline for OCR extraction accuracy by capturing all structured information from real-world receipts.
E2: Direct Policy Reasoning
Apply organisational policy directly to structured accounting data without structured reasoning.
E3: Structured CoT
Test whether structured chain-of-thought reasoning improves policy interpretation and verdict accuracy.
E4: Context-Sensitive Reasoning
Evaluate how project phase context affects the same transaction's compliance outcome.
E5: Uncertainty Handling
Test whether the model recognises ambiguity and avoids forced confident verdicts on incomplete evidence.
E6: Adversarial Stress Test
Run the same transaction against all three policy variants simultaneously to surface policy-quality issues.
Ground Truth Annotation
Annotate each receipt with a human verdict and supporting reason. The reference dataset for evaluation.
Policy Editor
Define and version the three controlled policy variants used by E2–E6.
Research Configuration
Configure the API key, models, currency treatment and saved-result controls used across all experiments.