Receipt:
E1: OCR Extraction Baseline
Upload a receipt — the model extracts all available fields: vendor details, tax IDs, line items with prices, payment method, and more.
E2: Zero-shot Verdict (No Chain-of-Thought)
Upload a receipt — model gives a direct verdict without step-by-step reasoning.
E3: Chain-of-Thought Reasoning
Same as E2, but the model reasons step by step and shows its work.
E4: Phase-Sensitive Reasoning
Upload a receipt — see how the verdict changes across all 3 project phases side by side.
E5: Ambiguity & Uncertainty Handling
Upload an ambiguous receipt — does the model express uncertainty or force a confident verdict?
E6: Adversarial Policy Stress Test
Upload a receipt — see verdicts under all 3 policy variants simultaneously. Hallucination badges appear when the model misassesses policy clarity.
Ground Truth Annotation
Annotate receipts with human verdicts for paper metrics. Data saved to data/annotations/ground_truth.json.
Policy Editor
Edit the expense policies used by all experiments. Changes take effect immediately.
⚠ Internal contradictions are intentional for Experiment E6. Removing them will break the hallucination detection test.
Settings
Configure the API key and models used for all experiments.
Switch models to compare behavior across E2–E6.
Force a specific currency for verdict prompts, or "Any" to use what the receipt shows.
Experiment Results