💻 computer science

A Deterministic Framework for Pharmaceutical Authenticity Verification Using Constrained Multimodal AI

The paper introduces MedTruth, a deterministic dual-mode framework that constrains a multimodal AI model with fixed parameters and strict output schemas to reduce hallucinations from 15% to under 2%, enabling reliable pharmaceutical authenticity verification and safe patient interaction.

Original authors: Viral Dhoka, Pankaj Kunekar, Tanaya Thorat, Shreenath Utekar, Samruddhi Wankhede

Published 2026-10-13

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Viral Dhoka, Pankaj Kunekar, Tanaya Thorat, Shreenath Utekar, Samruddhi Wankhede

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Every year, millions of people around the world buy medicine that looks real but is not. In many places, a bottle of pills might contain nothing but chalk, or worse, a dangerous industrial chemical. While governments and large companies use complex tracking systems to stop these fakes from entering the supply chain, the person standing in a small shop with a suspicious package has no way to check if it is genuine. Recently, people have started asking general artificial intelligence chatbots to help them decide. They upload a photo of the medicine and ask if it is safe. However, these chatbots are designed to be helpful and conversational, not forensic. They often guess confidently, inventing details that are not there, or giving vague answers that do not help a patient in a moment of crisis.

Researchers at the Vishwakarma Institute of Technology in India have built a new system called MedTruth to solve this specific problem. Instead of trying to make a smarter chatbot, they built a system that forces a standard artificial intelligence to act like a strict inspector. The team took a powerful visual language model—one that can see images and read text—and wrapped it in a set of rigid rules. These rules prevent the computer from chatting or guessing. Instead, the system is forced to look at the medicine package in four specific steps: reading the text, checking the design details, comparing the drug name to its ingredients, and looking for official government marks. Only after completing this checklist does the system give a final score.

The results show that this rigid approach works far better than letting the computer talk freely. When the researchers tested the system on fifty different medicine packages, including thirty real ones and fifteen cleverly made fakes, the new system correctly identified the vast majority of them. It caught nearly all the genuine products and most of the fakes. More importantly, the system stopped making up facts. In a test with a standard, unregulated chatbot, the computer invented fake security features or misread patterns about fifteen percent of the time. With the new MedTruth system, that error rate dropped to less than two percent. The system also remained fast, taking only about two seconds to scan a photo and give an answer, which is quick enough for a person to use while standing in a store.

The secret to this success lies in how the system handles the artificial intelligence. The researchers split the tool into two separate modes that never mix. The first mode is a stateless scanner, meaning it has no memory of past conversations. It looks at the image, follows its strict checklist, and outputs a simple score. This prevents the computer from being influenced by what a user said earlier. If the package passes this strict check, the user can then switch to a second mode, a conversational assistant that can answer questions about the medicine. This assistant is only allowed to talk about the specific drug that just passed the test, and it is programmed to never give medical advice or tell a patient to stop taking a prescription.

By forcing the computer to slow down and follow a step-by-step process, the researchers turned a tool designed for open-ended conversation into a reliable verification instrument. The study suggests that for safety-critical tasks like checking medicine, the best way to use artificial intelligence is not to make it more creative, but to make it more constrained. The system does not replace laboratory testing or official government records, but it offers a practical way for ordinary people to spot obvious fakes when they have no other option. The researchers note that while the system is highly effective in their tests, it still relies on the quality of the photo and the presence of visible text, and it cannot verify if a license number is truly real if a counterfeiter has copied the format perfectly. Nevertheless, the work demonstrates that with careful design, artificial intelligence can be made to act with the consistency and precision required to protect public health.

Technical Summary: A Deterministic Framework for Pharmaceutical Authenticity Verification Using Constrained Multimodal AI

Problem Statement

Counterfeit pharmaceuticals pose a critical threat to global public health, with estimates suggesting up to 30% of medical products in regions with weak regulatory oversight are substandard or falsified. While institutional defenses like blockchain and RFID exist, they often fail to assist the end-user holding a suspicious package in resource-constrained settings.

The paper identifies a specific gap in the use of general-purpose conversational AI (e.g., ChatGPT, Gemini) for drug verification. These models, optimized for fluency and engagement rather than forensic precision, exhibit three recurring failure patterns when asked to judge authenticity:

Visual-Textual Dissociation: Models may read text via OCR-like reasoning but ignore physical cues like typography drift, color mismatches, or missing holographic seals.
False Confidence: Due to next-token generation optimizing for plausibility, models often assert authenticity without evidential basis.
Ambiguity Injection: Models frequently fallback to vague suggestions (e.g., "consult a pharmacist") rather than providing actionable judgments.

These failures are inherent to the probabilistic nature of transformer models, where temperature-controlled sampling prioritizes engagement over binary correctness.

Methodology

The authors propose MedTruth, a dual-mode framework that constrains a multimodal Large Language Model (LLM) to behave deterministically rather than open-endedly. The system utilizes Gemini 1.5 Flash and is implemented as a client-side single-page application to minimize latency.

Architectural Design

The system employs a state router to strictly separate two operational modes, preventing conversation history from biasing forensic scans:

Verification Mode (Stateless): A forensic inspection pipeline where the model acts as a structured classification function.
Dr. AI Mode (Stateful): A conversational assistant that activates only after a package passes verification, maintaining a short rolling memory for dosage and side-effect queries.

Constraint-Based Prompting

To transform the generative model into a scoring instrument, MedTruth applies four specific constraints during the Verification Mode:

Role Definition: The system prompt explicitly defines the model as a deterministic verification engine, forbidding conversational text.
Schema Enforcement: Output is forced into a strict JSON schema containing an authenticity score, safety recommendations, visual anomaly arrays, and regulatory compliance flags.
Temperature Suppression: The sampling temperature is fixed at 0.2 (vs. the creative default of 1.0) to concentrate probability mass on the most likely tokens, ensuring repeatable results.
Visual Chain of Thought: The model is forced through a four-stage sequential reasoning process before generating a verdict:
1. Extract visible text (drug name, batch, expiry) and design elements.
2. Compare observations against expected brand patterns (e.g., hologram refraction).
3. Cross-check the declared drug name against the listed active pharmaceutical ingredient (API) and dosage.
4. Verify mandatory regulatory markers (e.g., Rx symbol, license format).

Authenticity Scoring

The framework calculates a composite authenticity score ( $S_{auth}$ ) using a weighted geometric mean of three vectors:

Visual Trade Dress ( $V_m$ , weight 0.4): Typography, color fidelity, holograms, and layout.
Semantic Consistency ( $S_c$ , weight 0.4): Match between drug name, API, and plausible dosage.
Regulatory Compliance ( $R_c$ , weight 0.2): Presence and format of regulatory markers.

The use of a geometric mean ensures that a high score in one vector cannot mask a critical failure in another (e.g., perfect printing cannot override a wrong active ingredient). Scores map to three labels: AUTHENTICATED ( $\geq 0.75$ ), UNVERIFIED ( $0.50–0.75$ ), and SUSPICIOUS ( $< 0.50$ ).

Key Contributions

Dual-Mode Router Architecture: A design that isolates deterministic scanning from open-ended conversation to prevent cross-contamination of states.
Constraint Prompting Methodology: A combination of schema enforcement, low-temperature sampling, and staged reasoning that measurably reduces hallucination in general-purpose multimodal models without retraining.
Empirical Validation: A comparative study against a standard multimodal chatbot and a traditional OCR pipeline, demonstrating gains in accuracy, reliability, and actionable output.

Results

The system was evaluated on a dataset of 50 pharmaceutical packaging images (30 genuine, 15 adversarial counterfeits with subtle flaws, and 5 low-quality captures).

Accuracy: MedTruth achieved 92% overall verification accuracy, correctly identifying 28/30 genuine samples and 13/15 adversarial samples.
Hallucination Reduction: The hallucination rate dropped from 15% in the standard chatbot baseline (GPT-4o) to under 2% in MedTruth. The single failure in MedTruth occurred on a deliberately blurred image.
Latency: Mean verification latency was 2.1 seconds, placing it within a usable window for consumer tools despite the added structured reasoning steps.
Comparison: The standard chatbot failed primarily on adversarial samples with hidden semantic mismatches, while MedTruth's structured cross-checking successfully caught these errors. Google Lens (OCR only) provided no direct verdict, leaving interpretation to the user.

Significance and Claims

The paper claims that architectural constraints and structured prompting can make general-purpose multimodal AI viable for safety-critical domains where determinism is more important than open conversation.

Reliability over Scale: The work demonstrates that improving reliability does not necessarily require retraining larger models with massive labeled datasets; instead, careful system design and behavior constraints yield significant improvements.
Actionable Output: By converting a generative model into a scoring instrument with a strict JSON output, the system provides clear, actionable decisions (Authenticated/Unverified/Suspicious) rather than ambiguous conversational responses.
Limitations Acknowledged: The authors modestly note that the regulatory vector validates format conformity rather than the authenticity of specific license numbers (the "oracle problem"), and the dataset size (50 samples) is modest relative to global packaging diversity. They position MedTruth as a decision support system rather than a replacement for laboratory testing or professional assessment.

The study concludes that separating the verification process into a stateless, forensic task and a stateful, conversational task allows multimodal AI to be deployed in high-stakes environments with repeatable and trustworthy outcomes.