Agentic Discovery Tools and Trust in Industrial Formulation Science

Industrial chemistry encompasses the material world, including paints and coatings for walls, glass adhesives, phone adhesives, coatings for eyewear, and personal care products.
Formulation science is identified as the largest domain within industrial chemistry.
Approximately more than half of the $200,000,000,000$ annual Research and Development (R&D) spend is directed specifically toward formulated products.
Despite its economic scale, formulation science is the least accessible domain for agentic coscientists.
The primary barrier to entry is that the majority of success-determining knowledge exists in a tacit, unpublished, and proprietary form that is rarely shared in the public sphere.
Currently, coscientists assist partners in inventing products that are utilized by over $1,000,000,000$ people daily.

Trust is the most critical feature when building AI for scientific discovery, transcending abstract accuracy or isolated high-performance capabilities.
In this context, trust is defined as alignment with the specific professional domain of the user.
Chemists responsible for material market segments reaching $1,000,000,000$ people have exceptionally high expectations for the systems they use.
The breakdown of trust occurs when a system begins to reason in a "confidently wrong" manner.
When an agentic system fails, the user does not merely lose trust in the specific answer provided; they lose trust in the developer of the system.
The reputational damage caused by these failures is difficult to recover from.

Trust often breaks down due to simple, quiet misalignments with the domain rather than dramatic failures in complex reasoning.
A real-world example involves a formulation chemist searching for a specific patent for an agent to reason with.
In a comparison of two passages that appeared relevant to the query, one was correct and the other was not.
A leading embedding model ranked the correct answer in the $300^{\text{th}}$ position, placing it well outside the range typically checked by production re-ranking systems.
The similarity scores between the correct passage and the incorrect passage were effectively identical, with a difference of nearly $0$ .
This indicates that standard models lack "within-domain discrimination" for formulation chemistry.
For an agentic coscientist, a poor retrieval doesn't just result in a missing fact; it creates a "false belief" from which the agent begins to reason with high confidence.

To solve discrimination errors, it is necessary to examine the underlying factors that distinguish material passages.
While two passages may share similar material ontologies, functions, and final application domains, they contain nuances that a human formulation chemist would use to differentiate them.
By utilizing ontological signals, developers can create powerful training signals for within-domain discrimination.
Benchmarking reveals that commercial APIs typically score near $0$ on within-domain discrimination tasks in industrial chemistry.
Open-based models perform slightly better, but continued pre-training on general chemistry text fails to yield significant improvements.
The most effective solution is task-specific, ontologically structured contrastive training.
This method teaches the model the actual "shape" of domain chemistry and aligns it with the cognitive processes of actual users.
Implementing this approach can move a model from a near-zero performance level to a state that is meaningful for enterprise-scale operations.

Reasoning and discovery in high-uncertainty environments are only as effective as the alignment between the system and the problem space.
Scalable AI solutions must be "domain-native" to understand specialized scientific fields properly.
The public record of science is limited by survivorship bias, often missing the failures or nuanced data points that exist below the surface.
Key institutional knowledge currently resides within enterprise R&D departments.
Building trust with end-users is the primary mechanism for gaining access to this proprietary, high-value data, which is essential for the next generation of agentic systems.