PercepticOS organizes the work of researchers and AI agents around scientific evidence.

PercepticOS pairs researchers and AI agents in a shared research space to analyze hypotheses, cross-examine evidence, and document decisions.

In pharmaceutical research, scientific decisions rarely depend on a single publication or experimental result. Teams must compare potentially conflicting studies, examine internal data, consult specialists across departments, and preserve the reasoning behind the hypotheses they pursue. PercepticOS is designed to organize that work in an environment shared by researchers and artificial intelligence agents. Developed by Perceptic, the platform breaks scientific questions into testable claims, assigns evidence gathering to autonomous agents, and leaves the validation of conclusions to human experts. According to the company, its technology is already used by major pharmaceutical companies and biotechnology firms, with access to PercepticOS available by application. Starting With a Decision, Not a Conversation

Most AI assistants respond to questions by generating text. For complex scientific research, that format has limitations: citations may be difficult to verify, uncertainty can become mixed with conclusions, and successive conversations do not necessarily produce a research record that an entire team can reuse.

PercepticOS organizes the process differently. A researcher defines a question, identifies relevant public or internal sources, and chooses an AI model and reasoning depth. The system then creates a structured study divided into sections and claims precise enough to be examined against available evidence.

These claims are connected through relationships of dependency. A biological hypothesis may support a proposed therapeutic intervention, which in turn depends on clinical findings or the feasibility of identifying an appropriate patient population. Agents examine both the claims and their relationships, identifying evidence that supports or challenges them.

Researchers can revisit an individual claim without reconstructing the full history of a conversation. Each statement retains its status, sources, qualifications, and the human decisions that shaped its assessment. Autonomous Agents Explore Multiple Hypotheses

PercepticOS agents can investigate an individual claim or an entire section of a study. They search scientific literature, consult databases, and execute analyses in parallel. The platform also describes combining deductive reasoning from a hypothesis with inductive reasoning from the available data.

When evidence conflicts or does not support a clear conclusion, an agent can pause and ask a specialist to intervene. The request presents the available options and explains how the response could affect the study. That decision is subsequently recorded alongside the claim.

Perceptic emphasizes this division of responsibility. Agents conduct part of the literature review, verification, and analysis, but they do not have final authority to validate scientific conclusions. Researchers remain responsible for evaluating the evidence and deciding what can be accepted.

This also addresses an organizational challenge in research and development. Analyses frequently span departments, software systems, and research teams, with important reasoning becoming difficult to recover once a project moves forward. PercepticOS preserves revisions and their justifications so that subsequent studies can build on earlier work. Every Claim Links Back to Its Evidence

Evidence provenance is central to the platform's design. Claims can link to scientific articles, specific figures, quoted passages, or internal experimental results. Researchers can inspect those sources, understand their limitations, and determine whether they justify the conclusion under consideration.

Contradictory evidence remains visible as well. A finding that does not support the hypothesis is not automatically excluded from the synthesis. Agents are expected to present agreements, disagreements, and gaps in the available evidence for human review.

Perceptic illustrates this workflow with a study examining clonal hematopoiesis, interleukin-6 signaling, and cardiovascular risk. The example compares published research, including a study by Bick and colleagues in Circulation, to distinguish observed genetic associations from the clinical effects that might follow from an intervention.

Some evidence suggests a relationship between genetic variations affecting IL-6 signaling and cardiovascular outcomes in people with certain blood-cell mutations. However, those observations do not establish that an IL-6-targeting treatment would reduce cardiovascular events in that population.

This study is an illustrative product demonstration, not a clinical investigation conducted by PercepticOS. Its characters, decisions, and workflow events are fictional. The cited publications are real, but the scientific questions represented in the example remain open. Higher Scores Than the Underlying Model on Two Benchmarks

Perceptic has evaluated its system on BioMysteryBench and DrugDiscoveryBench, benchmarks involving scientific reasoning and drug discovery. For these experiments, it used GPT-6 Astra at high reasoning effort within its own agent framework.

On the Human-solvable portion of BioMysteryBench, PercepticOS reports 97.1%, compared with a published reference score of 89% for GPT-6 Astra. On Human-difficult problems, it achieves 43.8%, compared with 32%. On DrugDiscoveryBench, the platform scores 80.5%, against 68.7% for GPT-6 Astra.

These differences approach twelve percentage points in some categories. They suggest that breaking problems into claims, using tools, and coordinating multiple reasoning steps may improve results over those of a foundation model operating within another environment.

The comparison nevertheless has important limitations. PercepticOS results come from an internal evaluation, while baseline scores are drawn from external publications using potentially different settings and agent frameworks. The figures do not represent a controlled comparison of otherwise identical systems or establish general scientific superiority.

They also do not