A record built to be checked
A preprint proposes a way for small scientific research communities to keep an auditable history of work done by AI agents. Called Symposium, the proposal is presented as both a formal framework and a practical implementation for recording the operation of agents deployed in research communities.
At the center is a CommunityRecord, defined as a publication history made up of a temporally ordered set of Artifacts published by community Members. The design treats the research record as a sequence of contributions that can be inspected over time.
Published Artifacts are immutable. If a result needs correcting, revising or reassessing, the change is represented by a new publication rather than an edit to the earlier Artifact.
The framework also gives scientific arguments a defined structure. Claims are represented alongside the reasoning, evidence and assumptions used to support them, making those parts of an argument visible as separate elements of the record.
Trust depends on the job
Symposium's central idea is that trust should be judged in relation to a purpose, not reduced to a binary verdict or a single score attached to a claim or an agent. The relevant question is whether a piece of work is trustworthy for the task at hand.
The framework also puts a boundary around what may be cited as evidence. Evidence citations are limited to content that has been explicitly declared groundable through a Content Object, a formal way of marking material that can be used for grounding.
Its validation system is aimed at structural problems, including dangling citations, non-unique names and missing properties. It does not decide whether an Artifact is sloppy or whether an Argument represents flawed science; those judgments remain with Members.
That distinction matters because a record can be complete in form without settling the quality of the science inside it. Symposium is designed to expose the material that communities need to examine, while leaving judgments about scientific competence to the people who use the record.
The biology is an illustration, not a result
The paper uses a biological example to show how the record can hold claims, methods and readouts together. But the figures and examples use a synthetic, agent-generated CommunityRecord, and the paper makes no claim about the agent's capability or about the accuracy of the example biology, methods or reasoning.
Within that illustration, proliferation measured by confluence imaging is reported to be 70% to 85% lower over 10 days in the lines tested under suppression than in the uninduced condition.
The same example reports no significant increase in Annexin V staining at day 10 in any tested line relative to the uninduced condition. Following doxycycline withdrawal, proliferation is reported to be within 90% of the uninduced rate by day 7 in all tested lines.
A separate densitometry example reports uninduced MYC protein at 0.94 of parental within the calibration range, while induced protein is below a floor of 0.1 of parental at 72 hours. Because the induced value is below that floor, the example does not quantify the residual protein level.
These figures describe the contents of the synthetic example rather than a test of Symposium or a biological finding established by the paper. The source-study generalization to the tumor type is also identified as the authors' own interpretive step.
The hard questions come next
The document reports no empirical participants, experimental units or independent sample for evaluating the framework. It also reports no statistical analysis of Symposium's performance, so the proposal does not provide a measured estimate of how well the system works in practice.
That means the paper does not show that Symposium improves trust, accuracy, reproducibility, safety or speed in AI-assisted science. Nor does it show that structural validation ensures scientifically competent or correct Artifacts.
The next questions are practical: whether communities can judge Artifact quality and scientific competence reliably when the system checks only basic structure, how trust judgments should be defined for different purposes and stakes, and how the framework would perform at scale.
The reported project repository at github.com/ndexbio/symposium includes the Symposium specification, code, supporting infrastructure, examples and documentation. The implementation corresponding to the manuscript is identified as release v1.0, but availability of those materials does not establish adoption, usability or performance.
The work reports support from National Institutes of Health grants U24 CA269436, 5U24HG012107 and U19AI135990.
Paper data and sources
Original title: Symposium: Trust via Auditable Records for Communities of AI Scientist Agents
Authors: Dexter Pratt
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text